A method for identifying radar radiation sources

Abstract

The present invention discloses a method for identifying radar radiation sources. The method specifically includes the following steps: S1, extracting the feature vector of the time-domain signal to be identified; S2, inputting the feature vector into a trained XGBoost model, and the XGBoost model outputs the radar radiation source of the time-domain signal to be identified. According to the time-domain signal data of the received radar radiation source, a denoising method based on regional features is first proposed, and a self-supervised model is proposed to extract time-domain features, and then combined with ensemble learning for identification. The method disclosed by the present invention has high flexibility, does not depend on the dimension and sampling method of database samples, and can be identified in multiple different scenarios; moreover, the method has digital reasoning ability, can effectively process abnormal signals, and improve the robust performance of the model. In addition, the algorithm is not restricted by knowledge level and expert experience, and there is no need to pre-form an expert knowledge base, and the overall intelligence level is high.

Description

Technical Field

[0001] The present invention belongs to the technical field of signal source recognition. More specifically, the present invention relates to a method for identifying radar emitters. Background Art

[0002] Radar Emitter Recognition (RER) technology is a crucial part of electronic warfare, the core of electronic support measures, and a key system in radar countermeasure systems. The emitter recognition technology provides a basis for the next step of situation assessment, threat assessment, and decision-making adjustment by obtaining the system, purpose, model, and other signals of the target emitter from the reconnaissance signal.

[0003] The conventional feature technology of emitter signals mainly refers to the characteristic parameters directly measured by signal detection. The method based on feature matching first needs to construct a database that matches the signal to be identified, and then processes and analyzes the intercepted radar emitter signal to obtain its characteristic parameters. This feature matching method has a simple idea, is easy to understand, has a small computational load, good real-time performance, and is easy to implement in engineering. However, the key of the feature matching algorithm lies in the search strategy and the matching algorithm. Its flexibility is poor, it highly depends on the sample data of the matching database, and it can achieve better recognition performance only in specific scenarios. Moreover, the recognition method based on feature matching does not have digital reasoning ability, cannot process abnormal signals, and is difficult to cope with the complex and changeable modern battlefield electromagnetic environment.

[0004] The recognition method based on an expert system designs inference rules by experienced radar experts based on systematic radar knowledge, and uses the characteristics and parameters of the intercepted signal to make inference and judgment to obtain unknown radar emitter information. This system has a certain logical reasoning ability, can process new system radar signals with uncertainty, identify unknown radar signals, analyze and infer information such as its radar system and attributes, and thus formulate targeted countermeasure strategies. Although the recognition accuracy and flexibility of this method have been improved compared with the feature matching method, and the functions implemented are more complex, the expert system has high requirements for information fusion algorithms, and the reasoning ability of the system is significantly restricted by the knowledge level and expert experience, and the overall intelligence level is not high. Summary of the Invention

[0005] The present invention provides a method for identifying radar emitters, aiming to improve the above problems.

[0006] The present invention is implemented as follows. A method for identifying radar emitters, the method specifically includes the following steps:

[0007] S1. Extract the feature vector of the time-domain signal to be identified;

[0008] S2. Input the feature vector into the trained XGBoost model, and the XGBoost model outputs the radar emitter of the time-domain signal to be recognized.

[0009] Further, the training process of the XGBoost model is specifically as follows:

[0010] Perform radar emitter annotation on the time-domain signal from which the feature vector has been extracted to form samples;

[0011] Train the XGBoost model based on the training samples, test the XGBoost model based on the test samples, and after the recognition accuracy of the XGBoost model reaches the set accuracy threshold, the XGBoost model training is completed.

[0012] Further, the feature vector of the time-domain signal is formed by horizontally splicing feature A, feature B, and feature C.

[0013] Further, the extraction method of feature A is specifically as follows:

[0014] (1) Divide the time-domain signal into n segments, randomly scramble the n segments m times, and splice the n segments after each scrambling to form m scrambled time-domain signals

[0015] (2) Calculate the Hamming distance between the time-domain signal and the m scrambled time-domain signals, and label the time-domain signal and the m scrambled time-domain signals based on the Hamming distance;

[0016] (3) Input the time-domain signal and the m scrambled time-domain signals into the autoencoder, and use their corresponding labeled values as the actual outputs;

[0017] (4) Define the cross-entropy between the actual output and the predicted output of the autoencoder as the loss function, and optimize the loss function to obtain feature A.

[0018] Further, the extraction process of feature B is specifically as follows:

[0019] (1) Use the K-Means clustering algorithm to cluster the time-domain signal r times, and the values of K are a1, a2,..., a r ;

[0020] (2) Use the time-domain signal as the input of the autoencoder, and use the r encodings of each time-domain signal as the actual output;

[0021] (3) Define the cross-entropy between the actual output and the predicted output of the autoencoder as the loss function, and optimize the loss function to obtain feature B.

[0022] Further, the extraction process of feature C is specifically as follows:

[0023] (1) Modify the random-length data at random positions in the time-domain signal using the "missing data" method to form a modified time-domain signal. Modify it w times to form w modified time-domain signals.

[0024] (2) Input the time-domain signal and the w modified time-domain signals into an autoencoder, and add a decoder to obtain the output signal 1 of the time-domain signal and the output signals 2 of the w modified time-domain signals.

[0025] (3) Calculate the sum of the Euclidean distances between the output signal 1 and the w output signals 2, and use the sum of the Euclidean distances as a loss function. Optimize the loss function to obtain the feature C.

[0026] Furthermore, before extracting the feature vector from the time-domain signal, it also includes:

[0027] Encode the time-domain signals of different radiation sources.

[0028] Denoise the encoded time-domain signals.

[0029] Furthermore, the method for denoising the time-domain signal is specifically as follows:

[0030] Divide the time-domain signal into regions based on a sliding window, and obtain the feature points of each region, including: peak value, extreme value, average value, median value, and the outlier point farthest from the average value.

[0031] Calculate the sum of the feature point distances between two regions, and detect whether the sum of the distances is less than the distance threshold. If the detection result is yes, delete one of the regions.

[0032] Traverse any two regions, and splice the remaining regions to form the denoised time-domain signal.

[0033] Furthermore, it includes 7 different types of radiation source signals, namely: single carrier frequency, linear frequency modulation, non-linear frequency modulation, binary phase coding, quaternary phase coding, binary frequency coding, and quaternary frequency coding.

[0034] The self-supervised radiation source signal feature extraction algorithm proposed by the present invention is different from the traditional feature matching algorithm, has high flexibility, does not depend on the dimension and sampling method of the database samples, and can be recognized in multiple different scenarios; moreover, this algorithm has digital reasoning ability, can effectively process abnormal signals, and improve the robustness of the model. This algorithm is also different from the feature extraction algorithm based on an expert system, is not restricted by the knowledge level and expert experience, and does not require a pre-formed expert knowledge base, with a high overall degree of intelligence. Description of the Drawings

[0035] Figure 1Seven types of radar emitter codes provided by the embodiments of the present invention, where (A) is a single carrier frequency, (B) is a linear frequency modulation, (C) is a non-linear frequency modulation, (D) is a binary phase encoding, (E) is a quadrature phase encoding, (F) is a binary frequency encoding, and (G) is a quadrature frequency encoding;

[0036] Figure 2 Flowchart of the denoising method for the time-domain signal provided by the embodiments of the present invention;

[0037] Figure 3 Schematic diagram of the framework of the autoencoder provided by the embodiments of the present invention;

[0038] Figure 4 Schematic diagram of the self-supervised model provided by the embodiments of the present invention;

[0039] Figure 5 Schematic diagram of the XGBoost framework provided by the embodiments of the present invention. Detailed implementation manners

[0040] The following further details the specific implementation manners of the present invention by describing the embodiments with reference to the accompanying drawings, so as to help those skilled in the art have a more complete, accurate, and in-depth understanding of the inventive concept and technical solutions of the present invention.

[0041] Based on the time-domain signal data of the received radar emitters, the present invention first proposes a denoising method based on regional features, and proposes a self-supervised model to extract time-domain features, and then combines ensemble learning for identification. The radar emitter identification method provided by the embodiments of the present invention includes the following steps:

[0042] S1. Receive the time-domain signals of different radar emitters

[0043] The present invention includes seven different emitter signals, namely: continuous wave (CW), linear frequency modulation (LFM), non-linear frequency modulation (NLFM), binary phase shift keying (BPSK), quadrature phase shift keying (QPSK), binary frequency shift keying (BFSK), and quadrature frequency shift keying (QFSK), as Figure 1 shown.

[0044] S2. For the time-domain signal Perform denoising to obtain the denoised time-domain signal

[0045] Since noise is random, the characteristics of the radiation source signal are not easily recognizable in regions with low signal-to-noise ratio, causing interference to the recognition of the entire signal. The present invention proposes a denoising method based on regional similarity. Different from traditional dimensionality reduction, considering periodicity and time series Figure 2 The following is a flowchart of the denoising method for the time-domain signal provided by the embodiment of the present invention. The method is as follows:

[0046] Based on a sliding window for the time-domain signal Perform regional division to obtain the characteristic points of each region, including: peak value, extreme value, average value, median value, and outlier point farthest from the average value;

[0047] Calculate the sum of the distances of the characteristic points of two regions. The smaller the sum of the distances of the characteristic points, the higher the similarity between the two regions, and the larger the sum of the distances of the characteristic points, the lower the similarity between the two regions.

[0048] Detect whether the sum of the distances is less than the distance threshold. If the detection result is yes, delete one of the regions;

[0049] Traverse any two regions and splice the remaining regions to form the denoised time-domain signal.

[0050] Sort the window regions according to similarity and eliminate signals with high similarity. This calculation method can be independent of the length of the radiation source signal and can well eliminate redundant and low signal-to-noise ratio sequence signals.

[0051] In the embodiment of the present invention, the method for determining the distance threshold is as follows:

[0052] Arrange all the sums of the distances of the characteristic points in ascending order, and determine the distance threshold of the sum of the distances of the characteristic points based on the removal ratio. For example, there are 100 sums of the distances of the characteristic points and they are arranged in ascending order. If 20% is to be removed, then set the sum of the distances of the characteristic points ranked 20th as the distance threshold to retain 80% of the regional signals.

[0053] S3. Extract the eigenvectors of the denoised time-domain signal Extract the eigenvectors;

[0054] In the embodiment of the present invention, the eigenvector is formed by horizontally splicing Feature A, Feature B, and Feature C. In the embodiment of the present invention, in combination with Figure 3 The provided autoencoder structure diagram is used to illustrate the extraction process of Feature A, Feature B, and Feature C;

[0055] In the embodiment of the present invention, the extraction method of feature A is specifically as follows:

[0056] (1) Segment the denoised time-domain signal into n segments, randomly scramble the n segments m times, and splice the n segments after each scrambling to form m scrambled time-domain signals

[0057] (2) Calculate the Hamming distance between the time-domain signal and the m scrambled time-domain signals , and mark the time-domain signal and the m scrambled time-domain signals based on the Hamming distance. The greater the Hamming distance, the greater the marking value.

[0058] (3) Input the time-domain signal and the m scrambled time-domain signals into the autoencoder, and use their corresponding marking values as the actual output;

[0059] (4) Define the cross-entropy between the actual output and the predicted output of the autoencoder as the loss function, optimize the loss function, implement self-supervised model training, and obtain feature A.

[0060] Segment the denoised time-domain signal (original time-domain signal) into 10 segments according to a 10% ratio, randomly shuffle the 10 segments 10 times, and splice them into a time-domain signal (scrambled time-domain signal) after each shuffle to form 10 scrambled time-domain signals; mark them with 0 - 100 (form a 101-dimensional marking vector), 0 represents the original time-domain signal, and the larger the value, the greater the difference between the scrambled time-domain signal and the original signal; input the original time-domain signal and the scrambled time-domain signal together into the autoencoder as the input, and use their corresponding markings (0 - 101) as the actual output; define the cross-entropy between the actual output and the predicted output as the loss function, optimize the loss function, implement self-supervised model training, and obtain feature A.

[0061] In the embodiment of the present invention, the extraction process of feature B is specifically as follows:

[0062] (1) Use the K-Means clustering algorithm to cluster the denoised time-domain signal r times, and the values of K are successively a1, a2,..., a r . In the same clustering process, the time-domain signals in different clusters are marked with the same length but different numerical encodings, and in different clustering processes, the time-domain signals are marked with encodings of different lengths;

[0063] Since each denoised time-domain signal To participate in r clustering operations, the encoding under the current K value is obtained in each clustering process. Therefore, each denoised time-domain signal has r labels. If clustering is performed 3 times and K is taken as 10, 100, and 200 respectively, all radiation source signals will have three types of encodings, which are 0-9, 0-99, and 0-199 respectively.

[0064] (2) Use the denoised time-domain signal as the input of the autoencoder, and the r encodings of each time-domain signal are used as the actual output;

[0065] (3) Define the cross-entropy between the actual output and the predicted output of the autoencoder as the loss function, optimize the loss function, and implement self-supervised model training to obtain feature B.

[0066] In the embodiments of the present invention, the extraction process of feature C is specifically as follows:

[0067] (1) Modify the data with random length at random positions in the time-domain signal using the "missing data" method to form a modified time-domain signal Modify it w times to form w modified time-domain signals The value range of the random length is: 1-10%;

[0068] Using the "missing data" method to modify means updating the signal value to 0.

[0069] (2) Input the time-domain signal and the w modified time-domain signals into the autoencoder together, and add a decoder to obtain the output signal of the time-domain signal and the output signals of the w modified time-domain signals

[0070] (3) Calculate the sum of the Euclidean distances between the output signal and the w output signals and use the sum of the Euclidean distances as the loss function; optimize the loss function to implement self-supervised model training to obtain feature C.

[0071] In these methods, both the encoder and the decoder are processed using a recurrent neural network, which can not only achieve the purpose of extracting non-linear features but also reduce the dimension of the data. The hidden space formed by the encoder can represent its latent features. Considering the temporal and periodic characteristics of radar radiation source signals, long short-term memory network units LSTM are used in the neurons of the encoder to solve the problems of gradient disappearance and gradient explosion in sequence problems. Horizontally splice feature A, feature B, and feature C to obtain the time-domain signal The eigenvector, Figure 4 Among those shown, the self-supervised model A, self-supervised model B, and self-supervised model C are respectively used for the extraction of feature A, feature B, and feature C.

[0072] S4. For the time-domain signal of the extracted eigenvector Perform the annotation of the radar emitter to form a sample. Train the XGBoost model based on the training sample, and test the XGBoost model based on the test sample. After the recognition accuracy of the XGBoost model reaches the set accuracy threshold, the XGBoost model training is completed.

[0073] S5. Extract the eigenvector of the time-domain signal to be recognized, input the eigenvector into the trained XGBoost model, and the XGBoost model outputs the radar emitter corresponding to the time-domain signal.

[0074] The method for extracting the eigenvector of the time-domain signal to be recognized is the same as above. The time-domain signal before eigenvector extraction needs to be denoised, and the denoising method is also the same as above. XGBoost is an ensemble learning model. As Figure 5 shown, the weights of the weak classifiers are summed to obtain the recognition result. The advantage of such ensemble learning lies in its high generalization performance, maintaining the timeliness of features, and at the same time, it can perform parallel computing to improve efficiency. The ratio of the training set to the test set is 80%:20%, the learning rate is 0.057, the number of weak classifiers is 500, and the maximum depth of the decision tree is 6.

[0075] The present invention has been described by way of example. Obviously, the specific implementation of the present invention is not limited by the above methods. As long as various non-substantive improvements are made by adopting the method concept and technical solution of the present invention, or the concept and technical solution of the present invention are directly applied to other occasions without improvement, they are all within the protection scope of the present invention.

Claims

1. A method for identifying radar radiation sources, characterized in that, The method specifically includes the following steps: S1. Extract the feature vector of the time-domain signal to be recognized; S2. Input the feature vector into the trained XGBoost model, and the XGBoost model outputs the radar emitter of the time-domain signal to be recognized; The feature vector of the time-domain signal is formed by horizontally splicing feature A, feature B, and feature C; The specific method for extracting feature A is as follows: (1) Segment the time-domain signal into n segments, randomly scramble the n segments m times, and splice the n segments after each scrambling to form m scrambled time-domain signals S i 3 ; (2) Calculate the Hamming distance between the time-domain signal and m scrambled time-domain signals, and label the time-domain signal and m scrambled time-domain signals based on the Hamming distance; (3) Input the time-domain signal and m scrambled time-domain signals into the autoencoder, and use their corresponding labeled values as the actual output; (4) Define the cross-entropy between the actual output and the predicted output of the autoencoder as the loss function, and optimize the loss function to obtain feature A; The specific process for extracting feature B is as follows: (1) Use the K-Means clustering algorithm to cluster the time-domain signal r times, and the values of K are a1, a2,..., a r respectively. In the same clustering process, the time-domain signals of different clusters are marked with encodings of the same length but different values, and in different clustering processes, the time-domain signals are marked with encodings of different lengths, obtaining r encodings of the denoised time-domain signal (2) Use the time-domain signal as the input of the autoencoder, and use the r encodings of each time-domain signal as the actual output; (3) Define the cross-entropy between the actual output and the predicted output of the autoencoder as the loss function, and optimize the loss function to obtain feature B; The specific process for extracting feature C is as follows: (1) Modify the random-length data at random positions in the time-domain signal using the missing data method to form a modified time-domain signal. Modify it w times to form w modified time-domain signals S i 4 ; (2) Input the time-domain signal and w modified time-domain signals into the autoencoder together, and add a decoder to obtain the output signal 1 of the time-domain signal and the output signals 2 of the w modified time-domain signals; (3) Calculate the sum of the Euclidean distances between the output signal 1 and the w output signals 2, and use the sum of the Euclidean distances as the loss function, and optimize the loss function to obtain feature C; Among them, modifying by using the missing data method means updating the signal value to 0.

2. The radar radiation source identification method according to claim 1, wherein, The training process of the XGBoost model is specifically as follows: Label the radar emitter of the time-domain signal with the extracted feature vector to form a sample; Train the XGBoost model based on the training samples, test the XGBoost model based on the test samples, and after the recognition accuracy of the XGBoost model reaches the set accuracy threshold, the XGBoost model training is completed.

3. The radar radiation source identification method according to claim 1, characterized in that, Before extracting the feature vector of the time-domain signal, it also includes: Denoise the time-domain signals of different emitters.

4. The radar radiation source identification method according to claim 3, characterized in that The specific method for denoising the time-domain signal is as follows: Based on a sliding window, divide the time-domain signal into regions, and obtain the feature points of each region, including: peak value, extreme value, average value, median value, and the outlier farthest from the average value; Calculate the sum of the feature point distances between two regions, and detect whether the sum of the distances is less than the distance threshold. If the detection result is yes, delete one of the regions; Traverse any two regions, and splice the remaining regions to form the denoised time-domain signal.

5. The radar radiation source identification method according to claim 1, characterized in that It includes 7 different emitter signals, namely: single carrier frequency, linear frequency modulation, non-linear frequency modulation, binary coding, quaternary coding, dual frequency coding, and quadruple frequency coding.

Images