Multifunctional radar working mode identification method

Abstract

The invention discloses a multifunctional radar working mode recognition method, and relates to the field of multifunctional radar mode recognition, and the method comprises the steps: obtaining a radar pulse signal at a target, and the radar pulse signal at the target is transmitted by a radar with a to-be-recognized working mode. Pulse features are extracted according to the radar pulse signals at the target, and a pulse graph structure is constructed according to the radar pulse signals at the target. And building a graph convolutional neural network model based on the pulse features and the pulse graph structure. And inputting a radar pulse signal to be identified into the graph convolutional neural network model, and determining the working mode of the radar. According to the method, the graph convolutional neural network model can be built by combining the time features and the space features of the pulses, the working mode recognition requirement of the radar is met in the non-cooperative battlefield environment, and the recognition accuracy of the working mode is improved.

Description

Technical Field

[0001] The present invention relates to the field of radar pattern recognition, and particularly to a method for recognizing the working modes of a multi-functional radar. Background Art

[0002] A multi-functional radar is a radar system that can perform multiple tasks simultaneously. It has the characteristics of agile waveform and beam changes, and can adaptively adjust the working mode and resource scheduling according to factors such as the change of the battlefield environment and the target priority to obtain better detection results. With the increasingly complex electromagnetic environment, the mapping relationship between radar pulse signals and working modes becomes more difficult to learn, posing a huge challenge to the pattern recognition of multi-functional radars.

[0003] The existing technologies for the method of recognizing the working modes of multi-functional radars mainly include two categories: one is the working mode recognition based on prior knowledge information. This method collects or organizes a large amount of information such as signal characteristics and parameter ranges of the radar in different working modes in advance. During the actual recognition process, the received signals are analyzed and inferred, relevant features are extracted, and then these feature parameters are compared with the existing prior knowledge to achieve the recognition of the working modes of multi-functional radars; the other is the working mode recognition based on traditional machine learning and deep learning. This method is based on model data, and automatically extracts and selects suitable signal features through data analysis to obtain high-dimensional and more discriminative signal features to achieve the recognition of the working modes of multi-functional radars.

[0004] However, the first method requires a large amount of prior information for modeling. For a non-cooperative battlefield environment, it is unrealistic to accurately obtain the prior information of the electromagnetic environment. The second method belongs to the working mode recognition in a fixed scenario. In the face of a new working scenario, its generalization ability may be insufficient and accurate recognition cannot be performed. Summary of the Invention

[0005] In view of the above problems, the present invention provides a method for recognizing the working modes of a multi-functional radar, which overcomes or at least partially solves the problem of difficult recognition of the working modes of a multi-functional radar in a non-cooperative battlefield environment.

[0006] The present invention provides a method for recognizing the working modes of a multi-functional radar, including: obtaining the radar pulse signal at the target, where the radar pulse signal at the target is sent by the radar whose working mode is to be recognized. Extracting pulse features according to the radar pulse signal at the target, and constructing a pulse graph structure according to the radar pulse signal at the target. Building a graph convolutional neural network model based on the pulse features and the pulse graph structure. Inputting the radar pulse signal to be recognized into the graph convolutional neural network model to determine the working mode of the radar.

[0007] In the embodiments of the present invention, the time characteristics of the pulse can be determined according to the extracted pulse characteristics, and there is no need to obtain the prior information of the electromagnetic environment. According to the constructed pulse graph structure, the spatial characteristics of the pulse can be determined, and the problem of multifunctional radar operating mode recognition is modeled as a graph classification problem. By combining the time characteristics and spatial characteristics of the pulse, a graph convolutional neural network model can be built to meet the radar operating mode recognition requirements in a non-cooperative battlefield environment and improve the recognition accuracy of the operating mode.

[0008] In an alternative manner, the pulse characteristics include pulse width, pulse repetition period, amplitude, duty cycle, and frequency.

[0009] In an alternative manner, according to the radar pulse signal at the target, constructing a pulse graph structure includes: determining a single pulse as a node, and constructing a distance matrix according to the distance of each element in two nodes. Based on the distance matrix, the shortest cumulative distance is determined. According to the shortest cumulative distance between nodes, the scanning process of the radar with the working mode to be recognized is reconstructed, and thus the pulse graph structure is constructed.

[0010] In the embodiments of the present invention, the pulse graph structure constructed based on dynamic time programming can effectively extract the characteristics between nodes, construct the connection relationship between nodes according to the shortest cumulative distance between nodes, and use it as the adjacency matrix for subsequent construction of the graph convolutional neural network model.

[0011] In an alternative manner, before extracting the pulse characteristics according to the radar pulse signal at the target, the method further includes: segmenting the radar pulse signal at the target.

[0012] In the embodiments of the present invention, by segmenting the radar pulse signal at the target, the pulse characteristics can be better extracted.

[0013] In an alternative manner, the graph convolutional neural network model uses a rectified linear activation function.

[0014] In an alternative manner, after obtaining the radar pulse signal at the target, the method further includes: dividing the obtained radar pulse signal at the target into a training set, a validation set, and a test set. The training set is used to train the graph convolutional neural network model, the validation set is used to verify and optimize the graph convolutional neural network model, and the test set is used to detect the recognition ability of the graph convolutional neural network model.

[0015] The method of the present invention only needs to estimate the echo signal and its pulse characteristics, without other state information of the environment, overcoming the disadvantage that the working mode recognition method based on prior information requires accurate estimation of electromagnetic environment information; this method reconstructs the multi-functional radar scanning information by using the correlation between pulses, accurately captures the dependence relationship between pulses and makes reasonable use of it. At the same time, this method adaptively extracts the correlation between pulses by using the dynamic time warping method, overcoming the disadvantages that the feature extraction of the working mode recognition method based on traditional machine learning is highly subjective and it is difficult to capture the relationship between pulses; this method constructs a discriminative structural feature to reduce the dependence of the model on the training data, overcoming the disadvantage that the generalization ability of the working mode recognition method based on deep learning is poor.

[0016] The above description is only an overview of the technical solution of the present invention. In order to be able to understand the technical means of the present invention more clearly, it can be implemented according to the content of the specification. And in order to make the above and other purposes, features and advantages of the present invention more obvious and understandable, the specific embodiments of the present invention are specifically exemplified below. Brief Description of the Drawings

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0018] Figure 1 It is a schematic flow chart of a method provided for some embodiments of the present invention.

[0019] Figure 2 It is a schematic diagram of the construction of the simulation model.

[0020] Figure 3 It is a flow chart of the radar working mode transition.

[0021] Figure 4 It is the training and verification results of six network models during the recognition process.

[0022] Figure 5 It is the simulation result of the comparison of the recognition accuracies of six network models in a non-cooperative battlefield environment. Detailed Description of the Embodiments

[0023] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0024] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which this invention belongs; the terms used in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0025] The terms "comprising" and "having" and any variations thereof in the specification and claims of the present invention and the accompanying drawings are intended to cover but not exclude other elements. The word "a" or "an" does not exclude the presence of a plurality.

[0026] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in connection with the embodiments can be included in at least one embodiment of the present invention. The phrase "embodiments" appearing in various places in the specification is not necessarily referring to the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art will explicitly and implicitly understand that the embodiments described herein can be combined with other embodiments.

[0027] In addition, the terms "first", "second", etc. in the specification and claims of the present invention or the above-mentioned drawings are used to distinguish different objects and are not used to describe a specific order, and may explicitly or implicitly include one or more of such features.

[0028] In the description of the present invention, unless otherwise stated, the meaning of "a plurality" refers to two or more (including two), and similarly, "a plurality of groups" refers to two or more groups (including two groups).

[0029] Figure 1 It is a schematic flowchart of a method for identifying the working modes of a multifunctional radar provided for some embodiments of the present invention. Refer to Figure 1 , a method for identifying the working modes of a multifunctional radar provided for some embodiments of the present invention can be used to identify the working modes of a multifunctional radar in a non-cooperative battlefield environment.

[0030] As Figure 1 shown, a method for identifying the working modes of a multifunctional radar provided for some embodiments of the present invention includes the following steps 101 to 105:

[0031] Step 101: Obtain the radar pulse signal at the target location.

[0032] Among them, the radar pulse signal at the target location is sent by the radar of the working mode to be recognized. The radar pulse signal at the target location refers to the radar pulse signal sent by the radar of the working mode to be recognized received at the target location. This radar pulse signal is the radar pulse signal after propagation attenuation.

[0033] Step 102: Extract pulse features according to the radar pulse signal at the target location.

[0034] Specifically, the pulse features include pulse width, pulse repetition period, amplitude, duty cycle, and frequency. Extracting pulse features according to the radar pulse signal at the target location means extracting parameters such as the pulse width, pulse repetition period, amplitude, duty cycle, and frequency of the pulse according to the radar pulse signal at the target location and normalizing them as pulse features.

[0035] Step 103: Construct a pulse graph structure according to the radar pulse signal at the target location.

[0036] Step 104: Build a graph convolutional neural network model based on the pulse features and the pulse graph structure.

[0037] Step 105: Input the radar pulse signal to be recognized into the graph convolutional neural network model to determine the working mode of the radar.

[0038] In this embodiment, the time features of the pulse can be determined according to the extracted pulse features, and there is no need to obtain prior information of the electromagnetic environment. The spatial features of the pulse can be determined according to the constructed pulse graph structure, and the problem of multi-functional radar working mode recognition is modeled as a graph classification problem. Combining the time features and spatial features of the pulse can build a graph convolutional neural network model to meet the requirements of radar working mode recognition in a non-cooperative battlefield environment and improve the recognition accuracy of the working mode.

[0039] In some embodiments, constructing a pulse graph structure according to the radar pulse signal at the target location includes: determining a single pulse as a node, and constructing a distance matrix according to the distance between each element in two nodes. Determine the shortest cumulative distance based on the distance matrix. Reconstruct the scanning process of the radar of the working mode to be recognized according to the shortest cumulative distance between nodes, and thus construct a pulse graph structure.

[0040] Specifically, a dynamic time warping algorithm can be used to construct a pulse graph structure.

[0041] Assume that two nodes are X and Y, with lengths m and n respectively. Calculate the distance between each element in the two nodes to construct a distance matrix D, where D(i,j) represents the distance between x i and y j Among them, x iLet the pulse be an element in X, and y j be an element in the pulse Y. Define the cumulative distance matrix M to store the shortest cumulative distance from the starting node to each node. The cumulative distance M(i,j) can be calculated by the following formula:

[0042] M(i,j) = (i,j) + min{M(i - 1,j), M(i,j - 1), M(i - 1,j - 1)}

[0043] Finally, M(m,n) is the shortest cumulative distance DTW between the two nodes, representing the minimum matching cost between the two nodes considering time warping. According to the shortest cumulative distance DTW between the nodes, the scanning process of the radar for the working mode to be recognized can be reconstructed. When the shortest cumulative distance DTW between the obtained nodes is relatively high, it can be considered that the two nodes are adjacent in space. Based on this, the adjacency matrix A can be constructed according to the spatial information of the beam scanning, and the pulse diagram structure is represented by the adjacency matrix A.

[0044] In this embodiment, the pulse diagram structure constructed based on dynamic time programming can effectively extract the features between the nodes, construct the connection relationship between the nodes according to the shortest cumulative distance DTW between the nodes, and use it as the adjacency matrix A for subsequent construction of the graph convolutional neural network model.

[0045] Specifically, the process of constructing the graph convolutional neural network model based on the pulse features and the pulse diagram structure is as follows:

[0046] Construct a two-layer feature extraction network to extract pulse features. Represent the pulse features with the first feature matrix X, X = {x 1,2 ,..., N}, X ∈ R N×M , where M represents the number of features, and x 1,2 represents each node. There are a total of N nodes, and each node has M features. These M features are the pulse width, pulse repetition period, amplitude, duty cycle, and frequency, etc. The adjacency matrix A ∈ R N×N , and the constructed graph convolutional neural network model can be mainly represented by the feature matrix H (l+1) of the (l + 1)-th layer. The process of obtaining H (l+1) is as follows:

[0047] H (0) =

[0048]

[0049] #

[0050] where σ is the activation function, H (l+1) is the feature matrix of the (l + 1)-th layer of the graph convolutional neural network model, is the degree matrix, Indicates that the adjacency matrix A between nodes has added a self-connection I N , where I N is the identity matrix. Indicates a normalization operation. This normalization takes into account the degrees of nodes, making the propagation of information on the graph more balanced. W (l) is the graph node weight parameter matrix. The feature matrix H of each node at the l+1 layer (l+1) is obtained by weighted summation of the feature matrices of its own and neighbor nodes at the l layer. In this way, the node can fuse its own and neighbor information and make full use of the dependencies between nodes. Moreover, a graph convolutional neural network model is used to extract the spatial characteristics of the multifunctional radar scanning waveform, comprehensively considering the overall and local characteristics between nodes. The graph convolutional neural network model can fully identify the structural differences of different categories through multiple rounds of iteration.

[0051] In some embodiments, the graph convolutional neural network model uses a rectified linear activation function, that is, the activation function σ uses a rectified linear activation function. This speeds up the gradient calculation and update speed, promotes the rapid convergence of the graph convolutional neural network model, and significantly shortens the training time.

[0052] Input the feature matrix H of the 3rd layer of the last convolutional layer (3) into the fully connected layer. Through linear layer and batch normalization processing, a softmax classifier is used for classification, and finally the classification result is obtained through the output layer, thereby successfully and accurately identifying the working modes of the multifunctional radar.

[0053] In some embodiments, before extracting pulse features from the radar pulse signal at the target, the method further includes: segmenting the radar pulse signal at the target. In this embodiment, by segmenting the radar pulse signal at the target, pulse features can be better extracted. For example, the radar pulse signal at the target can be segmented by wavelet transform. Wavelet transform is a method of converting a time-domain signal into a frequency-domain signal to extract features, which can better express the stationary and non-stationary parts of the signal. The detail coefficients obtained by wavelet transform can roughly determine the conversion points of the working modes of the multifunctional radar. To avoid the unreliability of a single parameter, wavelet transform is used to comprehensively test multi-dimensional parameters, and the positions of the conversion points are determined from multiple dimensions, laying a foundation for subsequent identification of the working modes of the multifunctional radar.

[0054] In some embodiments, after obtaining the radar pulse signal at the target, the method further includes: dividing the obtained radar pulse signal at the target into a training set, a validation set, and a test set. The training set is used to train the graph convolutional neural network model, the validation set is used to verify and optimize the graph convolutional neural network model, and the test set is used to detect the recognition ability of the graph convolutional neural network model.

[0055] Specifically, a training set is used to train the graph convolutional neural network model. The cross-entropy loss function is calculated for the classification results output by the output layer to quantify the difference between the data distribution of the radar working mode predicted by the model and the data distribution of the real radar working mode. The gradient descent optimization algorithm is used to update the relevant parameters of the graph convolutional neural network model. For example, the relevant parameters of the graph convolutional neural network model can be the learning rate, weight decay, etc. The validation set is used to validate and optimize the graph convolutional neural network model, and the test set is used to detect the recognition ability of the graph convolutional neural network model. Through the final recognition results, the performance of the graph convolutional neural network model is analyzed. When it is detected that the performance of the graph convolutional neural network model provided by the present invention is excellent, it can be used in actual use. By inputting the radar pulse signal to be recognized into this graph convolutional neural network model, the working mode of the radar can be determined.

[0056] In the simulation experiment of the present invention, the pattern recognition ability and adaptability of the traditional deep learning method convolutional neural network model (Convolutional Neural Networks, CNN) (taking CNN-12 as an example), LeNet network model, VGGNet network model, residual convolutional network model ResNet (taking ResNet-18 as an example), conventional graph convolutional neural network model GCN, and the graph convolutional neural network model provided by the present invention (abbreviated as DTW-GCN) will be compared and analyzed.

[0057] The present invention mainly models the change of the radar working mode in a complex electromagnetic environment and simulates and verifies the performance of the multi-functional radar working mode recognition method proposed by the present invention. Among them, the radar working mode generation strategy in the complex electromagnetic environment does not change with time.

[0058] In the simulation experiment, it is assumed that there are differences in the transmission parameters of radars in different working modes. The radar working modes are divided into four types: search mode, tracking mode, search-while-tracking mode, and search-plus-tracking mode. The characteristic parameters of each working mode are shown in Table 1. Table 1 shows the transmission parameters of 4 phased array radar working modes, including frequency (RF), pulse width (PW), power (P), pulse repetition frequency (PRF), and duty cycle (DC).

[0059] Table 1 Transmission parameters of each radar working mode

[0060] Working mode RF (GHz) PW (μs) P (KW) PRF (KHz) DC Search 2.6-5 35-50 40-50 7-13 0.2-0.5 Tracking 2.8-3.8 10-25 18-30 15-25 0.1-0.2 Search while tracking 2.7-3.3 15-30 5-10 6-16 0.2-0.3 Search plus tracking 3-4.6 5-25 14-20 4-20 0.3-0.4

[0061] In the simulation experiment, to make the model results conform to the non-cooperative battlefield environment, the transmission parameters of the above working modes are used for simulation modeling to realize the radar detection process of radar scanning environment - detecting targets - changing working modes, and taking the radar emission signal in this process as the research object to realize the radar working mode recognition in the non-cooperative battlefield scenario.Figure 2 It is a schematic diagram for the construction of a simulation model. Assume that there is a radar in the environment. The scanning range of the radar is [-40°, 40°] in the azimuth direction and [0°, 30°] in the elevation direction, and it does not change with time. The scanning mode is circular scanning. The false alarm probability during the detection process is P fa = 10 -4 , and the detection probability is P d > 99%. Taking the radar as the origin, a rectangular coordinate system is constructed. Assume that the target is located at (6500, 2100, 0) m and moves at a speed of (-300, -400, 0) m / s. The radar first conducts a spatial scan to search for the target. If the target is detected, it enters the next working mode Figure 3 It is a flow chart for the transfer of the radar working mode. Considering the spatial propagation loss of the radar signal captured by the receiver of the jammer, it is difficult to obtain the radar transmission power. Therefore, frequency, pulse width, pulse amplitude, pulse repetition frequency, and duty cycle are used as pulse characteristics, and dynamic time warping and graph convolutional neural network are combined to realize working mode recognition

[0062] Specifically, in the simulation environment, the network uses the Adam optimizer, the learning rate is set to 0.001, the batch size is set to 128, and each training is carried out for 150 rounds. To ensure the effectiveness of the experiment, each deep learning network is trained 10 times, and shaded filling is performed according to its standard deviation. The mean value of the convergent model in 10 trainings is taken as the value of the training curve Figure 5 It is the training and verification results of six network models during the recognition process. As Figure 4 can be seen, the VGGNet network model shows relatively serious overfitting, and the average accuracy of the validation set and the training set differs by about 10%. The CNN also has this problem to a certain extent. The accuracy of the training set and the validation set of the DTW-GCN of the present invention is basically consistent, which is better than the above two network models. From the perspective of the standard deviation of the validation curve, the CNN-12 network model, the ResNet-18 network model, and the VGGNet network model have large fluctuations and poor model stability, and it is necessary to train repeatedly many times to find the optimal structure. Compared with the other 5 networks of the present invention, the standard deviation is the smallest and the model is the most stable. From the perspective of the recognition rate in the training and validation stages, all 6 models ensure an accuracy rate of more than 90%. The present invention is slightly better than the recognition effect of the conventional graph convolutional neural network model GCN, and is completely convergent, and the recognition rate is the best among the 6 networks, which proves the importance of the shortest cumulative distance DTW between nodes, that is, the pulse-to-pulse correlation in working mode recognition. It can also be seen from Table 2 below that the recognition accuracy of the present invention is the highest

[0063] Table 2 Optimal recognition results of the network model for the validation set

[0064]

[0065]

[0066] Considering that in a non - cooperative battlefield environment, it is often impossible to obtain all the radar pulse signals at the target, the integrity of the intercepted radar pulse signals at the target has an important impact on identifying the working mode of the radar. Therefore, the above - mentioned 6 network models are selected for testing. The test set is generated independently of the training set. A test set with an interval of 5% and a missing - pulse ratio of 0 - 50% is set up to verify the generalization ability of each network model.

[0067] Figure 5 It is the simulation result of the recognition accuracy comparison of six network models in a non - cooperative battlefield environment. It can be seen that since the integrity of the pulse data has a great influence on the recognition accuracy of the network model, when the missing - pulse ratio increases, the recognition ability of each network model changes. Figure 5 It can be seen that the recognition performance of the network model provided by the present invention decreases slowly as the missing - pulse ratio increases. However, the network model provided by the present invention fully explores the correlation between the data among the pulses, inversely calculates the spatial features during the radar scanning process, is less affected by the missing pulses, and is superior to other network models, proving that the multi - function radar working mode recognition method provided by the present invention also has good adaptability in the case of pulse data loss in a non - cooperative battlefield environment.

[0068] Those skilled in the art can understand that although some of the embodiments herein include certain features included in other embodiments, the combination of the features of different embodiments means that it is within the scope of the present invention and forms different embodiments. For example, in the claims, any one of the claimed embodiments can be used in any combination.

[0069] As mentioned above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying multifunctional radar operating modes, characterized in that The method includes: Obtaining a radar pulse signal at a target, where the radar pulse signal at the target is sent by a radar with a working mode to be identified; Extracting pulse features according to the radar pulse signal at the target; Constructing a pulse graph structure according to the radar pulse signal at the target; Building a graph convolutional neural network model based on the pulse features and the pulse graph structure; Inputting the radar pulse signal to be identified into the graph convolutional neural network model to determine the working mode of the radar.

2. The method according to claim 1, wherein The pulse features include pulse width, pulse repetition period, amplitude, duty cycle, and frequency.

3. The method according to claim 1, wherein The constructing of the pulse graph structure according to the radar pulse signal at the target includes: Determining a single pulse as a node; Constructing a distance matrix according to the distance between each element in two nodes; Determining the shortest cumulative distance based on the distance matrix; Reconstructing the scanning process of the radar with the working mode to be identified according to the shortest cumulative distance between nodes, thereby constructing a pulse graph structure.

4. The method according to claim 1, wherein Before extracting the pulse features according to the radar pulse signal at the target, the method further includes: segmenting the radar pulse signal at the target.

5. The method according to claim 1, wherein The graph convolutional neural network model uses a rectified linear activation function.

6. The method according to claim 1, wherein After obtaining the radar pulse signal at the target, the method further includes: Dividing the obtained radar pulse signal at the target into a training set, a validation set, and a test set; The training set is used to train the graph convolutional neural network model, the validation set is used to verify and optimize the graph convolutional neural network model, and the test set is used to detect the recognition ability of the graph convolutional neural network model.

Images