A method for multifunctional radar working mode recognition based on smoothed graph signals

Abstract

The present invention relates to the field of electronic reconnaissance technology, and in particular to a method for identifying working modes of a multifunctional radar based on smoothed graph signals. First, the clustering results of a self-organizing map algorithm are used to represent global signal samples in the form of smoothed graph signals. Since the clustering results of the self-organizing map algorithm can not only directly reflect the topological relationship between clusters, but also reflect the spatial distribution of samples, and adjacent clusters are also adjacent in the sample space, the graph signal generated according to the cluster neighbor relationship has smoothness. By first generating a smoothed graph signal from the global sample, the problem of unbalanced working mode samples is solved. Then, a graph attention network is used to perform node-level data fusion and classification on the graph signal, thereby completing the task of identifying the working mode of the multifunctional radar under the condition of insufficient supervision data.

Description

Technical Field

[0001] The present invention relates to the technical field of electronic reconnaissance, and in particular to a multifunctional radar working mode recognition method based on smooth graph signals. Background Art

[0002] Multi-Function Radar (MFR) is widely used in electronic warfare due to its multiple operating modes, high intelligence, and strong reliability. This poses significant challenges to radar reconnaissance and countermeasures. The probability of interception of MFR signals is low, resulting in fewer signal samples being intercepted in a short period of time. Without sufficient countermeasure pressure, some of the MFR's operating modes may be hidden. Radar operating modes are a crucial factor in determining the threat level of radar systems and form the basis for electronic offensive and defensive decision-making.

[0003] Existing MFR operating mode recognition methods mainly include syntactic model-based methods, data fusion-based methods, and deep learning-based methods. Syntactic model-based methods classify operating modes by constructing corresponding automata based on the grammatical type of radar operating mode transitions. This method has strong theoretical advantages, but it is a forward modeling process and difficult to apply to non-cooperative parties. Data fusion-based methods first calculate the membership of the signal under test to a certain operating mode, then fuse the recognition results from multiple platforms or at different times to ultimately determine the operating mode to which the signal under test belongs. Although this method can achieve good recognition results, it requires sufficient prior information about the MFR signal. Deep learning-based methods recognize MFR operating modes by analyzing pulse characteristics and pulse group structure. However, the supervised data required to train deep learning models is difficult to obtain. In addition to the difficulty in obtaining large-scale supervised data, the sample set often exhibits an imbalance in operating mode samples due to the large difference in the number of samples captured between common MFR modes (such as search) and less common modes (such as tracking). Summary of the Invention

[0004] The purpose of the present invention is to provide a multi-function radar working mode recognition method based on smoothed graph signals, which can improve the technical problems of insufficient supervision data and unbalanced working mode samples in the existing multi-function radar working mode recognition tasks.

[0005] To achieve the above object, the present invention provides a method for identifying a multifunctional radar operating mode based on a smoothed graph signal, comprising the following steps:

[0006] Data preprocessing;

[0007] Use SOM algorithm to cluster global signal samples;

[0008] Generate smooth graph signals through the clustering results of SOM;

[0009] The smooth graph signal is subjected to node classification.

[0010] Among them, during the data preprocessing process, the data is standardized by the mean and standard deviation.

[0011] Among them, in the process of clustering global signal samples using the SOM algorithm, the distance between the input signal sample and the center of each cluster is first calculated. Specifically, the Euclidean distance is used to calculate the distance. The cluster with the smallest distance is the winning cluster. Then, the centers of the winning cluster and its neighboring clusters are adjusted to finally complete the topological mapping of the data set.

[0012] In the process of generating smooth graph signals through the clustering results of SOM, based on the conditions of full connection of samples within the cluster and partial connection of samples between adjacent clusters, one cluster is selected as the main cluster in each iteration, and the samples of the sample set contained in the main cluster are fully connected until all clusters have been set as main clusters.

[0013] In the process of node classification of the smooth graph signal, the feature vectors of all nodes of the smooth graph signal are input into the graph attention network, and the output layer of the graph attention network obtains the working mode to which the sample belongs by taking the average of multi-head attention.

[0014] The present invention provides a multi-function radar working mode recognition method based on smoothed graph signals. First, the clustering results of the self-organizing map algorithm are used to represent global signal samples in the form of smoothed graph signals. Since the clustering results of the self-organizing map algorithm can not only directly reflect the topological relationship between clusters, but also reflect the spatial distribution of samples, and adjacent clusters are also adjacent in the sample space, the graph signal generated according to the cluster neighbor relationship is smooth. By first generating a smoothed graph signal from the global samples, the problem of unbalanced working mode samples is solved. Then, a graph attention network is used to perform node-level data fusion and classification on the above-mentioned graph signal to complete the multi-function radar working mode recognition task under the condition of insufficient supervision data. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0016] Figure 1 The present invention is a flow chart of a method for identifying a multifunctional radar working mode based on a smoothed graph signal.

[0017] Figure 2 It is a schematic diagram of a confrontation scenario according to a specific embodiment of the present invention.

[0018] Figure 3 3 is a graph showing how the recognition results of a specific embodiment of the present invention change with the number of supervisory data in four different noise scenarios. DETAILED DESCRIPTION

[0019] The following describes embodiments of the present invention in detail, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and are not to be construed as limiting the present invention.

[0020] See also Figures 1 to 3 The present invention proposes a method for identifying a multifunctional radar working mode based on a smoothed graph signal, comprising the following steps:

[0021] S1: data preprocessing;

[0022] S2: Use SOM algorithm to cluster global signal samples;

[0023] S3: Generate smooth graph signals through the clustering results of SOM;

[0024] S4: performing node classification on the smooth graph signal.

[0025] Specifically, the data is standardized during the data preprocessing process.

[0026]

[0027] Where mean(·) and std(·) are the mean and standard deviation respectively.

[0028] In the process of clustering global signal samples using the SOM algorithm, the distance between the input signal sample and the center of each cluster is first calculated. The cluster with the smallest distance is the winning cluster. Then, the centers of the winning cluster and its neighboring clusters are adjusted to finally complete the topological mapping of the data set.

[0029] SOM is shown in Algorithm 1.

[0030]

[0031]

[0032] in Represents the neighborhood function centered on cluster u, v (row u, column v)

[0033]

[0034] Where σ is the neighborhood factor, represents the Euclidean distance between clusters u,v and i,j. The time complexity of this algorithm is about O(EN x rc), where N x is the number of input data.

[0035] Smooth graph signals are generated by clustering results from SOM. Since topologically adjacent clusters are also adjacent in the sample space, the transition between clusters is smooth. Therefore, the present invention provides a SOM-based smooth graph signal generation algorithm based on two aspects: full connectivity of samples within a cluster and partial connectivity of samples between adjacent clusters. The algorithm process is as follows:

[0036] (i) Select a cluster as the main cluster and initialize the connection ratio of the main cluster and neighbor clusters to r main 、r vice , the adjacency matrix The dimension is N x The identity matrix of .

[0037] (ii) Let the index of the sample set contained in the main cluster be where num main is the number of samples in the cluster. Then these samples are fully connected, that is, the corresponding items in the adjacency matrix are set to 1, A(idx main ,idx main )=1.

[0038] (iii) Calculate the distance between the samples in each neighboring cluster and the center of the main cluster, and take the indexes of the samples with the smallest distance. where N n is the number of neighbor clusters, num i is the number of samples contained in the ith neighbor cluster, and · is rounded up.

[0039] (iv) From idx vice Select a value from the set and let it be i v , randomly select num in the main cluster main ×r main The index of the samples Then execute A(i v ,idx′ main )=1. Repeat this step until idx vice All items in have undergone step (iv).

[0040] (v) Return to step (ii) until all clusters have been set as primary clusters.

[0041] Since the clustering results are uncertain, it is not convenient to directly analyze their time complexity. Therefore, assuming that the number of samples in each cluster is equal, the time complexity of the algorithm is approximately O(N x N n).

[0042] The smooth graph signal is subjected to node classification:

[0043] GAT is constructed by stacking several graph attention layers. The details of each layer are basically the same, and its input is the feature vector h=h1,h2,...,h of all nodes in the graph signal. N , Where N is the total number of nodes and F is the node feature dimension. First, the attention coefficient (i.e. weight) of the neighboring node is obtained using the following formula:

[0044]

[0045] where α ij represents the attention coefficient of node j to node i, is the first-order neighbor node set of node i, and the LeakyReLU function is used to introduce nonlinearity. is the weight vector of a single-layer feedforward network, is a learnable linear transformation matrix. a and W are obtained by gradient descent. T represents transposition and || represents concatenation. Then the output of a single graph attention layer is obtained using the following formula:

[0046]

[0047] in σ(·) is the activation function, and K is the number of attentions. The GAT output layer takes the average of the multi-head attentions to get the final result, that is, the working mode to which the sample belongs

[0048]

[0049] Furthermore, the present invention also proposes specific embodiments for simulation verification:

[0050] Scenario description:

[0051] Confrontation scenarios such as Figure 2 shown. Figure 2 (a) and (b) represent the pulse sequences transmitted by two MFRs during a certain beam dwell time period. Figure 2 (c) represents the interleaved pulse sequence that a certain UAV may receive, and after processing, it is obtained as follows Figure 2 (d) and (e) are signal samples. In real scenarios, these signal samples may have defects, such as pulse loss. Figure 1 (d) is shown in the filled box with slash lines; the interference pulse is as follows Figure 2 The leftmost filled box in (e) shows the measurement noise. Figure 2 (e) is shown in the box filled with horizontal lines.

[0052] Take the Mercury radar, which has five modes: search, acquisition, non-adaptive tracking, range resolution, and track-and-hold, as an example. The Mercury radar uses radar phrases to represent its operating modes. The radar phrases corresponding to each operating mode are shown in Table 2.

[0053] Table 2 MFR operating modes and their corresponding radar phrases

[0054]

[0055] As shown in Table 2, a radar phrase ph contains 4 radar words

[0056] ph=w a ,w b ,w c ,w d ,1 a,b,c,d 9 (6)

[0057] A radar word w has several pulses

[0058] w=p1,,p l ,,p L ,1 l L (7)

[0059] Since non-cooperative parties can usually only infer the intention of an MFR by analyzing its signal characteristics, the carrier frequency (RF), pulse width (PW), and pulse repetition interval (PRI) can provide a lot of key information about the MFR status. Therefore, the pulse is expressed as

[0060] p=rf,pw,pri (8)

[0061] Where rf, pw, and pri are RF, PW, and PRI respectively. The radar phrase can be rewritten as

[0062]

[0063] in Indicates the lth in the kth radar word k Since the present invention introduces the drone cluster to intercept the MFR signal, the global signal sample at the receiving end is

[0064]

[0065] in represents the mth radar phrase received by UAV n.

[0066] Finally, the present invention proposes a MFR working mode recognition method f(·), that is, a multifunctional radar working mode recognition method based on smoothing graph signals.

[0067]

[0068] in is the classification result of the mth radar phrase received by drone n.

[0069] The simulation scenario assumes the number of MFRs N r =3, number of drones N u = 100, signal reception is considered successful if the drone is within any MFR beam. This invention assumes the radar word parameters for each MFR are shown in Table 3. To reduce the chance of accidental deviation in experimental results, this invention stipulates that in each simulation, each MFR randomly selects a subrange of characteristic parameters within the maximum parameter range. The characteristic parameters refer to RF, PW, and PRI, and this subrange is used as the parameter range for that simulation. All random variables follow a uniform distribution. Furthermore, to ensure that the duration of each radar word is essentially the same, radar words w1-w3 are each 30 pulses, w4-w6 are each 40 pulses, and w7-w9 are each 50 pulses.

[0070] Table 3 MFR radar word parameter settings

[0071]

[0072] Let the number of clusters per row and column of SOM be r = c = 25. To reduce unnecessary calculations, the connection ratio of the main cluster and the neighbor cluster is r main =r vice =0.01, the number of neighbor clusters N n = 4. The GAT network structure is, in order, Dropout layer, graph attention layer (number of attention layers K = 8, output feature dimension F′ = 20), ELU, Dropout layer, graph attention layer (K ​​= 1, F′ = 15), ELU, and softmax, where the Dropout ratio is 0.2.

[0073] To simulate real-world conditions, the simulation considered three types of sample defects, as shown in Table 4. Because the PRI is the difference between the arrival times (TOA) of two adjacent pulses, the present invention influences the PRI by adding measurement noise to the TOA. The measurement noise follows a Gaussian distribution with a mean of 0 and a standard deviation as shown in Table 4.

[0074] Table 4 Parameter settings for different scenarios

[0075]

[0076] In order to discuss the recognition effect of the multi-function radar working mode recognition method based on smoothing graph signal when the supervision data is small, the number of supervision data of each category n is simulated. s=1, 2, 3, 4, 5, 10, 15, 20, 25. To discuss the impact of sample defects on recognition results, experiments were conducted on the four scenarios listed in Table 4. The number of intercepted signal samples in each simulation was approximately 3500, and the imbalance (the average number of samples in the category with the most samples to the least) was approximately 10:1. Accuracy and F1 value (commonly used in category imbalance scenarios) were used as evaluation indicators. For each of the above cases, 100 simulation experiments were conducted and the average results were taken. Figure 3 As shown in the figure, the accuracy and F1 value in each scenario increase with n. s The increase in the number of GAT training processes is due to the improvement in the stability of the GAT training process. s After reaching a certain value, the growth tends to slow down.

[0077] For scenario 1, if Figure 3 (a). When n s =15, that is, the number of supervised data accounts for about 6.43%, and the accuracy of the multi-function radar working mode recognition method based on smoothing graph signal is 96.67%, while many related methods need to take n s ≥980 to ensure that the accuracy reaches this level. In addition, when n s = 25, the F1 value of the multifunctional radar working mode recognition method based on the smoothed graph signal is 98.01%. The relevant literature simulates scenes with the same imbalance and achieves the same recognition effect, but the n is taken as s =50000. Meanwhile, there are 15 types of samples in the simulation of the present invention, while there are 6 types in the literature.

[0078] like Figure 3 As shown, as sample defects become more severe, the generated image signal and the GAT training process are increasingly disturbed, resulting in a downward trend in both accuracy and F1 value. However, even in scenario 4, the multi-function radar operating mode recognition method based on smoothed image signals still achieved an accuracy of 81.27% and an F1 value of 79.2%, demonstrating that the present invention can overcome certain sample defects.

[0079] The above disclosure is only a preferred embodiment of the present invention, and certainly cannot be used to limit the scope of the rights of the present invention. Ordinary technicians in this field can understand that all or part of the processes of the above embodiment and equivalent changes made in accordance with the claims of the present invention are still within the scope of the invention.

Claims

1. A method for identifying the working mode of a multifunctional radar based on a smoothed graph signal, characterized in that: The following steps are involved: Data preprocessing; Use SOM algorithm to cluster global signal samples; Generate smooth graph signals through the clustering results of SOM; In the process of generating smooth graph signals through SOM clustering results, based on the conditions of full connection of samples within a cluster and partial connection of samples between adjacent clusters, one cluster is selected as the main cluster in each iteration, and the samples in the sample set contained in the main cluster are fully connected until all clusters have been set as the main cluster; Performing node classification on the smooth graph signal; In the process of node classification of the smooth graph signal, the feature vectors of all nodes of the smooth graph signal are input into the graph attention network, and the output layer of the graph attention network obtains the working mode to which the sample belongs by taking the average of multi-head attention.

2. The multifunctional radar operating mode recognition method based on smoothed graph signal according to claim 1, characterized in that: During data preprocessing, the data were standardized using mean and standard deviation.

3. The multifunctional radar operating mode recognition method based on smoothed graph signal according to claim 1, characterized in that: In the process of clustering global signal samples using the SOM algorithm, the distance between the input signal sample and the center of each cluster is first calculated. Specifically, the Euclidean distance is used to calculate the distance. The cluster with the smallest distance is the winning cluster. Then, the centers of the winning cluster and its neighboring clusters are adjusted to finally complete the topological mapping of the data set.

Images