Multiple radar emitter sorting and identification method based on hierarchical clustering and time-frequency convolution

Abstract

The present invention discloses a multi-radar radiation source sorting and identification method based on hierarchical clustering and time-frequency convolution, which can solve the intra-class clustering and inter-class aliasing phenomena, and efficiently complete multi-category sorting and identification. Compared with traditional sorting algorithms, it can effectively solve the problems of intra-class feature distribution and inter-class feature aliasing, and the processing speed of the sorting algorithm is improved. In view of the problems of high overlap of inter-class feature distribution range and poor separability of original PDW features, periodic features between different categories are extracted based on slow time dimension TFA, which improves the recognition effect.

Description

Technical Field

[0001] The present invention belongs to the field of radar countermeasure technology, and in particular relates to a multi-radar radiation source sorting and identification method based on hierarchical clustering and time-frequency convolution. Background Art

[0002] Radar emitter identification is a crucial component of electronic attack and defense systems. Its key task is to separate and identify various emitter information from a stream of aliased pulses. Pulse Descriptor Words (PDWs), as readily available features, are often used for identification.

[0003] Successful sorting is a prerequisite for accurately identifying radiation sources. Especially in complex electromagnetic scenarios with multi-class signal aliasing, PDWs with adjacent arrival times may belong to different radiation sources, making it difficult to extract statistical and multi-processing domain features from the PDWs. This severely limits the range of available feature extraction methods and identification approaches. Clustering is a common radiation source sorting method that aggregates categories according to specific rules, thereby extracting and splicing single-class time series samples. However, as the electromagnetic environment becomes increasingly complex, the number of pulses surges, and intra-class clustering and inter-class aliasing become increasingly severe. Conventional clustering algorithms struggle to address these issues.

[0004] With the rapid development of artificial intelligence, PDW-based recognition methods have transitioned from template matching to machine learning classifiers, and are now transitioning to deep learning classifiers. Deep learning can further extract deep features for classification and recognition through neural networks, building on raw data or manually extracted features. Time-frequency analysis (TFA), a common non-stationary signal analysis technique, has been demonstrated by numerous researchers to offer significant advantages when combined with convolutional neural networks for radiation source identification. However, the integration of PDW and TFA remains underdeveloped. Summary of the Invention

[0005] In response to the above problems and current research trends, this paper proposes a multi-radar emitter sorting and identification method based on hierarchical clustering and time-frequency convolution, which can solve the intra-class clustering and inter-class aliasing phenomena, and efficiently complete multi-category sorting and identification.

[0006] 1. Multi-radar radiation source sorting and identification method, including:

[0007] Step S1, performing initial clustering based on the feature probability distribution function, includes:

[0008] Step S1.1, do N on the pulse flow s times uniform downsampling;

[0009] Step S1.2, fitting the characteristic distribution of the downsampled pulse stream;

[0010] Step S1.3, set the density threshold h p Divide the feature distribution and set it above the density threshold h p The sample points are clustered according to continuity;

[0011] Step S1.4: Calculate the posterior probability P of the remaining sample point i belonging to cluster c c,i , select the cluster c with the largest feature probability density as the cluster to which the sample point i belongs, and obtain the initial cluster set C0;

[0012] Step S2, merging similar clusters based on the similarity of inter-cluster feature distribution, includes:

[0013] Step S2.1, traverse the cluster set C0, and determine whether the PRI samples of cluster c are of the same type to be merged based on whether they have periodicity: if they have periodicity, proceed to step S5, otherwise proceed to the next step;

[0014] Step S2.2, calculate the feature distribution similarity matrix D between each cluster in the cluster set C0;

[0015] Step S2.3: merge the similarity matrices that are smaller than the threshold h s The merged cluster set C2 is obtained;

[0016] Step S3, performing anti-aliasing based on TOA period extraction and template matching, includes:

[0017] Step S3.1, calculate the TOA feature density r of each cluster in the merged cluster C2 c2,w , determine whether the cluster is aliased by checking whether a mutation occurs. If it is an aliased cluster, proceed to the next step. If not, proceed directly to step S5;

[0018] Step S3.2: Select a pure sample segment from the aliasing cluster and extract the periodic template of each component from it; then perform steps S3.2.1 and S3.2.2 on each aliasing sample segment in the aliasing cluster in turn:

[0019] Step S3.2.1, determining the matching starting sample point k of the aliased sample segment;

[0020] Step S3.2.2: Starting from the matching starting sample point k, perform forward and reverse matching on the aliased sample segment with the periodic template of each component. A successful match indicates that the periodic template of the current component belongs to the same category. This process is repeated until the periodic template matching of all components in the current aliased segment is completed, resulting in the anti-aliasing cluster set C3.

[0021] Step S4, constructing a training data set and training the EfficientNet network;

[0022] In step S5, the data to be identified are sorted using the methods of steps S1 to S3, and the trained EfficientNet network is applied to identify the sorted data set to obtain the radiation source type identification results of each sample.

[0023] Preferably, in step S1.2, a KDE Gaussian kernel density function is used to fit the probability density of the DOA-RF joint features of the downsampled pulse stream.

[0024] Preferably, in step S2.1, the method of determining whether the c-th cluster is a cluster of the same type to be merged based on whether the PRI samples of the c-th cluster have periodicity includes:

[0025] First, set the mutation point screening threshold Select |{Δ 2 TOA c,i}|>h cp,c of As a sequence of mutation points, periodic discrimination is equivalent to {Δ 2 TOA c,i} The arithmetic progression judgment:

[0026]

[0027] Among them, N cp,c is the number of mutation point samples, h c is the periodicity discrimination threshold, modenum(·) is the number of mode samples:

[0028]

[0029] Where [·] is an indicator function, which takes the value 1 when the internal condition is met and 0 otherwise.

[0030] Preferably, in step S3.1, the method of determining whether the cluster is aliased by checking whether a mutation occurs includes:

[0031] First, each cluster is divided into multiple overlapping windows, and the feature density of adjacent windows is calculated. and And set a threshold h d , determine whether the cluster meets the following conditions. If so, aliasing occurs; otherwise, aliasing does not occur:

[0032]

[0033] Among them, the sample density is obtained by setting the window length W in the TOA dimension. r , using unit step sliding window calculation:

[0034]

[0035] Among them, N c,w is the total number of samples in the w-th window of the c-th cluster, TOA c,i is the i-th sample point in the c-th cluster.

[0036] Preferably, the step S4 specifically includes:

[0037] Step S4.1: first, segment the data set consisting of the pure pulse stream, perform time-frequency transformation on each segment, obtain a time-frequency graph, and set its category as the label;

[0038] In step S4.2, the dataset is divided into a training set and a validation set in a ratio of 8:2. The training set is fed into the EfficientNet network for training, and the validation set is used to evaluate the training results until they meet the requirements.

[0039] The present invention has the following beneficial effects:

[0040] 1. This paper proposes a hierarchical clustering-based sorting method. Compared with traditional sorting algorithms, it can effectively solve the problems of intra-class feature distribution and inter-class feature aliasing, and the sorting algorithm processing speed is improved.

[0041] 2. This paper proposes a recognition method based on time-frequency convolution. To address the problems of high overlap and poor separability of the original PDW features, the slow-time dimension TFA is used to extract periodic features between different categories, improving the recognition effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is the overall framework diagram of the radiation source sorting and identification method based on hierarchical clustering and time-frequency convolution;

[0043] Figure 2 (a) Figure 2 (b) Figure 2 (c) Figure 2 (d) Figure 2 (e) Distribution diagrams of the five characteristics of data: PRI, RF, PW, PA and DOA;

[0044] Figure 3 (a) Figure 3 (b) Figure 3 (c) Figure 3 (d) Figure 3 (e) Figure 3 (f) Figure 3 (g) Figure 3 (h) Figure 3 (i) Figure 3 (j) Figure 3 (k) Figure 3 (l) Figure 3 (m), Figure 3 (n), Figure 3 (o), Figure 3 (p), Figure 3 (q), Figure 3 (r), Figure 3 (s), Figure 3 (t) PRI cycle remainder diagrams of each category in the training set and test set respectively;

[0045] Figures 4(a), 4(b), 4(c), 4(d), 4(e), and 4(f) are the DOA-RF distribution, t-SNE of 5-dimensional features, density clustering, clustering, merging, and anti-aliasing results, respectively;

[0046] Figure 5 PRI time-frequency diagrams of each category in the training set and test set;

[0047] Figure 6 Confusion matrix of identification results for the training set;

[0048] Figure 7(a), Figure 7(b), and Figure 7(c) are the confusion matrices of the recognition results after initial clustering, merging, and de-aliasing, respectively;

[0049] Figure 8 The following is a graph showing the changes in accuracy, Rand index, and time required for clustering under different downsampling multiples. DETAILED DESCRIPTION

[0050] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is clearly and completely described below in conjunction with the embodiments of the present application and the accompanying drawings.

[0051] Sorting part:

[0052] Applicable data: Multi-category aliased pulse streams (multi-category aliased pulse sequences), each pulse includes five features: carrier frequency (Radio Frequency, RF), pulse width (Pulse Width, PW), amplitude (Pulse Amplitude, PA), time of arrival (Time of Arrival, TOA), and direction of arrival (Direction of Arrival, DOA).

[0053] Figure 1 This is the overall framework of the radiation source sorting and identification method based on hierarchical clustering and time-frequency convolution. The overall method is divided into two parts: sorting and identification, in which sorting is further divided into three sub-parts: initial clustering, merging, and anti-aliasing:

[0054] Step S1: Use KDE (Kernel Density Estimation) method to perform initial clustering:

[0055] Step S1.1, do N on the pulse flow s Multiple uniform downsampling to reduce the amount of data to be processed;

[0056] Step S1.2, use the KDE Gaussian kernel density function to fit the probability density of the DOA-RF joint feature of the downsampled pulse stream, formula

[0057] as follows

[0058]

[0059] Where N' is the total number of samples after downsampling of the test set, h1 and h2 are the window lengths used to smooth DOA and RF respectively, and DOA i and RF i is the DOA and RF of the ith sample, K(·) is the kernel function, and the Gaussian kernel function is usually selected:

[0060]

[0061] in,

[0062] Step S1.3, set the density threshold h based on experience p Divide the feature distribution, cluster the sample points above the threshold according to continuity, and use the KDE method to calculate the 2D feature joint probability density f of each central cluster. c , the formula is as follows

[0063]

[0064] Among them, N c is the number of samples in cluster c, DOA c,i and RF c,i are the DOA and RF of the i-th sample in the c-th cluster.

[0065] Step S1.4: Calculate the posterior probability P of the remaining sample point i in the test set belonging to cluster c c,i , select the cluster c with the largest posterior probability as the cluster to which the sample point i belongs, and obtain the initial cluster set C1;

[0066] The formula for calculating the posterior probability is as follows:

[0067]

[0068] Step S2: Merge similar clusters based on KL divergence (KLD):

[0069] Step S2.1, traverse the cluster set C1, and determine whether the PRI samples of cluster c are of the same type to be merged based on whether there is periodicity. First, set the mutation point screening threshold Select |{Δ 2 TOA c,i}|>h cp,c The point set is used as the mutation point sequence, and then the mutation sequence is used to determine whether it is periodic. The specific judgment formula is as follows:

[0070] Periodicity discrimination is equivalent to {Δ 2 TOA c,i} The arithmetic progression judgment:

[0071]

[0072] Where, Δ represents the differential operation; TOA c,i represents the TOA of the i-th sample of the c-th cluster; Δ 2 TOA c,i Indicates TOA c,i The second-order difference of cp,c is the number of mutation point samples, h c is the periodicity discrimination threshold, modenum(·) is the number of mode samples: Indicates the index of the first sampling point of the cp segment in the c cluster in the original sequence; Represents an index The point set consisting of the difference of ;

[0073]

[0074] Where [·] is an indicator function, which takes the value 1 when the internal condition is met and takes the value 0 otherwise. mode(·) represents the mode;

[0075] If formula (6) is satisfied, PRI is periodic and step S5 is executed. Otherwise, the following steps are executed:

[0076] Step S2.2, calculate the DOA-PW joint distribution similarity matrix D between clusters C1;

[0077]

[0078] in,

[0079]

[0080] Step S2.3, the similarity matrix is ​​less than the threshold h s The categories will be merged to obtain the merged cluster set C2;

[0081] Step S3: perform anti-aliasing based on TOA period extraction and template matching:

[0082] Step S3.1: Perform density mutation check on each merged cluster set C2. First, each cluster is divided into multiple overlapping windows, and the density of the adjacent windows is calculated. and And set a threshold h d , determine whether the cluster is aliased:

[0083]

[0084] Among them, the sample density is obtained by setting the window length W in the TOA dimension. r , using unit step sliding window calculation:

[0085]

[0086] Among them, N c,w is the total number of samples in the w-th window of the c-th cluster, TOA c,i is the i-th sample point in the c-th cluster.

[0087] If equation (9) is satisfied, aliasing occurs. The density can be used to divide the pure segment into pure segment and the aliased segment. The low-density segment is the pure segment and the high-density segment is the aliased segment. The following steps are performed. If there is no aliasing, proceed directly to step S5.

[0088] Step S3.2: Extract each component periodic template from each pure sample segment

[0089]

[0090] in, represents the period of the p-th pure segment of the c-th cluster, is the number of cycles.

[0091] Step S3.3: perform the following processing steps on each aliasing segment in sequence.

[0092] Step S3.3.1 determines the matching starting sample point. Traverse the current aliasing segment and use each TOA value as the starting point to match the TOA forward template. The sample corresponding to the point with the minimum loss is the forward matching starting position k:

[0093]

[0094] in, N c,a Indicates the total number of samples in the a-th aliasing segment in the c-th cluster.

[0095] Step S3.3.2, from TOA c,a,k Start forward and reverse matching, and repeat steps S3.3.1-S3.3.2 until all categories of periodic template matching in the current aliasing segment are completed, and the anti-aliasing cluster set C3 is obtained;

[0096] Identification part:

[0097] Step S4: construct a training data set and train the network.

[0098] Step S4.1: Construct a clean dataset containing 5-dimensional pulse description words for network training. First, perform N rec After point sliding window processing, STFT is performed to obtain a time-frequency diagram, which is labeled as its category;

[0099] The STFT formula is as follows:

[0100]

[0101] Among them, x(n) is the discrete sample to be processed, w(n) is the window function, W stft is the window length, N fft is the number of Fourier transform points, T and F are the time and frequency sampling units respectively. The window function used in this paper is the Hanning window, and the window length is adjusted according to the periodicity of PRI.

[0102] In step S4.2, the dataset is divided into a training set and a validation set in a ratio of 8:2. The training set is fed into the EfficientNet network for training, and the validation set is used to evaluate the training results.

[0103] In step S5, the test set has the same categories as the training data, but with aliasing. After sorting, a nearly pure pulse cluster set C3 is obtained, which is input into the trained EfficientNet network for testing to obtain the radiation source category of each sample.

[0104] Example:

[0105] In this embodiment, a total of 10 radiation source signals are set and sorted and identified in sequence:

[0106] The sample distribution of the training data set and the test set is as follows Figure 2 shown.

[0107] In the training set, samples of each emitter type were collected separately, eliminating aliasing, loss, errors, and interference. Samples of each emitter type were sorted by arrival time and collected for a total of 10 seconds. The sample size for each category was approximately 120,000 to 220,000, for a total of approximately 1.4 million samples. The test set, with samples of the ten emitter types intermixed based on arrival time, was collected for a total of 16 seconds. The sample size for each category was approximately 180,000 to 350,000, for a total of approximately 2.06 million samples. Note that each TOA contains only one emitter type, corresponding to one class label.

[0108] from Figure 2As can be seen intuitively in the figure, for all five features, except for category 4, there is significant overlap between the features of each emitter in both the training and test sets, making it difficult to distinguish one emitter category from all the others. However, DOA and RF can distinguish at least two categories in both the training and test sets, demonstrating local separability, and therefore these two features were selected for classification.

[0109] The periodic difference of PRI makes it possible to purify the aliased samples. The PRI period comparison of the training dataset and the test dataset is as follows: Figure 3 shown.

[0110] The test set is sorted and identified. The algorithm simulation parameters are shown in Table 1

[0111] Table 1

[0112]

[0113] Sorting part:

[0114] The sorting accuracy is evaluated using the Rand Index, which represents the degree of overlap between two clusters:

[0115]

[0116] Among them, TP represents the number of cases where two similar sample points are in the same cluster, TN represents the number of cases where two non-similar sample points are in two clusters respectively, FP represents the number of cases where two non-similar sample points are in the same cluster, and FN represents the number of cases where two similar sample points are in two clusters respectively.

[0117] Figure 4 shows the DOA-RF distribution, t-SNE of the 5-dimensional features, density clustering, clustering, merging, and anti-aliasing results. Because the RF distribution range of category 4 is significantly different from that of other categories and is easy to sort, the clustering results for this category are not shown for the sake of aesthetic feature visualization. Figure 4(a) shows the original DOA-RF distribution. The 2-dimensional features of all categories are distributed as horizontal strips. Except for category 7, the feature distributions of each category have a certain degree of overlap. The overlap between categories 3 and 6, and between categories 2 and 10 is relatively large. In addition, there is feature clustering in categories 1, 2, 8, and 9.

[0118] Figure 4(c) is a density clustering diagram after executing step S1.2, which intuitively presents the changes in the 2D feature probability density of the test set samples. Among them, the background color represents the density change, the blue dots represent the sample points after downsampling, and the red dotted line represents the probability threshold. According to the probability threshold, the clusters are divided in the probability density contour map, and then the posterior probability of the sample points belonging to each cluster is calculated. The cluster with the maximum posterior probability is selected to complete the clustering of the sample points. The clustering result after executing step S1.4 is shown in Figure 4(d), which is divided into 11 clusters in total. Categories 1, 2, 8, and 9 with feature clustering are all mistakenly divided into two clusters, and categories 2, 3, 6, and 10 with feature aliasing are all mistakenly merged into one cluster, with a Rand index of 70.89%.

[0119] After performing the merging process in step S2.3, the 11 clusters were merged into 7 clusters, and the Rand index increased to 82%, achieving intra-cluster merging for the four categories, as shown in Figure 4(e). After performing the anti-aliasing process in step S3.4, the 7 clusters were reduced to 9 clusters, and the Rand index further increased to 92.24%, essentially achieving anti-aliasing for the four categories. The main source of sorting error is the inaccurate classification of edge aliased samples.

[0120] Identification part:

[0121] Without loss of generality, the confusion matrix and accuracy metrics are used to evaluate the recognition performance.

[0122] The time-frequency diagram obtained after executing step S4.1 is as follows Figure 5 As shown, there are obvious differences between the categories.

[0123] First, without considering the wrong sorting, the confusion matrix after executing step S4.2 is as follows Figure 6 As shown in the figure, the recognition accuracy of the time-frequency convolution is 99.21%. For categories 3 and 10 without periodicity, the recognition accuracy of the proposed method is 98.42% and 94.47%, respectively. The recognition accuracy of the remaining categories reaches 100%, demonstrating that the proposed method has excellent recognition performance for periodic sequences and also has certain recognition potential for categories without periodicity.

[0124] Combined with the proposed classification method, the recognition results are shown in Figure 7. When only KDE clustering is performed, the confusion matrix is ​​shown in Figure 7(a), and the recognition accuracy is 37.42%. We focus on the impact of inter-class feature aliasing and intra-class feature clustering on the recognition effect:

[0125] For categories 3 and 6, as well as categories 2 and 10, where aliasing is more serious, the recall rates are all less than 10%, namely 5.19%, 9.77%, 6.02% and 6.77% respectively;

[0126] For categories 1, 2, 8, and 9 where feature clustering occurs, except for category 8, the recall rates are all less than 10%, namely 0%, 6.02%, 69.62%, and 1.36%, respectively.

[0127] It can be seen from this that both feature aliasing and feature clustering will seriously deteriorate the recognition performance.

[0128] After adding merging, recognition accuracy increased by 33.94% to 71.36%, as shown in Figure 7(b). For categories 1, 2, 8, and 9, which exhibit feature clustering, recall increased to 99.45%, 48.92%, 77.77%, and 85.31%, respectively, representing increases of 39.15% to 94.26%. This demonstrates that merging can effectively improve recognition performance. However, the recall improvement for category 2 was limited due to aliasing.

[0129] After further adding anti-aliasing processing, recognition accuracy increased by another 20.22% to 91.58%, as shown in Figure 7(c). For categories 3, 6, 2, and 10 with aliasing, recall increased to 76.22%, 97.22%, 89.21%, and 93.16%, respectively, representing increases of 40.29% to 87.45%, demonstrating that anti-aliasing algorithms can also effectively improve recognition performance.

[0130] Figure 8 The impact of different downsampling factors on accuracy, Rand index, and clustering time is demonstrated. It can be seen that when downsampling is performed at a factor of 2 to 100, recognition accuracy remains around 90%, while processing time reaches a low of around 58 seconds at a factor of 50 to 100. As the downsampling factor increases, recognition accuracy rapidly deteriorates to around 40%. This is because excessive downsampling results in the loss of a large number of sample points, making it impossible to form a valid central cluster. This makes the attribution of all samples highly unreliable, leading to more complex intra-class clustering and inter-class aliasing. This, in turn, complicates subsequent merging and anti-aliasing, ultimately deteriorating recognition performance. Furthermore, the need for frequent anti-aliasing may increase processing time. Therefore, the downsampling factor should be kept to a minimum to ensure recognition performance while reducing processing time.

[0131] In summary, the multi-radar emitter sorting and identification method based on hierarchical clustering and time-frequency convolution proposed in the present invention can effectively solve the intra-class clustering and inter-class aliasing phenomena, and can efficiently complete multi-category sorting and identification.

Claims

1. A multi-radar radiation source sorting and identification method, characterized in that: include: Step S1, performing initial clustering based on the feature probability distribution function, includes: Step S1.1, do N on the pulse flow s times uniform downsampling; Step S1.2, fitting the characteristic distribution of the downsampled pulse stream; Step S1.3, set the density threshold h p Divide the feature distribution and set it above the density threshold h p The sample points are clustered according to continuity; Step S1.4: Calculate the posterior probability P of the remaining sample point i belonging to cluster c c,i , select the cluster c with the largest feature probability density as the cluster to which the sample point i belongs, and obtain the initial cluster set C0; Step S2, merging similar clusters based on the similarity of inter-cluster feature distribution, includes: Step S2.1, traverse the cluster set C0, and determine whether the PRI samples of cluster c are of the same type to be merged based on whether they have periodicity: if they have periodicity, proceed to step S5, otherwise proceed to the next step; The method of determining whether the c-th cluster is a cluster of the same type to be merged based on whether the PRI samples of the c-th cluster have periodicity includes: First, set the mutation point screening threshold Select |{Δ 2 TOA c,i }|>h cp,c of As a sequence of mutation points, periodic discrimination is equivalent to {Δ 2 TOA c,i } The arithmetic progression judgment: Among them, N c is the number of samples in cluster c; N cp,c is the number of mutation point samples, h c is the periodicity discrimination threshold, modenum(·) is the number of mode samples: Where [·] is an indicator function, which takes the value 1 when the internal condition is met and takes the value 0 otherwise; mode(·) represents the mode; Step S2.2, calculate the feature distribution similarity matrix D between each cluster in the cluster set C0; Step S2.3: merge the similarity matrices that are smaller than the threshold h s The merged cluster set C2 is obtained; Step S3, performing anti-aliasing based on TOA period extraction and template matching, includes: Step S3.1, calculate the TOA feature density of each cluster in the merged cluster C2 Determine whether the cluster is aliased by checking whether a mutation occurs. If it is an aliased cluster, proceed to the next step. If not, proceed directly to step S5. Step S3.2: Select a pure sample segment from the aliasing cluster and extract the periodic template of each component from it; then perform steps S3.2.1 and S3.2.2 on each aliasing sample segment in the aliasing cluster in turn: Step S3.2.1, determining the matching starting sample point k of the aliased sample segment; Step S3.2.2: Starting from the matching starting sample point k, perform forward and reverse matching on the aliased sample segment with the periodic template of each component. A successful match indicates that the periodic template of the current component belongs to the same category. This process is repeated until the periodic template matching of all components in the current aliased segment is completed, resulting in the anti-aliasing cluster set C3. Step S4, constructing a training data set and training the EfficientNet network; In step S5, the data to be identified are sorted using the methods of steps S1 to S3, and the trained EfficientNet network is applied to identify the sorted data set to obtain the radiation source type identification results of each sample.

2. The multi-radar radiation source sorting and identification method according to claim 1, characterized in that: In step S1.2, the KDE Gaussian kernel density function is used to fit the DOA-RF joint feature probability density of the downsampled pulse stream.

3. The multi-radar radiation source sorting and identification method according to claim 1, characterized in that: The step S4 specifically includes: Step S4.1: first, segment the data set consisting of the pure pulse stream, perform time-frequency transformation on each segment, obtain a time-frequency graph, and set its category as the label; In step S4.2, the dataset is divided into a training set and a validation set in a ratio of 8:

2. The training set is fed into the EfficientNet network for training, and the validation set is used to evaluate the training results until they meet the requirements.

Images