A radar signal synchronous distributed sorting and positioning method based on drone clusters

Abstract

The present invention belongs to the technical field of radar signal sorting, and proposes a method for synchronous distributed sorting and positioning of radar signals based on drone clusters. In order to sort and locate radar signals more accurately and efficiently and taking into account the limited computing power and storage space of a single drone, a method based on radar signal image mapping is proposed to compress the storage space occupied by radar signals, and a compressed PP‑YOLOE model and a data-parallel distributed collaborative computing method are used to reduce the storage space of the model and accelerate the calculation speed of the model. Experimental results show that the method proposed in the present invention is more accurate in sorting and positioning radar signals, and compared with other methods, the present invention has low memory usage and fast running speed. The present invention provides a new method for radar signal sorting and positioning, which provides a new solution for radar signal sorting and positioning in a drone cluster environment.

Description

Technical Field

[0001] The present invention relates to the field of radar signal sorting technology, and in particular to a radar signal synchronous distributed sorting and positioning method based on an unmanned aerial vehicle cluster. Background Art

[0002] With the rapid increase in the number of various electronic countermeasure devices, the signal density of the electromagnetic threat environment has reached the order of millions. Modern radars are developing towards multi-functional and multi-purpose directions, and their waveforms are becoming more complex and their regularity is gradually decreasing. These have put forward higher requirements on the accuracy, real-time and reliability of sorting and positioning methods. Traditional methods use structured data learning features to sort and locate radar emitters. Clustering and random forests are commonly used algorithms in the study of machine learning sorting algorithms. Wang et al. published a work titled "Multiple-parameter radar signal sorting using support vector clustering and similitude entropy index" in Circuits, Systems, and Signal Processing in 2014. It was found that the traditional support vector clustering (SVC) based on radar signal sorting methods has a high time complexity, and the disadvantage of traditional effectiveness indicators is that they cannot effectively represent the best sorting. To this end, the authors proposed a new cone cluster labeling (CCL) based on the sorting method and a new index (similitude entropy, SE). The random selection of the initial clustering center makes it easier to fall into the local optimal clustering, which is not suitable for non-spherical data. To address this problem, Feng et al. proposed a K-means clustering radar signal sorting algorithm based on data field in their work "Radar signal sorting algorithm of k-means clustering based on data field" published in ICCC in 2017. This algorithm can automatically obtain the initial cluster center and cluster number. However, the K-Means algorithm is sensitive to noise and requires the number of clusters to be manually set, which is a bottleneck that is difficult to break through in terms of accuracy. In recent years, many heuristic optimization algorithms have emerged. Reddy et al.'s work "Optimization of k-means algorithm: Ant colony optimization" published in ICCMC in 2017 uses the ant colony algorithm to optimize the K-Means algorithm. Li et al.'s work "A new clustering and sorting algorithm for radar emitter signals" published in Journal of Physics: Conference Series in 2020 proposed an improved K-Means clustering algorithm based on data domain. Since the information representation of structured data is fixed and limited, its accuracy is difficult to improve. In addition, traditional methods can only focus on classification or clustering tasks, and therefore cannot simultaneously sort and locate radar signals.Therefore, it is necessary to design an efficient synchronous signal sorting and positioning method in detail to meet the needs of high-precision signal sorting and positioning. Summary of the invention

[0003] In view of some shortcomings of existing research, the present invention proposes a method for synchronous distributed sorting and positioning of radar signals based on drone clusters, which uses drone clusters, radar signal image mapping, target detection synchronous sorting and positioning, and data parallel distributed collaborative computing to sort and locate radar signals. This method takes into account the limited computing power and storage capacity of a single drone and the number advantage of drones in a cluster environment. It is adapted to the drone cluster environment in terms of data and model, achieves high-precision radar signal sorting and positioning, and further utilizes the drone cluster environment to improve the sorting efficiency; maps structured radar signal data into image radar signal data, uses a compressed PP-YOLOE model to sort and locate image radar signal data, and uses data parallel distributed collaborative computing methods to accelerate the model's computing speed. The present invention provides a new solution to the problem of high-precision sorting and positioning of radar signals in a drone cluster environment.

[0004] The technical solution of the present invention:

[0005] A method for synchronous distributed sorting and positioning of radar signals based on drone clusters, the steps are as follows:

[0006] Step 1: Randomly determine a drone head node in the drone cluster environment, and the other drone nodes are computing nodes;

[0007] Step 2: Use the computing node constructed in step 1 to collect radar signal structured data and send it to the drone head node; the drone head node receives and caches the radar signal structured data sent by the computing node, and converts it into image radar signal data after a certain time interval; the drone head node distributes the image radar signal data to the computing node; the computing node discards the previously collected radar signal structured data that grows linearly with time, and saves the image radar signal data;

[0008] Step 3, the drone head node broadcasts the compressed PP-YOLOE model structure (an improved single-stage Anchor-free target detection model based on the YOLO model) and the initial weights of the PP-YOLOE model to the computing node, and the computing node calculates the forward propagation loss value based on the imaged radar signal data in step 2 and the PP-YOLOE model structure, and returns the back-propagation gradient value to the drone head node; the drone head node integrates the loss gradient value to optimize the PP-YOLOE model weight, and then distributes the optimized PP-YOLOE model weight to each computing node; the computing node iteratively calculates the forward propagation loss value and returns the loss gradient value, and the drone head node iteratively integrates the loss gradient and optimizes the PP-YOLOE model until the PP-YOLOE model converges;

[0009] Step 4: The drone head node distributes the converged PP-YOLOE model obtained in step 3 to the computing node; the computing node uses the converged PP-YOLOE model to sort and locate the latest collected radar signals to obtain the category of the radar signal and the location of the radar signal radiation source.

[0010] The step 2 is specifically as follows:

[0011] 2.1) The computing node collects radar signal structured data through the radar signal receiver; the radar signal structured data includes five parameters: arrival time TOA, arrival direction DOA, pulse repetition interval PRI, radio frequency RF and pulse width PW; the distance between the radar signal radiation source and the radar signal receiver is calculated by the arrival time, and the arrival time and arrival direction are used as R and θ parameters in the polar coordinate system to locate the radar signal radiation source; the pulse repetition interval, radio frequency and pulse width are used to sort the radar signal; after collecting the radar signal structured data, the computing node sends it to the drone head node;

[0012] 2.2) After a certain time interval, the drone head node caches a certain amount of radar signal structured data, and then uses the radar signal image mapping method to map it into image radar signal data; a single radar signal structured data (DOA, TOA, PRI, RF, PW) is represented as a vector x∈R 5 , because in unit time t, when the radar signal receiver captures m radar signals, the radar signal matrix is ​​X∈R m×5 To present these series of radar signals; in addition, use; superscript x (i) To represent the i-th sample in the matrix X, the subscript x j to represent the jth eigenvector x in the matrix; the final radar signal matrix is ​​expressed as

[0013] X=(x (1) , x(2) , ..., x (i) , ..., x (m) ),

[0014] in The radar signal matrix X is equivalent to the structured data table of the radar signal in unit time t;

[0015] Then, the five parameters of radar signal structured data, DOA, TOA, PRI, RF, and PW, are mapped into five image parameters, width, height, red, green, and blue for image representation; for each vector x (i) , which is mapped to a pixel in the image and is represented as

[0016]

[0017]

[0018]

[0019] Therefore, the entire structured data table, matrix X, is mapped into an image matrix I

[0020] I=(p (1) , p (2) , ..., p (i) , ..., p (m) ),

[0021] The mapping methods include normalization and inverse normalization; normalization is expressed as

[0022]

[0023] Where Min(X) is a matrix, each column of which is the minimum constant value within the range of the radar signal matrix X; Max(X) is a matrix, each column of which is the maximum constant value within the range of the radar signal matrix X; inverse normalization is the inverse function of normalization, denoted as

[0024] N -1 (X, I)=X⊙(Max(I)-Min(I))+Min(I),

[0025] where ⊙ is element-wise matrix multiplication;

[0026] Finally, combining the normalization and inverse normalization equations, the mapping function is

[0027] M(X, I) = N -1 (N(X), I),

[0028] 2.3) The drone head node distributes the mapped image radar signal data to each computing node to ensure the consistency of the training data in the subsequent distributed collaborative computing.

[0029] The step 3 is specifically as follows:

[0030] 3.1) After each computing node has received consistent training data, the drone head node distributes the compressed PP-YOLOE model structure and the initial weights of the PP-YOLOE model to the computing nodes to start the distributed collaborative computing process;

[0031] 3.2) The computing node randomly selects a batch of n training data and uses binary cross entropy to calculate the loss value of the forward propagation. The formula is as follows

[0032]

[0033] Among them, X is the radar signal data, y is the type corresponding to the radar signal data, W is the weight that the model needs to learn, h(X) is the mapping value of input X through the nonlinear activation function, and λ is the regularization hyperparameter; after obtaining a batch of loss Loss(X, y, W), the computing node uses the chain rule to calculate the gradient value of back propagation and returns it to the drone head node;

[0034] 3.3) After the drone head node receives the loss gradient of each computing node, it uses the formula

[0035]

[0036] Integrate the gradient loss values ​​of a batch, and then use the stochastic gradient descent algorithm to optimize the PP-YOLOE model weights of the batch;

[0037]

[0038] Among them, η is the learning rate of the optimization algorithm;

[0039] 3.4) The computing node iteratively calculates the forward propagation loss value and returns the gradient value of the loss. The head node iteratively integrates the loss gradient and optimizes the model until the PP-YOLOE model converges.

[0040] The step 4 is specifically as follows:

[0041] 4.1) The head node broadcasts the trained converged model to the computing nodes. The head node distributes the trained model to the computing nodes once and uses the trained converged model for prediction. There is no need to iteratively update the model as in training.

[0042] 4.2) The computing node sorts and locates the newly collected radar signals according to the received converged model structure and model weights. Since the model weights need to be updated iteratively during training, the communication loss is large, so the distributed expansion can reduce the training time nonlinearly; during prediction, since only one communication is required to obtain the trained model weights, the distributed expansion can nearly linearly improve the reasoning speed.

[0043] The generation process of the radar signal structured data is as follows;

[0044] First, simulate the data of signal transmitters; set the number of signal transmitters in each scene from 1 to 7, with an average of 4;

[0045] Randomly determine the number and type of simulated signals, generate the arrival direction and arrival time of each signal transmitter between 0 and 1, and generate the pulse repetition interval, radio frequency and pulse width in the parameter interval corresponding to the signal type;

[0046] Then the error value is added to the data of the signal transmitter, and each signal transmitter data is expanded into a plurality of signal data of the actual received data;

[0047] Error e error A normal distribution with mean 0 and standard deviation σ of random values ​​in the range of 0.02 to 0.05:

[0048]

[0049] where e error ∈R 5 , represents the error value of arrival time, arrival direction, pulse repetition interval, radio frequency and pulse width; assuming that the radar transmitter signal vector Generate a vector of radar data:

[0050]

[0051] The simulated data are all the signals received by the drone swarm in unit time. The task is to find the actual signal transmitter and identify its type based on these signals.

[0052] The first two parameters of the radar signal structured data, the arrival direction and the arrival time, are used as location information to identify the geographical location of the radar signal radiation source, and the remaining three parameters, the pulse repetition interval, the radio frequency and the pulse width, are used to distinguish the type of the signal; the five parameters are generated into normalized values ​​of 0 to 1; the signals are divided into different types according to the parameter pulse repetition interval, the radio frequency and the pulse width values; for each parameter of different types, the length of its range is 0.4, but the left and right boundaries are different.

[0053] For the parameters of the five types of signals, the left boundary is selected from 0, 0.15, 0.3, 0.45 and 0.6; correspondingly, the right boundary is selected from 0.4, 0.55, 0.7, 0.85 and 1.0; in this case, the overlapping length of the value ranges of two adjacent intervals is 0.25, accounting for 62.5% of a single interval.

[0054] In the drone cluster environment, a drone head node is determined, and other drone nodes are called computing nodes. Because the model and data are adapted to the drone cluster environment, there is no significant difference in the computing pressure of the head node and the computing node, so the identities of the head node and the computing node are interchangeable. This feature can effectively avoid the problem of single point failure of the head node.

[0055] Beneficial effects of the present invention: The present invention proposes a method for synchronous distributed sorting and positioning of radar signals based on drone clusters and PP-YOLOE models. In order to sort and locate radar signals more accurately and efficiently and taking into account the limited computing power and storage space of a single drone, the present invention also proposes a method based on radar signal image mapping to compress the storage space occupied by radar signals, and uses a compressed PP-YOLOE model and a data-parallel distributed collaborative computing method to reduce the storage space of the model and accelerate the calculation speed of the model. Experimental results show that the method proposed in the present invention is more accurate in sorting and positioning radar signals, and compared with other methods, the present invention has low memory usage and fast running speed. The present invention provides a new method for radar signal sorting and positioning, which provides a new solution for radar signal sorting and positioning in a drone cluster environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a framework diagram of the radar signal synchronous distributed sorting and positioning method. The left side of the drone head node is the data collection and image mapping part, and the right side of the drone head node is the distributed collaborative computing part.

[0057] Figure 2 These are example diagrams of converting structured radar signal data into imaged radar signals. In (a)-(d), the density of radar signals increases successively, and the storage space saving rate also increases successively.

[0058] Figure 3 The following are comparison charts of different model algorithms applied in the framework. (a) is a comparison chart of the model's F-score and inference speed; (b) is a comparison chart of the model size and training time. Tradition is a combination of two traditional algorithms: Logistics Regression and Mean-shift.

[0059] Figure 4The following are the compression comparison charts of the PP-YOLOE model. (a) is the comparison chart of the F-score and inference speed of the model; (b) is the comparison chart of the model size and training time.

[0060] Figure 5 The following are the comparison charts of the distributed expansion of the Micro PP-YOLOE model. (a) is the comparison chart of the F-score and inference speed of the model; (b) is the comparison chart of the model size and training time.

[0061] Figure 6 The simulated detection diagrams of the image radar signal, (a)-(d) are Figure 2 (a)-(d) are the results of simulation detection corresponding to the figure; the additional boxes, circles and marks in the simulation detection figure compared to the example figure are all generated by model prediction. The type of radar signal is marked above the box in the figure, and the white circle in the middle of the box marks the location of the radar signal radiation source. DETAILED DESCRIPTION

[0062] In order to make the objectives, technical solutions and advantages of the present invention more clear, the specific implementation modes of the present invention will be further described in detail below.

[0063] The present invention provides a method for synchronous distributed sorting and positioning of radar signals based on a drone cluster, the method comprising:

[0064] Step 1):

[0065] In the drone cluster environment, a drone head node is determined, and other drone nodes are called drone computing nodes. Because the model and data are adapted to the drone cluster environment, there is no significant difference in the computing pressure of the drone head node and the drone computing node, so the identities of the drone head node and the drone computing node are interchangeable. This feature can effectively avoid the problem of single point failure of the drone head node.

[0066] Step 2):

[0067] The drone computing nodes constructed in step 1) are used to collect radar signal data and send them to the drone head nodes. The drone head nodes receive and cache the structured data sent by the computing nodes, and convert them into image data after a certain time interval. After that, the drone head nodes distribute these image data to the drone computing nodes. The drone computing nodes discard the previously collected structured data that grows linearly with time and save the image data.

[0068] The experiment produced a radar signal simulation data set for sorting and positioning experiments. The first two parameters of the radar signal, DOA and TOA, were used as location information to identify the geographical location of the signal source, and the remaining three parameters, PRI, PW, and RF, were used to distinguish the types of signals. In order to facilitate subsequent calculations and processing, these parameters were normalized to values ​​ranging from 0 to 1. The signals were divided into five categories according to the values ​​of the parameters PRI, PW, and RF. For each parameter of different types, the length of its range is 0.4, but the left and right boundaries are different. For the parameters of these five signals, their left boundaries can be selected from 0, 0.15, 0.3, 0.45, and 0.6. Correspondingly, their right boundaries can be selected from 0.4, 0.55, 0.7, 0.85, and 1.0. In this case, the overlapping length of the value ranges of two adjacent intervals is 0.25, accounting for 62.5% of a single interval, which increases the complexity of the classification task. In order to generate structured data for radar signals, the data of the signal transmitter is simulated first. The number of transmitters in each scene is set to 1 to 7, with an average of 4. The number and type of signals to be simulated are randomly determined, the DOA and TOA of each transmitter are generated between 0 and 1, and the remaining three parameters are generated in the parameter interval corresponding to the signal type. The error value is then added to the data, and each transmitter data is expanded to multiple signal data that is closer to the actual received data. Error e error A normal distribution with mean 0 and standard deviation σ of random values ​​in the range of 0.02 to 0.05:

[0069]

[0070] where e error ∈R 5 , represents the error value of five parameters. Assuming the radar transmitter signal vector A vector of radar data can be generated:

[0071]

[0072] The simulated data are all the signals received by the drone swarm in unit time. The task is to find the actual signal transmitter and identify its type based on these signals.

[0073] 2.1) The computing node collects radar signal structured data through the radar signal receiver. Radar signal structured data is represented by radar pulse descriptor word (PDW), which includes five main parameters: time of arrival (TOA), direction of arrival (DOA), pulse repetition interval (PRI), radio frequency (RF) and pulse width (PW). The distance between the radar signal radiator and the drone radar signal receiver can be calculated through TOA, so TOA and DOA can be regarded as R and θ parameters in the polar coordinate system to locate the radar signal radiator; PRI, RF and PW are used to sort the radar signal. After collecting the PDW data, the computing node sends it to the head node.

[0074] 2.2) After a certain time interval, that is, after a certain amount of PDW structured data is cached, the drone head node uses the radar signal image mapping method to map it into image radar signal data. A single radar signal (DOA, TOA, PRI, RF, PW) is represented as a vector x∈R 5 , so in unit time t, if the radar receiver captures m radar signals, the radar signal matrix X∈R m×5 To represent these series of radar signals, we use the superscript x (i) To denote the i-th sample in the matrix X, we use the subscript x j To represent the jth eigenvector x in the matrix. Therefore, the radar signal matrix is ​​expressed as

[0075] X=(x (1) , x (2) , ..., x (i) , ..., x (m) ),

[0076] in The radar signal matrix X is equivalent to the structured data table of the radar signal in unit time t.

[0077] Then the radar signal parameters (DOA, TOA, PRI, RF, PW) are mapped into image parameters (width, height, red, green, blue) for image representation. (i) , which is mapped to a pixel in the image and is represented as

[0078]

[0079]

[0080]

[0081] Therefore, the entire structured data table, matrix X, can be mapped into an image matrix I

[0082] I=(p(1) , p (2) , ..., p (i) , ..., p (m) ).

[0083] The mapping methods include normalization and inverse normalization. Normalization is expressed as

[0084]

[0085] Where Min(X) is a matrix whose columns are the minimum constant values ​​in the original matrix X; similarly, Max(X) is a matrix whose columns are the maximum constant values ​​in the original matrix X. Inverse normalization is the inverse function of normalization, denoted by

[0086] N -1 (X, I)=X⊙(Max(I)-Min(I))+Min(I),

[0087] Where ⊙ is the element-wise matrix multiplication, and I is the matrix data domain corresponding to the image after mapping.

[0088] Finally, combining the normalization and inverse normalization equations, we can get the mapping function as

[0089] M(X, I) = N -1 (N(X), I).

[0090] 2.3) The drone head node distributes the mapped image radar data to each drone computing node to ensure the consistency of the training data in the subsequent distributed collaborative computing.

[0091] Step 3):

[0092] The drone head node broadcasts the compressed PP-YOLOE model structure and initial model weights to the drone computing node. The computing node calculates the forward propagation loss value based on the imaged radar data generated in step 2) and the model structure, and returns the back-propagation gradient value to the drone head node. The drone head node integrates the gradient value of the loss to optimize the model weights, and then distributes the optimized model weights to each drone computing node. The drone computing node iteratively calculates the forward propagation loss value and returns the gradient value of the loss. The drone head node iteratively integrates the loss gradient and optimizes the model until the model converges;

[0093] 3.1) After ensuring that each UAV computing node has received consistent training data, the UAV head node distributes the compressed PP-YOLOE model structure and initialized weights to the UAV computing nodes to start the distributed collaborative computing process.

[0094] 3.2) The computing node randomly selects a batch size of n training data and uses binary cross entropy to calculate the loss value of the forward propagation. The formula is as follows

[0095]

[0096] After obtaining a batch of losses Loss(X, y, W), the drone computing node uses the chain rule to calculate the gradient value of back propagation And return it to the drone head node.

[0097] 3.3) After the drone head node receives the loss gradient of each drone computing node, it uses the formula

[0098]

[0099] Integrate the gradient loss values ​​of a batch, and then use the stochastic gradient descent algorithm to optimize the model weights of this batch.

[0100]

[0101] 3.4) After that, the drone computing node iteratively calculates the forward propagation loss value and returns the gradient value of the loss. The drone head node iteratively integrates the loss gradient and optimizes the model until the model converges.

[0102] Step 4):

[0103] The head node distributes the convergence model obtained in step 3) to the UAV computing node. The UAV computing node uses the convergence model to sort and locate the latest collected radar signals to obtain the category of the radar signal and the location of the radar signal radiation source.

[0104] 4.1) The drone head node broadcasts the trained converged model to the drone computing node. When using the trained model for prediction, the drone head node only needs to distribute the trained model to the drone computing node once, without the need to iteratively update the model as during training;

[0105] 4.2) The drone computing node sorts and locates the newly collected radar signals according to the received model structure and model weights. Since the model weights need to be updated iteratively during training, the communication loss is large, so the distributed expansion can reduce the training time nonlinearly; during prediction, since only one communication is required to obtain the trained model weights, the distributed expansion can improve the reasoning speed nearly linearly.

[0106] Figure 3 , 4 , 5 are comparison diagrams of different model algorithms. The markers of each model in the figure have four sizes, corresponding to Figure 2There are four different densities. The larger the mark, the greater the density. The dotted line in the figure is the average performance of the model under the four densities. Figure 3 It can be seen that PP-YOLOE has the highest accuracy, and other indicators are between the other models. Figure 4 The number of layers and channels of the Mini PP-YOLOE model are 67% and 75% of the baseline PP-YOLOE model, while Micro PP-YOLOE is compressed to 33% and 50%. As can be seen from the figure, with almost no loss of accuracy, the other three indicators of the Micro PP-YOLOE model have been greatly improved. Figure 5 In the case of no significant loss in accuracy, distributed expansion increases the inference speed nearly linearly and reduces the training time nonlinearly.

Claims

1. A method for synchronous distributed sorting and positioning of radar signals based on drone clusters, characterized in that: Here are the steps: Step 1: Randomly determine a drone head node in the drone cluster environment, and the other drone nodes are computing nodes; Step 2: Use the computing node constructed in step 1 to collect radar signal structured data and send it to the drone head node; the drone head node receives and caches the radar signal structured data sent by the computing node, and converts it into image radar signal data after a certain time interval; The drone head node distributes these imaged radar signal data to the computing node; the computing node discards the previously collected radar signal structured data that grows linearly over time and saves the imaged radar signal data; Step 3: The drone head node broadcasts the compressed PP-YOLOE model structure and the initial weights of the PP-YOLOE model to the computing node. The computing node calculates the forward propagation loss value based on the imaged radar signal data in step 2 and the PP-YOLOE model structure, and returns the backward propagation gradient value to the drone head node. The drone head node integrates the gradient value of the loss to optimize the PP-YOLOE model weights, and then distributes the optimized PP-YOLOE model weights to each computing node; The computing node iteratively calculates the loss value of the forward propagation and returns the gradient value of the loss. The drone head node iteratively integrates the loss gradient and optimizes the PP-YOLOE model until the PP-YOLOE model converges; Step 4: The drone head node distributes the converged PP-YOLOE model obtained in step 3 to the computing node; the computing node uses the converged PP-YOLOE model to sort and locate the latest collected radar signals to obtain the category of the radar signal and the location of the radar signal radiation source; The step 2 is specifically as follows: 2.1) The computing node collects radar signal structured data through the radar signal receiver; the radar signal structured data includes five parameters: arrival time TOA, arrival direction DOA, pulse repetition interval PRI, radio frequency RF and pulse width PW; the distance between the radar signal radiation source and the radar signal receiver is calculated by the arrival time, and the arrival time and arrival direction are used as R and θ parameters in the polar coordinate system to locate the radar signal radiation source; the pulse repetition interval, radio frequency and pulse width are used to sort the radar signal; after collecting the radar signal structured data, the computing node sends it to the drone head node; 2.2) After a certain time interval, the drone head node caches a certain amount of radar signal structured data, and then uses the radar signal image mapping method to map it into image radar signal data; a single radar signal structured data (DOA, TOA, PRI, RF, PW) is represented as a vector x∈R 5 , because in unit time t, when the radar signal receiver captures m radar signals, the radar signal matrix is ​​X∈R m×5 To present these series of radar signals; in addition, use; superscript x (i) To represent the i-th sample in the matrix X, the subscript x j to represent the jth eigenvector x in the matrix; the final radar signal matrix is ​​expressed as X=(x (1) ,x (2) ,...,x (i) ,...,x (m) ), in The radar signal matrix X is equivalent to the structured data table of the radar signal in unit time t; Then, the five parameters of radar signal structured data, DOA, TOA, PRI, RF, and PW, are mapped into five image parameters, width, height, red, green, and blue for image representation; for each vector x (i) , which is mapped to a pixel in the image and is represented as Therefore, the entire structured data table, matrix X, is mapped into an image matrix I I=(p (1) ,p (2) ,...,p (i) ,...,p (m) ), Mapping methods include normalization and inverse normalization; Normalized expression is Where Min(X) is a matrix, each column of which is the minimum constant value within the range of the radar signal matrix X; Max(X) is a matrix, each column of which is the maximum constant value within the range of the radar signal matrix X; inverse normalization is the inverse function of normalization, denoted as N -1 (X,I)=X⊙(Max(I)-Min(I))+Min(I), where ⊙ is element-wise matrix multiplication; Finally, combining the normalization and inverse normalization equations, the mapping function is M(X,I)=N -1 (N(X),I), 2.3) The drone head node distributes the mapped image radar signal data to each computing node to ensure the consistency of the training data in the subsequent distributed collaborative computing.

2. The method for synchronous distributed sorting and positioning of radar signals based on drone clusters according to claim 1 is characterized in that: The step 3 is specifically as follows: 3.1) After each computing node has received consistent training data, the drone head node distributes the compressed PP-YOLOE model structure and the initial weights of the PP-YOLOE model to the computing nodes to start the distributed collaborative computing process; 3.2) The computing node randomly selects a batch of n training data and uses binary cross entropy to calculate the loss value of the forward propagation. The formula is as follows Among them, X is the radar signal data, y is the type corresponding to the radar signal data, W is the weight that the model needs to learn, h(X) is the mapping value of input X through the nonlinear activation function, and λ is the regularization hyperparameter; after obtaining a batch of loss Loss(X,y,W), the computing node uses the chain rule to calculate the gradient value of back propagation and returns it to the drone head node; 3.3) After the drone head node receives the loss gradient of each computing node, it uses the formula Integrate the gradient loss values ​​of a batch, and then use the stochastic gradient descent algorithm to optimize the PP-YOLOE model weights of the batch; Among them, η is the learning rate of the optimization algorithm; 3.4) The computing node iteratively calculates the forward propagation loss value and returns the gradient value of the loss. The head node iteratively integrates the loss gradient and optimizes the model until the PP-YOLOE model converges.

3. The method for synchronous distributed sorting and positioning of radar signals based on drone clusters according to claim 1 or 2, characterized in that: The step 4 is specifically as follows: 4.1) The head node broadcasts the trained converged model to the computing nodes; the head node distributes the trained model to the computing nodes once and uses the trained converged model for prediction; 4.2) The computing node sorts and locates the newly collected radar signals according to the received converged model structure and model weights.

4. The method for synchronous distributed sorting and positioning of radar signals based on drone clusters according to claim 1 is characterized in that: The generation process of the radar signal structured data is as follows; First, simulate the data of signal transmitters; set the number of signal transmitters in each scene from 1 to 7, with an average of 4; Randomly determine the number and type of simulated signals, generate the arrival direction and arrival time of each signal transmitter between 0 and 1, and generate the pulse repetition interval, radio frequency and pulse width in the parameter interval corresponding to the signal type; Then the error value is added to the data of the signal transmitter, and each signal transmitter data is expanded into a plurality of signal data of the actual received data; Error e error A normal distribution with mean 0 and standard deviation σ of random values ​​in the range of 0.02 to 0.05: where e error ∈R 5 , represents the error value of arrival time, arrival direction, pulse repetition interval, radio frequency and pulse width; assuming that the radar transmitter signal vector Generate a vector of radar data: The simulated data are all the signals received by the drone swarm in unit time. The task is to find the actual signal transmitter and identify its type based on these signals.

5. The method for synchronous distributed sorting and positioning of radar signals based on drone clusters according to claim 1 or 2, characterized in that: The first two parameters of the radar signal structured data, the arrival direction and the arrival time, are used as location information to identify the geographical location of the radar signal radiation source, and the remaining three parameters, the pulse repetition interval, the radio frequency and the pulse width, are used to distinguish the type of signal; The five parameters are normalized to values ​​ranging from 0 to 1. The signals are classified into different types according to the values ​​of the parameters pulse repetition interval, radio frequency and pulse width. For each parameter of different types, the length of its range is 0.4, but the left and right boundaries are different.

6. The method for synchronous distributed sorting and positioning of radar signals based on drone clusters according to claim 5 is characterized in that: For the five parameters, the left boundary is selected from 0, 0.15, 0.3, 0.45 and 0.6; correspondingly, the right boundary is selected from 0.4, 0.55, 0.7, 0.85 and 1.0; in this case, the overlapping length of the value ranges of two adjacent intervals is 0.25, accounting for 62.5% of a single interval.

Images