A direction finding method and system based on a UAV cluster under a multi-factor error condition

Abstract

The application discloses a direction finding method based on a UAV cluster under a multi-factor error condition, and comprises the following steps: S10, in a laboratory, a network model considering four errors is reconstructed by considering position errors, radio frequency channel amplitude and phase errors, nonlinear processing errors and coupling interference errors respectively through a control variable method; S20, a UAV cluster carrying an electronic countermeasure reconnaissance load is deployed, signals from each target radiation source are perceived and received, and data are recorded; S30, the network model of S10 is used to correct errors of the data received by S20; S40, after the error correction data of S30 are received, a virtual-real separation is performed, and a neural network is reentered; S50, data in the neural network sequentially undergo six times of processing to obtain an output function, a complex function is obtained by combining real and imaginary parts of the output function, the complex function is multiplied by a steering vector to obtain a spatial spectrum, and each radiation source is direction found according to the spatial spectrum; and a corresponding system is also disclosed. The application improves the accuracy, reliability and flexibility of electronic countermeasure reconnaissance and simplifies a direction finding process.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electronic countermeasures, more particularly, to a direction finding method and system based on a UAV cluster under multi-factor error conditions. BACKGROUND

[0002] With the rapid development of modern electronic information technology, the complexity of the electromagnetic environment is increasing day by day, and electromagnetic situation awareness has become one of the key technologies to ensure information security and implement electronic countermeasures. Among them, the accurate direction finding and positioning of the target radiation source is not only an important part of electromagnetic situation awareness, but also a prerequisite for implementing precise strikes, electronic reconnaissance and anti-reconnaissance and other military operations. The traditional ground electronic countermeasure reconnaissance means is difficult to meet the needs of complex and variable battlefield environments due to factors such as terrain shielding and line-of-sight distance limitations.

[0003] In order to solve these problems, some existing technologies propose a UAV-borne direction finding de-aliasing method based on an interferometer, which effectively improves the direction finding accuracy and stability by optimizing the interferometer direction finding system. However, it cannot realize the simultaneous direction finding of multiple radiation sources. Some existing technologies realize direction finding based on a UAV cluster and an intelligent metasurface, which innovates the atomic norm direction finding method based on the compressed sensing sparse reconstruction system. However, it does not consider the influence of actual errors. Some existing technologies combine neural networks with direction finding methods, consider the existence of errors, and use network training methods to fit functions to eliminate the influence of errors. However, the error model is relatively single and has a gap with the actual situation; the input and output of the neural network are the measured values and the actual values respectively, and the characteristic information of the original data is not fully utilized. SUMMARY

[0004] In view of the above defects or improvement needs of the prior art, the present application provides a direction finding method and system based on a UAV cluster under multi-factor error conditions, which comprehensively considers the UAV position error, the radio frequency channel amplitude and phase inconsistency error, the non-linear processing error, and the coupling interference error between the reconnaissance loads of each UAV, and constructs a more accurate and practical network model; the accuracy and reliability of the UAV cluster in electronic countermeasure reconnaissance are significantly improved.

[0005] In order to achieve the above purpose, according to the first aspect of the present application, a direction finding method based on a UAV cluster under multi-factor error conditions is provided, comprising the following steps:

[0006] S10, network training, in the ideal state of the laboratory, considering the position error of the UAV, the radio frequency channel amplitude and phase inconsistency error, the non-linear processing error, and the coupling interference error between the reconnaissance loads of each UAV, a network model considering the four kinds of errors is constructed;

[0007] S20, data acquisition, deploy the unmanned aerial vehicle cluster and make it carry electronic countermeasure reconnaissance payload; the unmanned aerial vehicle senses and receives signals from each target radiation source while flying in the air, and records data received by the electronic countermeasure reconnaissance payload of each unmanned aerial vehicle;

[0008] S30, data correction, using the network model in S10 to correct errors of the data received in S20;

[0009] S40, virtual-real separation processing, after the ground station receives the error-corrected data in S30, the virtual-real separation is performed and the data is sent into the neural network;

[0010] S50, neural network processing, the data in the neural network sequentially passes through the full connection layer, tensor deformation, convolution layer, batch normalization, activation function and full connection layer to obtain the output function, the virtual part and the real part of the output function are combined to obtain the complex function, and the spatial spectrum is obtained by multiplying the complex function and the steering vector, and finally the direction finding of each radiation source is realized by peak searching according to the spatial spectrum.

[0011] Further, the method for constructing the network model considering four kinds of errors in S10 is:

[0012] S11, network training is performed in the laboratory without considering the position error of the unmanned aerial vehicle, the amplitude-phase inconsistency error, the nonlinear processing error and the coupling interference error between the reconnaissance payloads of each unmanned aerial vehicle, and ideal state data is received;

[0013] S12, training is performed only considering the position error of the unmanned aerial vehicle in S11, assuming that the position error obeys Gaussian distribution, the average value is 0, and the position error standard deviation σ per obeys uniform distribution, and the maximum value is σ per-max ;

[0014] S13, training is performed only considering the amplitude-phase inconsistency error in S11, assuming that the amplitude-phase inconsistency error obeys Gaussian distribution, the average value is 0, the amplitude standard deviation of the radio frequency channel is σ gain , and the phase standard deviation of the radio frequency channel is σ phase , the amplitude standard deviation of the radio frequency channel and the phase standard deviation of the radio frequency channel obey uniform distribution, and the maximum values are σ gain-max and σ phase-max , respectively;

[0015] S14, training is performed only considering the nonlinear processing error in S11, and the nonlinear error is described by tanh(σ nl ), wherein σ nl is the nonlinear error coefficient;

[0016] S15, training is performed only considering the coupling interference error in S11, and the coupling coefficient Cn，n′ Subject to uniform distribution, the maximum value of coupling coefficient is C max ;

[0017] S16, the training of considering the position error in S12, the amplitude and phase inconsistency error in S13, the non-linear processing error in S14 and the coupling interference error in S15 is carried out, and a network model considering the four kinds of errors is constructed.

[0018] Further, the method for recording the data received by each unmanned aerial vehicle electronic countermeasure reconnaissance payload in S20 is:

[0019] There are K radiation sources, and the direction of the incoming wave of each radiation source is θ k , wherein k=0, 1, …, K-1, k is the radiation source number, and the unmanned aerial vehicle has N unmanned aerial vehicles, taking the unmanned aerial vehicle at the edge of the cluster as a reference machine;

[0020] The distance between the nth unmanned aerial vehicle and the reference machine is d n , wherein n=0, 1, …, N-1;

[0021] At time t, the total data received by the reference machine electronic countermeasure reconnaissance payload from the K radiation sources is:

[0022]

[0023] In the formula, S k (t) is the data received by the reference machine electronic countermeasure reconnaissance payload at the kth radiation source at time t, and then the data received by the nth unmanned aerial vehicle electronic countermeasure reconnaissance payload without considering the error is x n (t):

[0024]

[0025] In the above formula, λ is the wavelength of the radiation source signal, w n (t) is noise, and j is a complex unit.

[0026] Further, the method for error correction of the received data in S30 is:

[0027] Considering the position error of the unmanned aerial vehicle, the distance between the nth unmanned aerial vehicle and the reference machine is d n will be directly changed;

[0028] Considering the amplitude and phase inconsistency error of the radio frequency channel, the radio frequency channel of the unmanned aerial vehicle electronic countermeasure reconnaissance payload will produce different amplitude gains, wherein the amplitude gain corresponding to the nth unmanned aerial vehicle is denoted as A n ; and different radio frequency channels of the unmanned aerial vehicle reconnaissance payload have different phase shifts, wherein the phase shift corresponding to the nth unmanned aerial vehicle is denoted as

[0029] The received data is processed by a function g to compensate for the non-linear processing error, wherein the function g does not have an additional impact on the noise part in the data during processing;

[0030] After considering the above three errors, the received data is p n (t) is obtained.

[0031]

[0032] Further, considering that there is also coupling interference between each unmanned aerial vehicle reconnaissance load, the coupling coefficient of the electronic countermeasure reconnaissance load of the nth unmanned aerial vehicle and then'th unmanned aerial vehicle is C n，n’ ,

[0033] Then the received data is changed from p n (t) to r n (t) is obtained.

[0034]

[0035] In the formula, A n′ is the amplitude gain of then'th unmanned aerial vehicle, is the phase shift of then'th unmanned aerial vehicle, and d n′ is the distance between then'th unmanned aerial vehicle and the reference machine.

[0036] Then the total data r(t) received by the N unmanned aerial vehicles is:

[0037] r(t) = [r0(t), r1(t),..., r N-1 (y)] T .

[0038] Further, the method for separating virtual and real in S40 is:

[0039] Separate the real part and the imaginary part of the total data r(t) received by the N unmanned aerial vehicles to obtain the neural input value y(t):

[0040] y(t) = [Real{r(t)}, img{r(t)}] T ,

[0041] In the formula, Real{r(t)} and img{r(t)} represent the real part and the imaginary part of r(t), respectively.

[0042] Further, the method for obtaining the output function in S50 is:

[0043] The input value Y of the neural network is the batch data obtained after the neural input function y(t) is accumulated for a plurality of fast shots, and then:

[0044] Y = [y(0), y(1), …, y(B-1)] T ,

[0045] In the formula, B is the batch quantity, the dimension of Y is B2X N, N is the total number of drone racks, and the method comprises the following steps:

[0046] S41, Y first passes through a full connection layer, and the output dimension of the full connection layer is BxFI, wherein F is the number of subsequent convolution layer filters; I is a value required for dimension expansion;

[0047] S42, after output from the full connection layer, tensor deformation is performed, and the output dimension is BxFXI

[0048] S43, then enter the convolution layer, and the output is processed by batch normalization and an activation function, the number of convolution layer is S, and the final output dimension is BxFI;

[0049] S44, the dimension BxFI is transformed into Bx2N through a full connection layer again, and an output function is obtained.

[0050] Further, the method for obtaining the spatial spectrum in S50 is:

[0051] The imaginary part and the real part of the output function are combined to obtain a complex function z, and the spatial spectrum f is:

[0052] f = |α H z|| 2 ,

[0053] In the formula, α H is a steering vector.

[0054] According to the second aspect of the application, a direction finding system based on a drone cluster under a multi-factor error condition is provided, comprising:

[0055] A network training module, under a laboratory ideal state, considers the position error of the drone, the amplitude-phase inconsistency error of the radio frequency channel, the non-linear processing error and the coupling interference error between the reconnaissance loads of each drone by the control variable method, and then constructs a network model considering the four errors;

[0056] A data acquisition module, which deploys a drone cluster and makes it carry electronic countermeasure reconnaissance loads; the drone senses and receives signals from each target radiation source while flying in the air, and records the data received by the electronic countermeasure reconnaissance loads of each drone;

[0057] A data correction module, which uses the network model in S10 to correct the errors of the data received in S20;

[0058] The virtual-real separation processing module separates the data after the error correction in S30 and sends the data into the neural network.

[0059] The neural network processing module obtains an output function through full connection layer, tensor deformation, convolution layer, batch normalization, activation function and full connection layer in the neural network, combines the virtual part and the real part of the output function to obtain a complex function, multiplies the complex function with a steering vector to obtain a spatial spectrum, and finally realizes direction finding of each radiation source by peak searching based on the spatial spectrum.

[0060] Overall, compared with the prior art, the above technical solutions conceived by the present application can achieve the following beneficial effects:

[0061] 1. The direction finding method of the present application considers the position error of the unmanned aerial vehicle, the amplitude-phase inconsistency error of the radio frequency channel, the non-linear processing error and the coupling interference error between each unmanned aerial vehicle reconnaissance load, and constructs a more accurate and practical network model; the accuracy and reliability of the unmanned aerial vehicle cluster in electronic countermeasure reconnaissance are improved.

[0062] 2. The direction finding method of the present application adopts a unique scheme with signal original data as input, retains the original characteristics of the signal, and extracts features through the convolution layer, which can more comprehensively capture signal information and avoid information loss and feature distortion; at the same time, through the tensor deformation technology, the complexity of data processing is reduced and the calculation efficiency is improved; in addition, the multiplication operation of the complex function at the output end and the steering vector directly obtains the spatial spectrum, without estimating the number of signal sources, further simplifying the direction finding process and enhancing the adaptability and flexibility of the system.

[0063] 3. The direction finding method of the present application gradually increases the consideration of error factors during network training, improves the generalization ability of the model, and also ensures the stability and accuracy of the model under different error conditions; in addition, the direction finding of multiple radiation sources is realized by peak searching of the spatial spectrum, which provides an efficient and accurate means for electronic reconnaissance and intelligence collection. BRIEF DESCRIPTION OF DRAWINGS

[0064] Figure 1 The measurement method schematic diagram provided for the preferred embodiment of the present application;

[0065] Figure 2 The information processing flow provided for the preferred embodiment of the present application;

[0066] Figure 3 The neural network architecture provided for the preferred embodiment of the present application;

[0067] Figure 4 The network training flow provided for the preferred embodiment of the present application;

[0068] Figure 5 Test results provided for the preferred embodiments of the present application;

[0069] Figure 6 A structure diagram of a direction finding system based on a UAV cluster under a multi-factor error condition according to an embodiment of the present application. DETAILED DESCRIPTION

[0070] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and should not be used to limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0071] Based on the above problems, the present application proposes a direction finding method based on a UAV cluster under a multi-factor error condition. This method combines laboratory research and practical application, and uses a network model to deal with various error sources such as UAV position error, radio frequency channel error, nonlinear processing error, and interference between UAVs.

[0072] In actual operation, the UAV cluster carries reconnaissance equipment to collect signal data from different radiation sources in the air. These data are then error-corrected by specially designed algorithms to improve the accuracy of the data. Then, the data are separated into real and imaginary parts and input into a neural network for further analysis.

[0073] The neural network outputs a function through multiple layers of processing, including full connection, convolution, normalization, and activation function steps. This function is combined with a steering vector to generate a spatial spectrum, which is used to determine the direction of the radiation source.

[0074] Specifically,

[0075] For reference Figure 1 and Figure 4 The present application relates to a direction finding method based on a UAV cluster under a multi-factor error condition, which comprises the following steps:

[0076] S10, network training, in a laboratory ideal state, by using the method of controlling variables, considering the position error of the UAV, the amplitude-phase inconsistency error of the radio frequency channel, the nonlinear processing error, and the coupling interference error between the reconnaissance loads of each UAV, and then constructing a network model considering the four errors.

[0077] The method for constructing the network model considering the four errors in S10 is:

[0078] S11, network training in the laboratory without considering the position error of the UAV, the amplitude-phase inconsistency error of the radio frequency channel, the non-linear processing error and the coupling interference error between the reconnaissance payloads of each UAV, and receiving data in an ideal state;

[0079] S12, training considering only the position error of the UAV in S11, assuming that the position error follows a Gaussian distribution with a mean of 0 and a position error standard deviation σ per , which follows a uniform distribution with a maximum value of σ per-max ;

[0080] S13, training considering only the amplitude-phase inconsistency error in S11, assuming that the amplitude-phase inconsistency error follows a Gaussian distribution with a mean of 0, a radio frequency channel amplitude standard deviation σ gain and a radio frequency channel phase standard deviation σ phase , which follows a uniform distribution with a maximum value of σ gain-max and σ phase-max ;

[0081] S14, training considering only the non-linear processing error in S11, the non-linear error is described by tanh(σ nl ), where σ nl is the non-linear error coefficient;

[0082] S15, training considering only the coupling interference error in S11, the coupling coefficient C n，n′ follows a uniform distribution with a maximum value of C max ;

[0083] S16, training considering the position error in S12, the amplitude-phase inconsistency error in S13, the non-linear processing error in S14 and the coupling interference error in S15, and constructing a network model considering the four errors.

[0084] S20, data acquisition, deploying a UAV cluster and making it carry electronic countermeasure reconnaissance payloads; the UAVs fly in the air, sense and receive signals from each target radiation source, and record the data received by the electronic countermeasure reconnaissance payloads of each UAV.

[0085] The method for recording the data received by the electronic countermeasure reconnaissance payloads of each UAV in S20 is as follows:

[0086] There are K radiation sources, and the direction of the incoming wave of each radiation source is θ k , where k=0, 1, …K-1, k is the radiation source number, and there are N UAVs, with the UAV at the edge of the cluster as the reference machine;

[0087] The distance between the nth UAV and the reference machine is d nwherein n = 0, 1, …, N-1;

[0088] At time t, the total data received by the electronic countermeasure reconnaissance payload of the reference machine from K radiation sources is:

[0089]

[0090] wherein S k (y) is the data received by the electronic countermeasure reconnaissance payload of the reference machine at time t at the kth radiation source, then the data received by the nth unmanned aerial vehicle electronic countermeasure reconnaissance payload without considering errors is x n (t):

[0091]

[0092] In the above formula, λ is the wavelength of the radiation source signal, w n (t) is noise, and j is a complex unit.

[0093] S30, data correction, using the network model in S10, error correction is performed on the data received in S20;

[0094] The method for performing error correction on the received data in S30 is:

[0095] Considering the position error of the unmanned aerial vehicle, the distance between the nth unmanned aerial vehicle and the reference machine is d n will be directly changed;

[0096] Considering the amplitude-phase inconsistency error of the radio frequency channel, the radio frequency channel of the unmanned aerial vehicle electronic countermeasure reconnaissance payload will produce different amplitude gains, wherein the amplitude gain corresponding to the nth unmanned aerial vehicle is denoted as A n ; and different unmanned aerial vehicle reconnaissance payload radio frequency channels have different phase shifts, wherein the phase shift corresponding to the nth unmanned aerial vehicle is denoted as

[0097] Considering the non-linear processing error, the received data is processed by a function g to compensate for the non-linear processing error, wherein the function g does not produce additional effects on the noise part in the data during processing;

[0098] After considering the above three errors, the received data is p n (t):

[0099]

[0100] Considering that there is also coupling interference between each unmanned aerial vehicle reconnaissance payload, the coupling coefficient of the electronic countermeasure reconnaissance payloads of the nth unmanned aerial vehicle and then'th unmanned aerial vehicle is C n，n′ ,

[0101] The received data is represented by p n (t) becomes r n (t):

[0102]

[0103] In the formula, A n′ is the amplitude gain of the nth unmanned aerial vehicle, is the phase shift of the nth unmanned aerial vehicle, d n′ is the distance between the nth unmanned aerial vehicle and the reference machine;

[0104] The total data r(t) received by the N unmanned aerial vehicles is:

[0105] r(t) = [r0(t), r1(t), …, r N-1 (t)] T .

[0106] Please refer to Figure 2 and Figure 3 , S40, virtual-real separation processing, after the ground station receives the error-corrected data in S30, virtual-real separation is performed and the data is sent into the neural network.

[0107] The method for performing virtual-real separation in S40 is:

[0108] The real part and the imaginary part of the total data r(t) received by the N unmanned aerial vehicles are separated to obtain the neural input value y(t):

[0109] y(t) = [Real{r(t)}, img{r(t)}] T ,

[0110] In the formula, Real{r(t)} and img{r(t)} represent the real part and the imaginary part of r(t), respectively.

[0111] S50, neural network processing, in the neural network, the data sequentially passes through the full connection layer, tensor deformation, convolution layer, batch normalization, activation function and full connection layer to obtain the output function, the imaginary part and the real part of the output function are combined to obtain a complex function, and the complex function is multiplied by a steering vector to obtain a spatial spectrum, and finally the spatial spectrum is searched for peaks to realize direction finding of each radiation source.

[0112] The method for obtaining the output function in S50 is:

[0113] The input value Y of the neural network is the batch processing data obtained after the neural input function y(t) is accumulated for multiple fast shots, and then:

[0114] Y = [y(0), y(1), …, y(B-1)] T ,

[0115] Where B is the batch size, the dimension of Y is B×2N, and N is the total number of drones. The specific steps include:

[0116] S41, Y first passes through the fully connected layer, and the dimension of the output of the fully connected layer is B×F·I, where F is the number of filters in the subsequent convolutional layer; I is the value required for dimensional expansion;

[0117] S42, after output from the fully connected layer, performs tensor deformation, and the output dimension is B×F×I

[0118] S43, then enters the convolution layer, the output is batch normalized and activated, the number of convolution layers is S, and the final output dimension is B × F·I;

[0119] S44, passes through the fully connected layer again, transforms the dimension B×F·I to B×2N, and obtains the output function.

[0120] The method for obtaining the spatial spectrum in S50 is:

[0121] Combining the imaginary part and the real part of the output function to obtain the complex function z, the spatial spectrum f is:

[0122] f=||α H z|| 2 ,

[0123] Where, α H is the guiding vector.

[0124] like Figure 6 As shown, as another aspect of the present invention, it also relates to a direction finding system based on a drone cluster under multi-factor error conditions, comprising:

[0125] The network training module, under ideal laboratory conditions, uses the control variable method to consider the UAV's position error, RF channel amplitude and phase inconsistency error, nonlinear processing error, and coupling and mutual interference error between the UAV's reconnaissance payloads, and then constructs a network model that takes into account these four errors.

[0126] The data acquisition module deploys a swarm of drones and enables them to carry electronic countermeasure reconnaissance payloads. When the drones are in flight, they sense and receive signals from each target radiation source and simultaneously record the data received by the electronic countermeasure reconnaissance payload of each drone.

[0127] A data correction module, using the network model in S10 to perform error correction on the data received in S20;

[0128] The virtual-real separation processing module, after receiving the error-corrected data in S30, the ground station performs virtual-real separation and then sends the data into the neural network;

[0129] The neural network processing module obtains an output function by sequentially passing data through a full connection layer, tensor deformation, a convolution layer, batch normalization, an activation function and a full connection layer in a neural network, combines an imaginary part and a real part of the output function to obtain a complex function, multiplies the complex function by a steering vector to obtain a spatial spectrum, and finally realizes direction finding of each radiation source by peak searching according to the spatial spectrum.

[0130] Embodiment 1

[0131] The test conditions are set as follows. The number of unmanned aerial vehicle racks N = 16, the number of radiation sources K = 3. The neural network hyperparameters: batch size B = 128, the number of convolution layer filters F = 2, the number of convolution layers S = 6. The error condition related parameters: the maximum standard deviation of unmanned aerial vehicle position error σ per-max = 0.2, the maximum standard deviation of radio frequency channel amplitude inconsistency σ gain-max = 0.5, the maximum standard deviation of radio frequency channel phase inconsistency σ phase-max = 0.2, the nonlinear error coefficient σ nl = 1, the maximum coupling coefficient C max = 0.1.

[0132] The test results are shown in Figure 5 The abscissa is the direction finding angle, and the ordinate is the spatial spectrum. It can be seen that the actual wave directions of each radiation source are -25°, -9.5° and 42.6°, and the direction finding results of the application are -24.8°, -9.9° and 42.5°, which are relatively close and can realize accurate direction finding under various error factors.

[0133] It is easy for those skilled in the art to understand that the above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modification, equivalent replacement and improvement within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A direction finding method based on drone clusters under multi-factor error conditions, characterized by: The following steps are involved: S10, network training, under ideal laboratory conditions, uses the control variable method to consider the UAV position error, RF channel amplitude and phase inconsistency error, nonlinear processing error, and coupling and mutual interference error between the UAV reconnaissance payloads, and then constructs a network model that considers these four errors; S20, data acquisition, deploying a swarm of drones and making them carry electronic countermeasure reconnaissance payloads; while the drones are in flight, they sense and receive signals from each target radiation source and simultaneously record the data received by the electronic countermeasure reconnaissance payloads of each drone; The method for recording the data received by each UAV electronic countermeasure reconnaissance payload in S20 is: There are K radiation sources, and the direction of the wave from each radiation source is θ k , where k = 0, 1, ... K-1, k is the radiation source number, there are N drones, and the drone at the edge of the cluster is the reference drone; The distance between the nth UAV and the reference UAV is d n , where n = 0, 1, ..., N-1; At time t, the total data received by the reference aircraft's electronic countermeasure reconnaissance payload from K radiation sources is: Where s k (t) is the data received by the electronic countermeasure reconnaissance payload of the reference aircraft at the kth radiation source at time t. Without considering the error, the data received by the electronic countermeasure reconnaissance payload of the nth UAV is x n (t): In the above formula, λ is the wavelength of the radiation source signal, w n (t) is the noise, j is a complex unit; S30, data correction, using the network model in S10 to perform error correction on the data received in S20; The method for performing error correction on the received data in S30 is: Considering the position error of the UAV, the distance between the nth UAV and the reference UAV is d n will change directly; Considering the amplitude and phase inconsistency error of the RF channel, the RF channel of the UAV electronic countermeasure reconnaissance payload will produce different amplitude gains, where the amplitude gain corresponding to the nth UAV is recorded as A n ; Different UAV reconnaissance payload RF channels have different phase shifts, where the phase shift corresponding to the nth UAV is recorded as Considering nonlinear processing errors, processing the received data by a function g to compensate for the nonlinear processing errors, wherein the function g does not generate additional effects on noise portions in the data during the processing; After considering the three errors, the received data is p n (t): S40, virtual-real separation processing, after receiving the error-corrected data in S30, the ground station performs virtual-real separation and then sends the data into the neural network; S50, neural network processing, in which the data passes through a fully connected layer, tensor transformation, convolutional layer, batch normalization, activation function, and fully connected layer in sequence to obtain an output function. The imaginary and real parts of the output function are combined to obtain a complex function. This complex function is then multiplied by the steering vector to obtain a spatial spectrum. Finally, the direction of each radiation source is achieved by finding the peak of the spatial spectrum.

2. The direction finding method based on drone clusters under multi-factor error conditions according to claim 1 is characterized in that: The method for constructing the network model considering the four errors in S10 is: S11: Network training is performed in the laboratory without considering the UAV position error, RF channel amplitude and phase inconsistency error, nonlinear processing error, and coupling and mutual interference error between the UAV reconnaissance payloads, and data in the ideal state is received; S12, perform training considering only the position error of the UAV in S11, assuming that the position error obeys Gaussian distribution, the average value is 0, and the standard deviation of the position error σ per Obeys uniform distribution, with a maximum value of σ per-max ; S13, perform training considering only the amplitude-phase inconsistency error in S11, assuming that the amplitude-phase inconsistency error obeys Gaussian distribution, with an average value of 0, and a standard deviation of the RF channel amplitude σ gain and the RF channel phase standard deviation σ phase Obeying uniform distribution, the maximum standard deviation of RF channel amplitude and RF channel phase are σ gain-max and σ phase-max ; S14, conduct training considering only the nonlinear processing error in S11, and the nonlinear error is expressed as tanh(σ nl ) description, where σ nl is the nonlinear error coefficient; S15, training is performed considering only the coupling mutual interference error in S11, and the coupling coefficient C n，n′ Obeying uniform distribution, the maximum coupling coefficient is C max ; S16, conduct training that comprehensively considers the position error in S12, the amplitude-phase inconsistency error in S13, the nonlinear processing error in S14, and the coupling mutual interference error in S15, and construct a network model that considers the four errors.

3. The direction finding method based on drone clusters under multi-factor error conditions according to claim 1 is characterized in that: Considering the coupling and mutual interference between the reconnaissance payloads of each UAV, the coupling coefficient of the electronic countermeasure reconnaissance payload of the nth UAV and the n′th UAV is C n，n′ , The received data is then n (t) becomes r n (t): Where A n′ is the amplitude gain of the n′th UAV, is the phase shift of the n′th UAV, d n′ is the distance between the n′th UAV and the reference aircraft; Then the total data r(t) received by N drones is: r(t)=[r0(t),r1(t),…,r N-1 (t)] T 。 4. The direction finding method based on drone clusters under multi-factor error conditions according to claim 3 is characterized in that: The method for performing virtual-real separation in S40 is: Separate the real and imaginary parts of the total data r(t) received by the N drones to obtain the neural input value y(t): y(t)=[Real{r(t)},img{r(t)}] T , Where Real{r(t)} and img{r(t)} represent the real and imaginary parts of r(t), respectively.

5. The direction finding method based on drone clusters under multi-factor error conditions according to claim 4 is characterized in that: The method for obtaining the output function in S50 is: The input value Y of the neural network is the batch data obtained after the neural input function y(t) is accumulated over multiple snapshots, then: Y=[y(0),y(1),…,y(B-1)] T , Where B is the batch size, the dimension of Y is B×2N, and N is the total number of drones. The specific steps include: S41, Y first passes through the fully connected layer, and the dimension of the output of the fully connected layer is B×F·I, where F is the number of filters in the subsequent convolutional layer; I is the value required for dimensional expansion; S42, after output from the fully connected layer, performs tensor deformation, and the output dimension is B×F×I S43, then enters the convolution layer, the output is batch normalized and activated, the number of convolution layers is S, and the final output dimension is B×F×I; S44, passes through the fully connected layer again, transforms the dimension B×F×I to B×2N, and obtains the output function.

6. The direction finding method based on drone clusters under multi-factor error conditions according to claim 5 is characterized in that: The method for obtaining the spatial spectrum in S50 is: Combining the imaginary part and the real part of the output function to obtain the complex function z, the spatial spectrum f is: f=||α H with|| 2 , Where a H is the guiding vector.

7. A direction-finding system based on drone swarms under multi-factor error conditions, used to implement the direction-finding method based on drone swarms under multi-factor error conditions as claimed in any one of claims 1 to 6, characterized in that: include: The network training module, under ideal laboratory conditions, uses the control variable method to consider the UAV's position error, RF channel amplitude and phase inconsistency error, nonlinear processing error, and coupling and mutual interference error between the UAV's reconnaissance payloads, and then constructs a network model that considers these four errors. The data acquisition module deploys a swarm of drones and enables them to carry electronic countermeasure reconnaissance payloads. When the drones are in flight, they sense and receive signals from each target radiation source and simultaneously record the data received by the electronic countermeasure reconnaissance payload of each drone. The data correction module uses the network model in S10 to perform error correction on the data received in S20; Virtual-real separation processing module: after receiving the error-corrected data from S30, the ground station performs virtual-real separation and then feeds the data into the neural network; In the neural network processing module, data passes through a fully connected layer, tensor transformation, convolutional layer, batch normalization, activation function, and fully connected layer in sequence to obtain an output function. The imaginary and real parts of the output function are combined to obtain a complex function. This complex function is then multiplied by the steering vector to obtain a spatial spectrum. Finally, the direction of each radiation source is achieved by finding the peak of the spatial spectrum.

Images