Distortion communication complex signal anti-interference intermediate frequency generation method based on complex generative adversarial network

Abstract

The application discloses a distortion communication complex signal anti-interference intermediate frequency generation method based on CVGAN. The core of the application is a special structure of the adversarial network, which solves the complex signal generation problem, so that the received distorted radio frequency signal can remove the interference through the network to generate an anti-interference intermediate frequency signal. Through the generated anti-interference intermediate frequency signal, the receiving end can complete the intermediate frequency signal integrity verification and radio frequency fingerprint identification. After the verification and identification of the intermediate frequency signal, the intermediate frequency signal can be extracted to the baseband to further complete the analysis of the baseband information.

Description

Technical Field

[0001] This invention relates to the field of radio communication technology, and more specifically to an intermediate frequency generation method for anti-interference of distorted communication complex signals based on complex generative adversarial networks. Background Technology

[0002] With the development and increasing number of wireless devices, various industries have begun to adopt wireless methods for information interconnection. To resist environmental interference, radio frequency electromagnetic interference, and communication channel fading, wireless devices typically use high-frequency signals to carry and transmit information. At the wireless communication signal receiver, a down-conversion technique is usually used to convert the received signal into an intermediate frequency (IF) signal, which falls between high frequency and baseband. Then, through a series of processing steps, the IF signal is down-converted again to a baseband signal. The baseband signal uses a 1s and 0s encoding mode, and the information carried by the communication signal can be analyzed using the baseband signal. Therefore, the stability and reliability of the received IF signal directly affect the accuracy and reliability of the subsequent baseband information analysis.

[0003] In real-world transmission environments, various electromagnetic interferences are common, causing distortion in radio frequency (RF) signals. This leads to abnormal recovery of the received intermediate frequency (IF) signal, making information decoding impossible and severely impacting the reliability and security of wireless communication, particularly in areas like air traffic control, drone control, and vehicle-to-everything (V2X) communication. Wireless RF signals typically face two challenges that need to be balanced: 1. High-frequency signals are highly susceptible to Doppler interference; the movement of objects can easily interfere with the RF signal waveform, preventing the receiver from interpreting the high-frequency information; 2. Lowering the frequency of high-frequency signals results in lower information carrying efficiency and poor resistance to other interference. Therefore, efficient utilization of RF signals from wireless devices requires the receiver to have extremely strong anti-interference capabilities at the IF level, enabling rapid recovery of distorted RF signals. This ensures the receiver can perform subsequent verification and interpretation of the RF signal, and then recover the baseband signal from the "clean" IF signal for correct decoding of the baseband information.

[0004] The drawbacks of directly recovering baseband signals: In patent CN 112202529 B, when recovering severely distorted wireless communication signals, the baseband information is analyzed by directly recovering the baseband signal. However, without anti-interference recovery of the intermediate frequency (IF) signal, the accuracy of directly recovering baseband transmission information is extremely low. First, if the check bit of the distorted IF signal is severely distorted, the signal cannot pass the integrity and security checks when analyzing the baseband signal. Second, in today's wireless devices, radio frequency (RF) identification is usually performed on the IF signal at the front end, and RF identification typically uses the received IF signal for identification. Therefore, the traditional and existing methods of directly recovering baseband information are highly unreliable. The biggest difficulties in recovering anti-interference IF signals from distorted RF signals are: 1. The randomness of IF waveform point values ​​is difficult to recover; 2. The phase relationship between the real and virtual paths of the complex signal is difficult to recover. Therefore, there is currently no anti-interference IF generation device based on distorted RF signals. Summary of the Invention

[0005] To address the aforementioned shortcomings in the existing technology, this invention provides a method for generating intermediate frequency (IF) signals for anti-interference in distorted communication based on complex generative adversarial networks, which solves the problem of generating IF signals for anti-interference based on distorted radio frequency signals.

[0006] To achieve the aforementioned objectives, the present invention employs the following technical solution: a method for generating intermediate frequency signals for anti-interference of distorted communication complex signals based on complex generative adversarial networks, characterized by comprising the following steps:

[0007] S1. Receive radio frequency complex signals from different terminals and of different types through the receiving terminal, downconvert all radio frequency signals to intermediate frequency, generate an interference intermediate frequency signal library as an interference dataset;

[0008] S2. Perform intermediate frequency (IF) signal simulation on commonly used signal types in the received signal to generate a standard IF signal library as a standard dataset.

[0009] S3. Align the signal lengths of the interference dataset and the standard dataset by upsampling or downsampling respectively, so that all intermediate frequency signals are kept at the same length.

[0010] S4. Establish an integrated adversarial network G. Input both the interfering intermediate frequency (IF) signal and the standard IF signal into the integrated adversarial network G. G extracts the interference waveform from the interfering IF signal based on the standard IF signal to form an anti-interference IF signal. G integrates two adversarial generator networks (A and B). Adversarial generator network A is input with a combination matrix of the real part of the standard signal and the real part of the interfering signal to extract the interference waveform of the real part of the interfering signal and generate an anti-interference real part waveform of the real part of the interfering signal. Adversarial generator network B is input with a combination matrix of the imaginary part of the standard signal and the imaginary part of the interfering signal to extract the interference waveform of the imaginary part of the interfering signal and generate an anti-interference imaginary part waveform of the imaginary part of the interfering signal.

[0011] S5. Establish a complex signal fusion equation to fuse the anti-interference real and imaginary waveforms into a real number expression for the complex signal;

[0012] S6. Establish a supervisor C to monitor the phase relationship between the real and imaginary anti-interference waveforms generated by A and B, ensuring that the real and imaginary parts always maintain the constraint relationship of the standard signal. The generated real and imaginary anti-interference parts are fused to obtain the real expression 1 of the anti-interference complex signal through a fusion equation. The real and imaginary parts of the standard intermediate frequency signal are fused to obtain the real expression 2 of the standard complex signal through a fusion equation. Input these two real expressions into C. C outputs 0 for expression 1 and 1 for expression 2.

[0013] S7. Joint training G and C form the complete structure of CVGAN. The output of supervisor C is used to repeatedly stimulate G to generate the real and imaginary parts of the intermediate frequency anti-interference complex signal until supervisor C can no longer distinguish between expression 1 and expression 2.

[0014] Furthermore, the process includes step S8: Signals transmitted by different devices are labeled according to the transmitting device, and strict interference removal is performed on these signals; these interference-removed data are divided into training and testing sets proportionally; artificially set interference is added to the testing set, with a signal-to-noise ratio ranging from 10dB to -20dB and a Doppler frequency shift ranging from 0 to the maximum frequency shift calculated by the maximum frequency shift formula; a deep learning network model D is trained using the standard training set to identify signal categories; the interference test set, through the anti-interference signal generated by the trained G model, is input into D, and the accuracy of generating the anti-interference complex signal is evaluated by the recognition accuracy of the complex signal generated by D against interference.

[0015] Furthermore: the adversarial generative network A consists of an Ag network model and an Ad network model, and the adversarial generative network B consists of a Bg network model and a Bd network model, where g represents the generator network and d represents the discriminator network.

[0016] Furthermore: the method for establishing the adversarial generative network A is as follows:

[0017] A1. Let the standard intermediate frequency signal be... The interfering intermediate frequency signal is The generated anti-interference intermediate frequency signal is ,in To generate interference factors extracted by the network learning process;

[0018] A2. Establish a boundary loss function for training the Ag network model and the Ad network model. Specifically:

[0019]

[0020] In the above formula, Output predicted values ​​for the network model. For the true value, for Values ​​in the array for , Here, n is the loss function, and n is the total length of the output values.

[0021]

[0022] ;

[0023]

[0024] In the above formula, An array consisting entirely of zeros represents false. Given the [generated signal, generated noise value] array of Ag, input the discriminant value obtained from the Ad network, and use... Calculation makes The values ​​in the vector approach 0, thus motivating Ad to correctly identify the generated values; An array of all 1s represents true. Given an array of [true signal value, true noise value] for Ag, input the discriminant value obtained from the Ad network, and then use... Calculation makes The value in is close to 1, enabling the Ad network to distinguish the true value of the signal from the noise; I is the real part of the standard signal. The real part of the interference signal;

[0025] A4. Establish the objective function for Ag. ;

[0026]

[0027] ;

[0028] In the above formula, Given the [generated signal, generated noise value] array of Ad, input the discriminant value obtained from the Ag network. This function reduces and The value of causes Ag to be misjudged as true by the Ad network; Given the [true signal value, true noise value] array of Ad, input the discriminant value obtained from the Ag network; through and The mutual pulling of Ag and Ad stimulates both Ag and Ad to become more robust;

[0029] A5. Establish an amplitude supervision function for the interference factor to ensure that the generated interference factor never exceeds the value of c, and calculate the amplitude loss value. ;

[0030]

[0031] In the above formula, In order to be in The average expected value of this function in the domain. Interference factor Dimodulo value;

[0032] A6. Establish the loss function calculated for generated data and real data. Where N is the total length of Y;

[0033] A7. Add a supervised interference factor generation function to make the anti-interference intermediate frequency waveform after removing the interference factor closer to the standard interference-free waveform, and calculate the loss function. ;

[0034]

[0035] In the above formula, In order to be in The mean expected value of the domain's supervised interference factor generation function. To generate waveforms Compared with the actual waveform For the difference value, For the generated interference factor Compared with actual interference factors The difference value;

[0036] A8. Train Ag and Ad until each loss function is stable to obtain the adversarial generative network A.

[0037] Further: Step A8 specifically includes:

[0038] A81. Take out a pair , The distribution is obtained and ;

[0039] A82. Will and Input Ad, calculate Ad and obtain the output value Ad through the calculation between Ag and Ad. By making If the value is smaller, update the parameters of Ad in reverse;

[0040] A83. Will and Input Ag, calculate Ag and Ad to obtain the output value of Ag. , as well as By making , as well as If the value is smaller, update the parameters of Ag in reverse;

[0041] A84. Repeat A82 and A83 until each loss function is stationary.

[0042] Furthermore, the method for establishing the adversarial generative network B is the same as the method for establishing the adversarial generative network A, and the input of B is the imaginary part of the standard signal. Q With the imaginary part of the interference signal .

[0043] Furthermore: the complex signal fusion equation is as follows Where I is the real part of the standard signal and Q is the imaginary part of the standard signal.

[0044] Furthermore: the method for establishing the supervisor C is as follows:

[0045] C1. Establish the real number expression of the standard signal ;

[0046] C2. Establish the real number expression for the interference signal. ,in, The real part of the interference signal. The imaginary part of the interference signal;

[0047] C3. Generate the real number expression for the interference factor ,in The interference factor generated for A Interference factor generated for B;

[0048] C4. Generate the anti-interference intermediate frequency real number expression ;

[0049] C5. Establish the objective functions for the Cg and Cd network models in the supervisory system C. , and loss function , ;

[0050]

[0051]

[0052]

[0053]

[0054]

[0055] In the above formula, and The values ​​are the secondary fusion values ​​after processing the real and imaginary parts of the interference signal respectively. "fake" represents the output value of the Cd supervised network model after the second fusion of [generated signal value, generated noise value], and "real" represents the output value of the Cd supervised network model after the second fusion of [real signal value, real noise value]. For network model Y fake , For network model Y real Its calculation principle is the same as that of Ad and Ag mentioned above. Interference factor Dimodulo value;

[0056]

[0057]

[0058] C6. Train Cg and Cd until each loss function is stationary, thus obtaining the supervisor C.

[0059] Further: Step C6 specifically includes:

[0060] C61. Take out a pair , ;

[0061] C62. Will , Input Cd, and calculate the output value of Cd using M, Ag, Bg and Cd. By making If the value is smaller, update the parameters of Cd in reverse;

[0062] C63. Will , Input Ag, and calculate using M, Ag, Bg, and Cd to obtain... , as well as By making , as well as If the value is smaller, update the parameters of Ag and Bg in reverse;

[0063] C64. Repeat C62 and C63 until each loss function is stationary.

[0064] The beneficial effects of this invention are as follows: This invention provides a complex-valued generative adversarial network (CVGAN) based on complex signals. The core of this network is a specially structured adversarial network that solves the problem of complex signal generation. This network removes interference from the received distorted radio frequency (RF) signal, generating an anti-interference intermediate frequency (IF) signal. Using this generated IF signal, the receiving end can perform IF signal integrity verification and RF fingerprinting. The verified and identified IF signal can be extracted to the baseband for further baseband information analysis. The proposed complex generative adversarial network can be effectively applied to the generation of distorted signal anti-interference IF signals. Under extreme interference (SNR = -20 dB with extreme Doppler shift), the generated IF signal still achieves an RF fingerprinting accuracy of 50%. Under moderate interference (SNR = -5 dB with extreme Doppler shift), the RF fingerprinting accuracy can reach over 90%. At this accuracy level, the anti-interference IF signal can successfully complete verification and identification, and also complete baseband information analysis. Meanwhile, based on complex generative adversarial networks, the capacity and efficiency of intermediate frequency complex signal radio frequency fingerprint recognition networks can be effectively reduced. Attached Figure Description

[0065] Figure 1 This is a system flowchart of the present invention;

[0066] Figure 2 This is a diagram showing the interference-affected ADS-B waveform and phase distribution in an embodiment of the present invention.

[0067] Figure 3 This is a diagram illustrating the architecture of Generative Adversarial Network A and Generative Adversarial Network B in an embodiment of the present invention.

[0068] Figure 4 A diagram of the traditional adversarial generative network structure;

[0069] Figure 5 This is a diagram of the overall structure of CVGAN in an embodiment of the present invention;

[0070] Figure 6 This is a flowchart of the dataset partitioning process in an embodiment of the present invention;

[0071] Figure 7This is a schematic diagram illustrating the accuracy evaluation of two data processing methods using the anti-interference intermediate frequency signal of CVGAN in an embodiment of the present invention.

[0072] Figure 8 This is a schematic diagram illustrating the system accuracy of different combinations of CVGAN in embodiments of the present invention;

[0073] Figure 9 This is a comparison chart of classifiers in an embodiment of the present invention. Detailed Implementation

[0074] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0075] A specific implementation example uses ADS-B signals. These signals vary in length and are PPM signals, exhibiting irregular phase distributions. This makes them more effective for robust evaluation of intermediate frequency recovery in complex signals. Figure 1 As shown, the specific process is as follows:

[0076] 1. Establish an interference dataset for ADS-B signal reception in a real-world environment. Signal acquisition is performed using a center frequency of 1090MHz and a sampling frequency of 150MHz. The signals are truncated and filtered based on the number of valid square waves, retaining only the ADS-B interrogation and response signals (filtered by the interval value of the number of valid square waves). The complete lengths of all signals are inconsistent. The waveform and phase distribution of the interfered ADS-B signal are shown below. Figure 2 As shown, the first row of signal time-domain images, from left to right, represents the base signal; A=0, SNR=-5dB; A=0, SNR=-10dB; A=0, SNR=-20dB; A=0.05, SNR=10dB; A=0.5, SNR=10dB; A=0.8, SNR=10dB. The second row is the signal phase diagram, from left to right, representing the base signal; A=0, SNR=-5dB; A=0, SNR=-10dB; A=0, SNR=-20dB; A=0.05, SNR=10dB; A=0.5, SNR=10dB; A=0.8, SNR=10dB. A=0.05 represents a frequency offset of 3.75mHz, 0.5 represents a frequency offset of 37.5mHz, and 0.8 represents a frequency offset of 60mHz.

[0077] 2. Simulate the main propagation types and lengths of ADS-B signals, and establish a standard dataset for the simulation;

[0078] 3. Establish an adversarial generative network A, as shown in Figure 1. Figure 3 As shown; specifically:

[0079] a. Let the basic signal be: The interference signal is: ;

[0080] b. Traditional Generative Adversarial Networks: Traditional generative adversarial networks generally consist of two network models, such as... Figure 4 As shown, the generator network and the discriminator network are respectively. The generator network is responsible for generating simulated data that is close to real data, and the discriminator is responsible for distinguishing whether it is real data. If it is simulated data generated by the generator network, it is judged as 0, and if it is real data, it is judged as 1. By training the discriminator network, the generator network is stimulated to generate more realistic simulated data. By training the generator network, the discriminator network can be stimulated to identify more realistic simulated data, until the two reach a loss balance.

[0081] The G designed in this invention consists of adversarial generative networks A and B. A is composed of Ag and Ad, and B is composed of Bg and Bd. g represents the generator network and d represents the discriminator network.

[0082] Our designed A:Ag network model consists of three convolutional networks, while the Ad network model consists of three fully connected layers.

[0083] c. The generated anti-interference intermediate frequency signal is: ,in To generate interference factors extracted by the network, the information entropy of the interference factors is much lower than that of the complete signal. Therefore, by extracting interference factors first through the generator network and then subtracting the interference factors, the resulting anti-interference intermediate frequency signal is not only faster to train, but also more accurate.

[0084] d. Establish boundary loss functions for training the g network and the d network:

[0085]

[0086] e. Establish the objective function of Ad. .

[0087] To monitor the generated anti-interference intermediate frequency signal, Ad typically identifies the input generated anti-interference signal as 0 and the standard signal from the base database as 1.

[0088] Traditional:

[0089] ;

[0090] in To simulate real-world data generation, the discriminator is trained to identify generated data as 0 and real data as 1.

[0091] f. Compared to the traditional GAN ​​function d, we made an optimization to enable the Ad function to better supervise the generated anti-interference signal. Ad design ( ):

[0092] ;

[0093] Among them, As a set of inputs, representing true, the Ad supervisor sets the output of this set to 1, representing the true standard signal and the true interference factor (interference signal minus standard signal); As one set of inputs, represented as false, causing Ad to supervise the output of this set to be 0. Therefore, the objective function of Ad is to output 1 and 0 for the two sets of inputs respectively:

[0094]

[0095]

[0096] The calculated data is output via Ad. The distance between the value and 0 is used to modify the parameters of the Ad network through continuous training to adjust the output value. By continuously approaching 0, the generated data can be judged. Similarly.

[0097] g. Optimize the input of the g function in traditional GANs; similarly to Ad, establish the objective function of Ag. This function enables the anti-interference intermediate frequency (IF) signal generated by Ag to deceive Ad, ensuring that Ad still outputs a 1 when the generated IF signal is input. Its purpose is to stimulate each other under the supervision of the two objective functions, Ag and Ad, thereby improving Ad's supervisory capabilities and Ag's generation capabilities.

[0098] ;

[0099]

[0100] h. Establish an amplitude supervision function for the interference factor to ensure that the generated interference factor never exceeds the value of c, where the value of c is determined by the average amplitude of the actual received signal dataset. In order to be in The average expected value of this function in the domain. The amplitude loss value is generated through training to make the interference factor... The second modulus value (amplitude value) does not exceed c:

[0101]

[0102] i. Establish loss functions for generated data and real data.

[0103] j. Add a supervisory interference factor generation function to make the anti-interference intermediate frequency waveform with interference factor removed closer to the standard interference-free waveform. In order to be in The average expected value of this function in the domain. Used to calculate and generate waveforms respectively Compared with the actual waveform The difference value, the generated interference factor Compared with actual interference factors The difference value is calculated to train Ag, and the Ag parameters are optimized in reverse to reduce the difference value, so that the generated anti-interference waveform is consistent with the standard waveform.

[0104]

[0105] k. Training Ag and Ad:

[0106] (1) Take out a pair , , distribution obtained and ;

[0107] (2) and Input Ad, calculate Ad and obtain the output value Ad through the calculation between Ag and Ad. ;

[0108] By making If the value is smaller, update the parameters of Ad in reverse;

[0109] (3) and Input Ag, calculate Ag and Ad to obtain the output value of Ag. , as well as ; by making , as well as If the value is smaller, update the parameters of Ag in the opposite direction;

[0110] (4) Repeat (2) and (3) until each loss value is stable.

[0111] 4. Establish an adversarial generative network B, as shown in Figure B. Figure 3 As shown, its setup process, network structure, loss function establishment, and training process are consistent with A; the input of B is... ;

[0112] 5. Establish the fusion equation M: ;

[0113] 6. Establish a supervisor C. The purpose of supervisor C is to strengthen the generators A and B, ensuring that the generated data (real and imaginary parts) always conform to the phase relationship and angular distribution between the real and imaginary parts of the actual data.

[0114] a. Standard real number expression for a signal: ;

[0115] b. Real number expression of the interference signal: ;

[0116] c. Real number expression for the generated interference factor: ;

[0117] d. Generated anti-interference intermediate frequency real number expression: ;

[0118] e. ;

[0119] f. Establish the objective functions for the Cg and Cd network models in supervisor C. , and loss function , ;

[0120]

[0121]

[0122]

[0123]

[0124]

[0125] g. Training process:

[0126] (1) Take out a pair , ;

[0127] (2) , Input Cd, and calculate the output value of Cd using M, Ag, Bg and Cd. ;

[0128] By making If the value is smaller, update the parameters of Cd in reverse;

[0129] (3) , Input Ag and Bg, and calculate using M, Ag, Bg, and Cd to obtain... ; by making If the value is smaller, update the parameters of Ag and Bg in the opposite direction;

[0130] (4) Repeat (2) and (3) until each loss value is stable.

[0131] 7. Repeat steps 3, 4, and 5 of the training process until all loss values ​​converge and stabilize, obtaining Ag and Bg. These are the real and imaginary generating functions of the intermediate frequency signal for receiving distorted RF signals to resist interference. Assume the intermediate frequency signal for receiving distorted RF signals is: The generated anti-interference intermediate frequency signal is: Its overall structure diagram is as follows: Figure 5 As shown.

[0132] 8. A diverse interference dataset was created from the collected signals. A portion of the signals was extracted and professionally labeled. The generated anti-interference intermediate frequency (IF) signal was then tested. The interference dataset extracted signals from 10 devices. For each device, 200 interference signals were extracted and professionally separated and labeled. Each signal achieved a 10dB noise level and no frequency offset. This dataset is called the standard acquisition dataset. The standard acquisition dataset was divided into 180 signals for each device, called the training dataset for classifier D. The remaining 20 signals from each device were mixed with -20dB to 10dB noise and a 0 to 60mHz frequency offset, resulting in a final test set containing various interferences, with 140 signals per device. The process is as follows: Figure 6 As shown.

[0133] 10. Evaluation Results:

[0134] The direct effect of CVGAN: Two data length alignment methods are used, namely data processing 1 and data processing 2. The test structure is as follows. Figure 7As shown, under cross-interference, the accuracy of the intermediate frequency signal not recovered by CVGAN in data processing method 1 is 44%, while that of CVGAN is 62%, an improvement of 18%. With a frequency offset of 37.5mHz, CVGAN improves the accuracy by 19%, 30%, and 20% at signal-to-noise ratios of 5dB, -10dB, and -20dB, respectively. With a signal-to-noise ratio of -10dB, CVGAN improves the accuracy by 17%, 25%, and 13% at frequency offsets of 3.75mHz, 37.5mHz, and 60mHz, respectively. Even with no frequency offset at 10dB, the accuracy is improved by 6.5%. Under cross-interference conditions, the accuracy of intermediate frequency (IF) signal recognition by data processing method 2 without CVGAN recovery was 30%, while that of CVGAN was 53%, an improvement of 23%. With a frequency offset of 37.5 MHz and signal-to-noise ratios (SNRs) of 5 dB, -10 dB, and -20 dB, CVGAN improved accuracy by 24%, 31%, and 5%, respectively. With an SNR of -10 dB and frequency offsets of 3.75 MHz, 37.5 MHz, and 60 MHz, CVGAN improved accuracy by 47%, 38%, and 5%, respectively. Even with no frequency offset at 10 dB, it improved accuracy by 10%. These data demonstrate that CVGAN can very well recover IF waveforms from simulated standard waveforms, obtaining highly interference-resistant IF waveforms. It exhibits excellent generation capabilities for both high SNR low-frequency offsets and low SNR high-frequency offsets, and can cope with randomly changing and heavily interfered environments.

[0135] The roles of CVGAN network A, B, and C: (e.g.) Figure 8 As shown, CVGAN1 consists only of A and B, without C supervision; CVGAN2 consists only of real numbers supervised by C; and CVGAN consists of A, B, and C simultaneously. It is evident that CVGAN with added C performs significantly better than CVGAN1 without C supervision. In principle, CVGAN1 supervises the extracted amplitude values ​​of the real and imaginary parts, while CVGAN2 supervises the constraint relationship between the real and imaginary parts of complex numbers. Therefore, our proposed CVGAN, targeting the characteristics of complex signals, prioritizes a strategy of separating the complex signal and extracting interference factors from the real and imaginary parts separately. Simultaneously, it proposes a supervisor to oversee the relationship between the real and imaginary parts, ensuring that the intermediate frequency complex signal recovered from the real and imaginary part data after interference factor extraction has extremely strong anti-interference capabilities, closely approximating the standard signal mode. This enables subsequent processes such as baseband signal recovery, baseband signal verification and authentication, and baseband information parsing. CVGAN reduces the capacity of the RFID classification network as follows: Figure 9 As shown.

Claims

1. A method for generating intermediate frequency signals for anti-interference of distorted communication complex signals based on complex generative adversarial networks, characterized in that, Includes the following steps: S1. Receive radio frequency complex signals from different terminals and of different types through the receiving terminal, downconvert all radio frequency signals to intermediate frequency, generate an interference intermediate frequency signal library as an interference dataset; S2. Perform intermediate frequency (IF) signal simulation on commonly used signal types in the received signal to generate a standard IF signal library as a standard dataset. S3. Align the signal lengths of the interference dataset and the standard dataset by upsampling or downsampling respectively, so that all intermediate frequency signals are kept at the same length. S4. Establish an integrated adversarial network G. Input both the interfering intermediate frequency (IF) signal and the standard IF signal into the integrated adversarial network G. G extracts the interference waveform from the interfering IF signal based on the standard IF signal to form an anti-interference IF signal. G integrates two adversarial generator networks (A and B). Adversarial generator network A is input with a combination matrix of the real part of the standard signal and the real part of the interfering signal to extract the interference waveform of the real part of the interfering signal and generate an anti-interference real part waveform of the real part of the interfering signal. Adversarial generator network B is input with a combination matrix of the imaginary part of the standard signal and the imaginary part of the interfering signal to extract the interference waveform of the imaginary part of the interfering signal and generate an anti-interference imaginary part waveform of the imaginary part of the interfering signal. S5. Establish a complex signal fusion equation to fuse the anti-interference real and imaginary waveforms into a real number expression for the complex signal; S6. Establish a supervisor C to monitor the phase relationship between the real and imaginary anti-interference waveforms generated by A and B, ensuring that the real and imaginary parts always maintain the constraint relationship of the standard signal. The generated real and imaginary anti-interference parts are fused to obtain the real expression 1 of the anti-interference complex signal through a fusion equation. The real and imaginary parts of the standard intermediate frequency signal are fused to obtain the real expression 2 of the standard complex signal through a fusion equation. Input these two real expressions into C. C outputs 0 for expression 1 and 1 for expression 2. S7. Joint training G and C form the complete structure of CVGAN. The output of supervisor C is used to repeatedly stimulate G to generate the real and imaginary parts of the intermediate frequency anti-interference complex signal until supervisor C can no longer distinguish between expression 1 and expression 2.

2. The intermediate frequency generation method for anti-interference of distorted communication complex signals based on complex generative adversarial networks according to claim 1, characterized in that, It also includes step S8, which involves labeling the data of signals transmitted by different devices according to the transmitting devices and strictly removing interference from these signals; dividing the interference-removed data into training set and test set according to the proportion; adding artificially set interference to the test set, with the signal-to-noise ratio ranging from 10dB to -20dB and the Doppler frequency shift ranging from 0 to the maximum frequency shift calculated by the maximum frequency shift formula; A deep learning network model D is trained using a standard training set to identify signal categories. An anti-interference test set, generated by the trained model G, is input into D. The accuracy of generating complex signals against interference is evaluated by the recognition accuracy of complex signals generated by D against interference.

3. The intermediate frequency generation method for anti-interference of distorted communication complex signals based on complex generative adversarial networks according to claim 1, characterized in that, The adversarial generative network A consists of an Ag network model and an Ad network model, and the adversarial generative network B consists of a Bg network model and a Bd network model, where g represents the generator network and d represents the discriminator network.

4. The intermediate frequency generation method for anti-interference of distorted communication complex signals based on complex generative adversarial networks according to claim 1, characterized in that, The method for establishing the adversarial generative network A is as follows: A1. Let the standard intermediate frequency signal be... The interfering intermediate frequency signal is The generated anti-interference intermediate frequency signal is ,in To generate interference factors extracted by the network learning process; A2. Establish a boundary loss function for training the Ag network model and the Ad network model. Specifically: In the above formula, Output predicted values ​​for the network model. For the true value, for Values ​​in the array for , Here, n is the loss function, and n is the total length of the output values. A3. Establish the objective function of the Ad network model. ; ; In the above formula, An array consisting entirely of zeros represents false. Given the [generated signal, generated noise value] array of Ag, input the discriminant value obtained from the Ad network, and use... Calculation makes The values ​​in the vector approach 0, thus motivating Ad to correctly identify the generated values; An array of all 1s represents true. Given an array of [true signal value, true noise value] for Ag, input the discriminant value obtained from the Ad network, and then use... Calculation makes The value in the value is close to 1, which enables the Ad network to identify the true value of the signal and noise; I represents the real part of the standard signal. The real part of the interference signal; A4. Establish the objective function for Ag. ; ; In the above formula, Given the [generated signal, generated noise value] array of Ad, input the discriminant value obtained from the Ag network. This function reduces and The value of causes Ag to be misjudged as true by the Ad network; Given the [true signal value, true noise value] array of Ad, input the discriminant value obtained from the Ag network; through and The mutual pulling of Ag and Ad stimulates both Ag and Ad to become more robust; A5. Establish an amplitude supervision function for the interference factor to ensure that the generated interference factor never exceeds the threshold c, and calculate the amplitude loss value. ; In the above formula, In order to be in The average expected value of this function in the domain. Interference factor Dimodulo value; A6. Establish the loss function calculated for generated data and real data. Where N is the total length of Y; A7. Add a supervised interference factor generation function to make the anti-interference intermediate frequency waveform after removing the interference factor closer to the standard interference-free waveform, and calculate the loss function. ; In the above formula, In order to be in The mean expected value of the domain's supervised interference factor generation function. To generate waveforms Compared with the actual waveform For the difference value, For the generated interference factor Compared with actual interference factors The difference value; A8. Train Ag and Ad until each loss function is stable to obtain the adversarial generative network A.

5. The intermediate frequency generation method for anti-interference of distorted communication complex signals based on complex generative adversarial networks according to claim 4, characterized in that, Step A8 specifically involves: A81. Take out a pair , The distribution is obtained and ; A82. Will and Input Ad, calculate Ad and obtain the output value Ad through the calculation between Ag and Ad. By making If the value is smaller, update the parameters of Ad in reverse; A83. Will and Input Ag, calculate Ag and Ad to obtain the output value of Ag. , as well as By making , as well as If the value is smaller, update the parameters of Ag in reverse; A84. Repeat A82 and A83 until each loss function is stationary.

6. The intermediate frequency generation method for anti-interference of distorted communication complex signals based on complex generative adversarial networks according to claim 5, characterized in that, The method for establishing the adversarial generative network B is the same as that for establishing the adversarial generative network A, and the input of B is the imaginary part of the standard signal. Q With the imaginary part of the interference signal .

7. The intermediate frequency generation method for anti-interference of distorted communication complex signals based on complex generative adversarial networks according to claim 6, characterized in that, The complex signal fusion equation is as follows: Where I is the real part of the standard signal and Q is the imaginary part of the standard signal.

8. The intermediate frequency generation method for anti-interference of distorted communication complex signals based on complex generative adversarial networks according to claim 7, characterized in that, The method for establishing the supervisor C is as follows: C1. Establish the real number expression of the standard signal ; C2. Establish the real number expression for the interference signal. ,in, The real part of the interference signal. The imaginary part of the interference signal; C3. Generate the real number expression for the interference factor ,in The interference factor generated for A Interference factor generated for B; C4. Generate the anti-interference intermediate frequency real number expression ; C5. Establish the objective functions for the Cg and Cd network models in the supervisory system C. , and loss function , ; In the above formula, and The values ​​are the secondary fusion values ​​after processing the real and imaginary parts of the interference signal respectively. "fake" represents the output value of the Cd supervised network model after the second fusion of [generated signal value, generated noise value], and "real" represents the output value of the Cd supervised network model after the second fusion of [real signal value, real noise value]. For network model Y fake , For network model Y real Its calculation principle is the same as that of Ad and Ag mentioned above. Interference factor Dimodulo value; ； C6. Train Cg and Cd until each loss function is stationary, thus obtaining the supervisor C.

9. The intermediate frequency generation method for anti-interference of distorted communication complex signals based on complex generative adversarial networks according to claim 8, characterized in that, Step C6 specifically involves: C61. Take out a pair , ; C62. Will , Input Cd, and calculate the output value of Cd using M, Ag, Bg and Cd. By making If the value is smaller, update the parameters of Cd in reverse; C63. Will , Input Ag, and calculate using M, Ag, Bg, and Cd to obtain... , as well as By making , as well as If the value is smaller, update the parameters of Ag and Bg in reverse; C64. Repeat C62 and C63 until each loss function is stationary.

Images