Method and system for mixed electromagnetic threat type identification based on deep learning in strong confrontation environment

Abstract

The application relates to a mixed electromagnetic threat mode recognition method and system based on deep learning in a strong confrontation environment, and the method comprises the following steps: acquiring a complete data set required for training a deep learning model; training a signal reconstruction model to reconstruct the original signal sent by a sending end from the disturbed signal; performing blind detection on the difference value signal to obtain the number of radiation sources; performing a pretreatment operation on the signal data; carrying out blind source separation of the mixed radiation source signal; converting the input modulated signal sequence into an IQ signal or an AP signal; using a radiation source signal automatic blind modulation mode recognition model to perform automatic blind modulation mode recognition; using a channel coding feature extraction module to construct feature vectors of various coding modes; and using a radiation source channel coding type recognition model to perform channel coding type recognition, so that intelligent recognition and monitoring of mixed illegal electromagnetic signals and interference can be realized in a strong confrontation environment.

Description

Technical Field

[0001] This invention relates to the reconstruction, blind detection and separation of electromagnetic signals from radiation sources, as well as the field of intelligent signal pattern recognition, and particularly to a method and system for identifying hybrid electromagnetic threat patterns based on deep learning in a highly adversarial environment. Background Technology

[0002] Wireless communication is currently the most widely used communication method. With the continuous improvement of wireless communication techniques and the widespread use of radio signals, the electromagnetic environment has become increasingly complex and variable. The electromagnetic environment has become a core environmental element of informatization and one of the factors that must be considered under future informatization conditions. The characteristics of the electromagnetic environment include numerous radiation sources, high distribution density, wide distribution range, and severe signal overlap. Furthermore, signal modulation is complex, parameters change rapidly and frequently, and the overall threat level of signals is very high, highlighting the research value of electromagnetic environment signals.

[0003] The environment today is filled with various communication signals, remote sensing signals, radar signals, and other radio signals. Signal receiving equipment often receives multiple signals. Detecting suspicious signals in a specific frequency band within this mixed electromagnetic environment, monitoring and managing them, and using this as the basis for interference detection and wireless environment analysis has become a widely researched topic. For example, in military applications, communication reconnaissance is a crucial part of electronic warfare, used to monitor electromagnetic spectrum activity on the battlefield, identify enemy signals as threats, intercept useful information, and react accordingly to interference signals, thus achieving better anti-interference effects. This ensures the normal operation of friendly equipment and provides information support while simultaneously damaging enemy equipment and weakening its support capabilities. In civilian applications, to ensure the normal operation of legitimate communications, relevant departments need to monitor civilian communication signals, effectively manage the electromagnetic spectrum, and prevent illegal radiation sources from using and interfering with the wireless spectrum.

[0004] Blind signal detection and separation technology refers to the process of detecting the presence of received signals and separating blind sources in non-cooperative communication when the receiver has no prior knowledge or only limited prior knowledge. This technology combines algorithmic research on blind signal detection and separation with electromagnetic environment signal analysis and identification. It searches, detects, intercepts, and measures non-cooperative radio electromagnetic signals to obtain their modulation methods, source coding types, and encryption / decryption methods, providing a basis for monitoring and countering electromagnetic signals in highly contested environments. Because the number of radiation sources and receivers is not limited to one, resulting in random and variable irregular signals, the detection of threat signals will be uncertain, potentially leading to false alarms or missed alarms. Therefore, the models constructed for identifying electromagnetic environment threat targets must have a certain degree of universality and specificity. Summary of the Invention

[0005] To overcome the problems existing in the prior art, this invention provides a method and system for identifying hybrid electromagnetic threat patterns based on deep learning in a highly adversarial environment.

[0006] Terminology Explanation:

[0007] An autoencoder is a neural network that reproduces the input signal as accurately as possible. Its output vector has the same dimension as the input vector, and the training label is the input vector. It often learns the representation of the data or effectively encodes the original data through hidden layers according to a certain form of the input vector. A denoising autoencoder is based on the traditional autoencoder. It injects noise into the input and then uses the noisy "corrupted" samples to reconstruct the noise-free "clean" input.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows:

[0009] A deep learning-based method for identifying hybrid electromagnetic threat patterns in highly adversarial environments includes the following steps:

[0010] (1) Construct a wide-area spatial electromagnetic signal database and obtain the complete dataset required for training deep learning models from the wide-area spatial electromagnetic signal database;

[0011] (2) Use the complete dataset obtained in step (1) to train a signal reconstruction model, reconstruct the interfered signal to obtain the original signal sent by the transmitter, and the difference between the output value of the signal reconstruction model and the input is the possible interference, illegal electromagnetic signal and noise, and the difference is used as the object of subsequent detection and identification.

[0012] (3) Perform blind detection on the difference mentioned in step (2) to obtain the number of radiation sources;

[0013] (4) Perform preprocessing operations on the signal data, i.e., the difference obtained in step (2), including zero-mean processing and spherical processing.

[0014] (5) Based on the number of radiation sources obtained in step (3), perform blind source separation of mixed radiation source signals for signals that are simultaneously mixed in time, space and frequency without prior knowledge.

[0015] (6) For the input signal sequence, i.e. the signal sequence obtained after blind source separation in step (5), convert it into the form of In-Phase / Quadrature (IQ) signal or Amplitude / Phase (AP) signal, and generate the data set required for training, verifying and testing the automatic blind modulation mode identification model of radiation source signal in the form of dual-channel sequence.

[0016] (7) Use the dataset constructed in step (6) to train an automatic blind modulation mode identification model for radiation source signals, and perform automatic blind modulation mode identification on each radiation source signal obtained after separating the unknown mixed radiation source signals;

[0017] (8) Construct a channel coding feature extraction module based on data randomness, and propose channel coding codeword features from the aspects of codeword weight, run length, depth and autocorrelation of codeword sequence, and construct feature vectors for various coding methods;

[0018] (9) Use the constructed feature vector as a dataset to train a radiation source channel coding type identification model, and perform channel coding type identification on each radiation source signal obtained after separating unknown mixed radiation source signals.

[0019] According to a preferred embodiment of the present invention, in step (2), the signal reconstruction model is a denoising autoencoder, wherein the encoder is constructed using, but not limited to, a stacked three-layer Long Short-Term Memory (LSTM) network, and the output dimension of each layer gradually decreases. The decoder is also constructed using, but not limited to, a stacked three-layer LSTM network and a fully connected layer, and the output dimension of each layer of the LSTM gradually increases. The denoising autoencoder is used to reconstruct and train the uncontaminated ideal signals in the complete dataset obtained in step (1), so that for all signals, the reconstructed output of the signal reconstruction model is the original transmitted signal.

[0020] According to a preferred embodiment of the present invention, in step (3), the source number detection algorithm of the Gell circle criterion is used to perform blind detection on the difference in step (2) to obtain the number of radiation sources. The source number detection algorithm of the Gell circle criterion GDE(k) is shown in equation (I):

[0021]

[0022] In equation (I), k = 1, 2, ..., M-1, M is the number of observed signals, and ρ k Let ρ be the radius of the k-th Gell-Hill circle. i Let GDE(k) be the radius of the i-th Gell circle and D be the correction factor. The method for determining the Gell circle criterion is to start traversing from k=1 and stop traversing when GDE(k) is less than 0 for the first time. At this time, the estimated number of sources is k-1.

[0023] According to a preferred embodiment of the present invention, in step (4), the zero-mean processing is achieved by subtracting the average value of the received signal, as shown in equation (II):

[0024]

[0025] In equation (II), x(t) is the difference signal in step (2). It is the average value of the difference signal in step (2), w(t) represents the signal after zero mean processing, t represents time, and E(w) represents the mean of w(t); after such processing, E(w) = 0, and the process is also centered.

[0026] Spherization processing involves using a spherization matrix V to process the received signal after zero-mean processing, as shown in equation (III):

[0027] z(t)=Vw(t) (III)

[0028] In equation (III), z(t) represents the signal after spheroidization.

[0029] According to a preferred embodiment of the present invention, in step (5), based on the number of radiation sources obtained in step (3), a blind source separation algorithm is used to perform mixed source separation on signals that are simultaneously mixed in time, space, and frequency. The Fast Independent Component Analysis (Fast-ICA) algorithm based on negative entropy is used. The larger the negative entropy, the stronger the non-Gaussianity of the signal. Since the probability density function is unknown, the approximate expression of the negative entropy J(y) is expressed as equation (IV):

[0030] J(y) = {E[G(y)]} - E[G(y)] gauss )]] 2 (IV)

[0031] In equation (IV), E(·) is the function for finding the mean, G(·) is a non-quadratic even function, and y gauss y is a zero-mean Gaussian random variable with the same variance as y, where y is the separated signal.

[0032] According to a preferred embodiment of the present invention, in step (6), the I-path is the in-phase component, which is considered as the modulation signal in cos(2πf c The projection of f onto t) c The carrier frequency is Q, and the Q-path is the quadrature component, which can be considered as the modulating signal at -sin(2πf) / 2π. c The projection onto t), the mathematical relationship between the modulated signal and the IQ signal, is shown in equation (V):

[0033] x(t)=I(t)cos(2πf c t)+Q(t)sin(2πf c t) (V)

[0034] In equation (V), x(t) is the modulation signal, and I(t) and Q(t) are the I-channel signal and the Q-channel signal, respectively;

[0035] The mathematical relationship between the AP signal and the IQ signal is shown in equations (VI) and (VII):

[0036]

[0037] P = arctan(Q / I) (VII)

[0038] In equations (VI) and (VII), I and Q are the I-channel signal and the Q-channel signal, respectively.

[0039] According to a preferred embodiment of the present invention, in step (7), the automatic blind modulation mode identification model for radiation source signals uses, but is not limited to, a multi-stream model combining a convolutional neural network (CNN) and a long short-term memory network (LSTM) with feature interaction. The automatic blind modulation mode identification model for radiation source signals includes two branches with the same structure. The input of one branch is an IQ format data signal, and the input of the other branch is an AP format data signal. Each branch is composed of three CNNs of different sizes and two LSTMs with different numbers of hidden layers to extract signal features. Then, the features learned in the two branches are interacted in pairs, flattened, and connected to three fully connected layers. Finally, a Softmax layer is added to the end to map the features to the classification categories.

[0040] A training method combining center loss and softmax loss as loss functions is used to train the automatic blind modulation mode identification model for radiation source signals. This results in smaller intra-class distances, enhanced intra-class discriminative power, and a more robust judgment result for samples with large intra-class variations. The center loss function... As shown in equation (VIII):

[0041]

[0042] In equation (VIII), m represents the batch size. Let x represent the k-th class center of the feature. i This represents the features output from the last fully connected layer before the Softmax layer;

[0043] Relative to x i gradient and update equation The calculations are shown in equations (IX) and (X):

[0044]

[0045]

[0046] If the condition is met, δ(condition) = 1; otherwise, δ(condition) = 0.

[0047] The joint supervision formula of Center Loss and Softmax Loss used to train the automatic blind modulation mode recognition model of radiation source signals to obtain discriminative and separation features. As shown in equation (XI):

[0048]

[0049] In formula (XI), It is Softmax Loss. It is Center Loss, and the hyperparameter λ is used to balance the two loss functions.

[0050] According to a preferred embodiment of the present invention, in step (8), the feature vectors of various encoding methods include depth spectrum feature parameters, code weight feature parameters, code weight similarity feature parameters, run length feature parameters, longest run correlation feature parameters, and sequence autocorrelation feature parameters, with each sample data corresponding to one feature vector.

[0051] The process for calculating depth spectral feature parameters is as follows:

[0052] Define a codeword c = (c1, c2, ..., c L The differential operation of ) is shown in equation (XII):

[0053]

[0054] In equation (XII), D is used. L-1 (c) indicates that the codeword c has undergone L-1 differentiation operations; the coding depth C is defined after the differentiation operations of the codeword are defined. d Satisfying D i (c) = [0 L-i The smallest non-negative integer i is the depth of codeword c, denoted as C. d Where L is the code length of codeword c, and if such i does not exist, then C d =L, let C be a set of codewords of length L, and use C d (i) represents the number of codewords of depth i in set C, and the depth spectral feature parameter D of the signal is obtained. p As shown in equation (XIII):

[0055]

[0056] The calculation of code weight feature parameters is as follows: The number of non-zero elements in a random codeword of length L varies within the range [0, L]. Let the uniform distribution probability of the code weight of the random codeword be P, and the code weight distribution probability of the encoded codeword be p. i Where i∈[0,L], the code weight feature parameter W for different codes is obtained by statistical averaging. t As shown in equation (XIV):

[0057]

[0058] The process for calculating the code weight similarity feature parameters is as follows: the encoded codeword length is L, and the code weight probability distribution of the encoded codewords is W={p0,p1,…p L}, where the elements are the probabilities of a codeword appearing with code weights of 0, 1, ..., L, and the probability p′ of a random codeword appearing with code weights of 0, 1, ..., L. i Represented using combinations, as shown in equation (XV):

[0059]

[0060] The probability distribution of code weights of random codewords is obtained as w = {p′0, p′1, ..., p′}. L Define the code overlap similarity feature parameter W. d As shown in equation (XVI):

[0061]

[0062] And ρ W,w Let w represent the Pearson correlation coefficient between W and w, so we can extend it to obtain equation (XVII):

[0063]

[0064] The process for calculating run characteristic parameters is as follows:

[0065] T i Let a be the total number of runs contained in a codeword a of length L, where a = (a1, a2, ..., a...). L ), then T i As shown in equation (XVIII):

[0066]

[0067] The run length characteristic parameter T was obtained through statistical calculation. r As shown in equation (XIX):

[0068]

[0069] The process for calculating the longest run correlation characteristic parameters is as follows:

[0070] The longest run length in a sequence of length L ranges from [0, L]. Let the probability distribution of the longest run in a random sequence of this length be given by... Divide the channel-coded codewords into N partitions, each partition having a length of L. Calculate and count the longest run length in each partition, and then calculate the probability distribution of the longest run length using these counts, denoted as [Probability of Longest Run Path Distribution]. The correlation between the longest run length distribution of coded codewords and the longest run length distribution of random sequences is calculated using the Pearson correlation coefficient, thus obtaining the longest run correlation characteristic parameter L. t As shown in equation (XX):

[0071]

[0072] The process for calculating the autocorrelation feature parameters of a sequence is as follows:

[0073] The codeword is divided into N partitions, each partition being L in length, and the code block c = (c1, c2, ..., c...). L ), where d represents the logical shift delay, and D(d) represents the number of distinct elements in the test sequence and the shifted sequence, as shown in equation (XXI):

[0074]

[0075] Define the time delay, which is set to 2 here, to obtain the sequence autocorrelation feature parameter AT. c As shown in equation (XXII):

[0076]

[0077] According to a preferred embodiment of the present invention, the radiation source channel coding type identification model uses, but is not limited to, a typical convolutional neural network with three convolutional layers of different sizes, two fully connected layers, and one Softmax layer.

[0078] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement a method for identifying hybrid electromagnetic threat patterns based on deep learning in a highly adversarial environment.

[0079] A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of a deep learning-based method for identifying hybrid electromagnetic threat patterns under strong adversarial environments are disclosed.

[0080] A deep learning-based hybrid electromagnetic threat identification system for high-intensity adversarial environments, the system comprising:

[0081] Central equipment: Used for configuration management, equipment status management, database management, alarm information management, standard identification, and visualization of the hybrid electromagnetic threat identification system;

[0082] Front-end monitoring equipment: used to collect mixed electromagnetic signals in the area in real time, perform preprocessing, sampling, and transformation of sample signals, and send the preliminary analyzed data to the central equipment;

[0083] Power supply equipment: Used to supply power to central equipment and front-end monitoring equipment.

[0084] According to a preferred embodiment of the present invention, the central equipment includes an application server and a workstation;

[0085] The application server is used to: deploy the trained signal reconstruction model, the automatic blind modulation mode identification model of radiation source signal, and the channel coding type identification model to the high-performance application server to achieve high-speed inference of the task and obtain real-time identification results;

[0086] The workstation is used to: ensure the normal operation of the system, configure the operating environment for the system, monitor the operating status of the equipment, issue warning information for failed or damaged equipment, connect the central equipment and the front-end monitoring equipment using PCIe (Peripheral Component Interconnect Express) cable, visualize the system identification results, store the signals collected in real time in the database, and issue warning information in a timely manner for the identified threat signals.

[0087] According to a preferred embodiment of the present invention, the front-end monitoring device includes a radio frequency circuit, an analog-to-digital conversion circuit, and a digital circuit;

[0088] The radio frequency circuit is used for: after the signal is received by the antenna, it passes through the bandpass filter for frequency band selection and the low noise amplifier for amplification. The radio frequency signal is first down-converted to the intermediate frequency by the mixer, and the signal is further channel-selected and amplified in the intermediate frequency band to collect electromagnetic waves in the target frequency band.

[0089] Analog-to-digital conversion circuits include analog-to-digital converters (ADCs) and their matching circuits, used to convert analog signals into digital signals;

[0090] Digital circuits include complex programmable logic devices, field-programmable gate arrays, and digital signal processors;

[0091] Complex Programmable Logic Devices (CPLDs) are used for: starting up small systems within front-end equipment; combining Field Programmable Gate Arrays (FPGAs) and Digital Signal Processors (DSPs) to implement FFT algorithms; performing time-frequency domain transformation on acquired sample signals; and processing data for communication interfaces with central equipment.

[0092] The beneficial effects of this invention are:

[0093] 1. This invention abandons the prediction model that uses the correlation between time series to predict signals, and uses an illegal electromagnetic signal interference detection model based on ideal signal reconstruction to better solve the problem of detecting illegal electromagnetic signal interference with time-frequency overlap under low signal-to-noise ratio conditions.

[0094] 2. This invention addresses the pain point of general methods, which mainly focus on the detection and identification of single signals, and proposes a hybrid electromagnetic threat identification method and system that utilizes blind source signal detection and separation technology in complex real-world environments with multiple illegal electromagnetic signals and interference.

[0095] 3. Before blind source signal separation begins, the present invention performs preprocessing operations on the signal data, including zero-mean processing and spherical processing, which can reduce the amount of computation and simplify the separation process.

[0096] 4. This invention converts the input sequence features into the form of IQ or AP signals and performs normalization processing, enabling the model to extract effective deep features from them.

[0097] 5. This invention addresses the limitation of feature selection in signal coding type classification by proposing a channel coding type identification method based on data randomness theory. Experiments are conducted using codeword sequences with different signal-to-noise ratios to enhance the effectiveness of features and improve classification accuracy.

[0098] 6. This invention addresses the issue that traditional models use Softmax Loss to train networks. It proposes a method that combines Center Loss and Softmax Loss for training pattern recognition networks. This method provides a more robust judgment result for samples with large intra-class variations, reduces intra-class distances, increases inter-class distances, and learns a discriminative feature. The resulting deep features typically have strong discriminative power and are easy to train and optimize. Attached Figure Description

[0099] Figure 1This is a flowchart of the hybrid electromagnetic threat identification method based on deep learning under strong adversarial environment of the present invention.

[0100] Figure 2 This is a schematic diagram of the signal reconstruction model.

[0101] Figure 3 Scatter plot of two uniformly distributed signals:

[0102] Figure 4 Schematic diagrams of IQ signals for some common modulation schemes;

[0103] Figure 5 This is a schematic diagram illustrating the method of generating a modulated signal using IQ signals;

[0104] Figure 6 A schematic diagram of the structure of a radiation source signal modulation mode identification model;

[0105] Figure 7 This is a schematic diagram illustrating the distribution of two-dimensional depth features for different types of data when using Softmax Loss.

[0106] Figure 8 A schematic diagram illustrating the distribution of two-dimensional depth features for different types of data when using both Center Loss and Softmax Loss.

[0107] Figure 9 A schematic diagram showing the modulation scheme identification results with or without center loss;

[0108] Figure 10 A schematic diagram of the structure of the radiation source channel coding type identification model;

[0109] Figure 11 This is a block diagram of a hybrid electromagnetic threat identification system. Detailed Implementation

[0110] The present invention will be further described below with reference to the accompanying drawings and embodiments, but is not limited thereto.

[0111] Example 1

[0112] A deep learning-based method for identifying hybrid electromagnetic threat patterns in highly adversarial environments, such as Figure 1 As shown, it includes the following steps:

[0113] (1) Based on the characteristics of electromagnetic threats in electronic warfare scenarios, a wide-area space electromagnetic signal database is constructed. The wide-area space electromagnetic signal database includes ideal communication signals, communication signals interfered with in strong confrontation environments, a variety of commonly used analog and digital modulation signals, and a variety of commonly used channel coding signals. With the help of big data related technologies, appropriate and sufficient data in the database are selected for labeling, including modulation category, channel coding category, and signal-to-noise ratio. The complete dataset required for deep learning model training is obtained from the wide-area space electromagnetic signal database.

[0114] (2) Use the complete dataset obtained in step (1) to train a signal reconstruction model, reconstruct the interfered signal to obtain the original signal sent by the transmitter, and the difference between the output value of the signal reconstruction model and the input is the possible interference, illegal electromagnetic signal and noise. The difference is used as the object of subsequent detection and identification.

[0115] (3) The difference signal obtained in step (2) is an unknown mixed signal. Before blind source separation, it is necessary to perform blind detection on the difference in step (2) to obtain the number of radiation sources. This is because the blind signal separation algorithm is derived based on the fact that the number of signal sources is known.

[0116] (4) In blind source separation, the signal data, i.e., the difference obtained in step (2), is generally preprocessed before separation begins, including zero-mean processing and spherical processing, to simplify the separation process.

[0117] (5) Based on the number of radiation sources obtained in step (3), perform blind source separation of mixed radiation source signals for signals that are simultaneously mixed in time, space and frequency without prior knowledge.

[0118] (6) In order to train the automatic blind modulation mode identification model of radiation source signal in the next step, the input signal sequence, i.e. the signal sequence obtained after blind source separation in step (5), is converted into the form of In-Phase / Quadrature (IQ signal) or Amplitude / Phase (AP signal) to generate the data set required for training, verifying and testing the automatic blind modulation mode identification model of radiation source signal in the form of dual-channel sequence.

[0119] (7) Use the dataset constructed in step (6) to train an automatic blind modulation mode identification model for radiation source signals, and perform automatic blind modulation mode identification on each radiation source signal obtained after separating the unknown mixed radiation source signals;

[0120] (8) Construct a channel coding feature extraction module based on data randomness by using the relevant content of data randomness theory. Propose channel coding codeword features from the aspects of codeword weight, run length, depth and autocorrelation of codeword sequence, and construct feature vectors for various coding methods.

[0121] (9) Use the constructed feature vector as a dataset to train a radiation source channel coding type identification model, and perform channel coding type identification on each radiation source signal obtained after separating unknown mixed radiation source signals.

[0122] Example 2

[0123] The difference between the hybrid electromagnetic threat identification method based on deep learning in a highly adversarial environment described in Example 1 and the method described in Example 1 is as follows:

[0124] In step (2), the signal reconstruction model is a denoise autoencoder. The encoder is constructed using, but not limited to, a stack of three layers of Long Short-Term Memory (LSTM) networks, with the output dimension of each layer gradually decreasing. The decoder is also constructed using, but not limited to, a stack of three layers of LSTM networks and a fully connected layer, with the output dimension of each layer of the LSTM gradually increasing. The denoise autoencoder is used to reconstruct and train the uncontaminated ideal signals in the complete dataset obtained in step (1), so that for all signals, the reconstructed output of the signal reconstruction model is the original transmitted signal.

[0125] In step (3), the blind source separation algorithm for mixed radiation source signals is basically based on the premise that the number of source signals is known. Therefore, blind source detection is performed on the difference after signal reconstruction, and blind source detection is performed on the signal difference in step (2) to obtain the number of radiation sources, which lays the foundation for blind source separation. It can also run in a colored noise environment and use the algorithm with the best performance by mixing various interferences. The source number detection algorithm using but not limited to the Gerschgorin Disk Estimator (GDE) is used. The source number detection algorithm using but not limited to the Gerschgorin Disk Estimator (GDE) is used to blindly detect the difference in step (2) to obtain the number of radiation sources. The source number detection algorithm of the Gerschgorin Disk Estimator (GDE) is shown in equation (I):

[0126]

[0127] In equation (I), k = 1, 2, ..., M-1, M is the number of observed signals, and ρ k Let ρ be the radius of the k-th Gell-Hill circle. i Let GDE(k) be the radius of the i-th Gell circle and D be the correction factor. The method for determining the Gell circle criterion is to start traversing from k=1 and stop traversing when GDE(k) is less than 0 for the first time. At this time, the estimated number of sources is k-1.

[0128] In step (4), the signal data undergoes preprocessing operations, including zero-mean processing and spherical processing, in order to reduce the computational load and simplify the separation process.

[0129] Zero-mean processing is achieved by subtracting the average value of the received signal, as shown in equation (II):

[0130]

[0131] In equation (II), x(t) is the difference signal in step (2). It is the average value of the difference signal in step (2), w(t) represents the signal after zero mean processing, t represents time, and E(w) represents the mean of w(t); after such processing, E(w) = 0, and the process is also centered.

[0132] Spherization processing involves using a spherization matrix V to process the received signal after zero-mean processing, as shown in equation (III):

[0133] z(t)=Vw(t) (III)

[0134] In equation (III), z(t) represents the signal after spheroidization.

[0135] After spheroidizing, the degrees of freedom of the observed signal matrix decrease, which reduces the computational load of subsequent separation algorithms. For example... Figure 3 As shown, consider two independent random signals uniformly distributed between [10,30] and [-40,-20] respectively, with a signal length of 1000. The scatter plots of the two uniformly distributed source signals are shown below. Figure 3 As shown in (a), the scatter plot of the signal after linear mixing by a randomly generated 2*2 matrix is ​​as follows. Figure 3 As shown in (b), the scatter plot of the mixed signal after zero-mean processing is as follows: Figure 3 As shown in (c), the scatter plot of the signal after spheroidizing preprocessing following zero-mean processing is as follows. Figure 3 As shown in (d), after the signal is preprocessed, the blind signal separation algorithm only needs to find the separation matrix in the set of orthogonal matrices. For a two-dimensional matrix, this means finding a definite rotation angle.

[0136] In step (5), based on the number of radiation sources obtained in step (3), a blind source separation algorithm is used to perform mixed source separation on signals that are simultaneously mixed in time, space, and frequency. The algorithm used is not limited to Fast Independent Component Analysis (Fast-ICA) based on negative entropy. The larger the negative entropy, the stronger the non-Gaussianity of the signal. Since the probability density function is unknown, the approximate expression of the negative entropy J(y) is expressed as Equation (IV):

[0137] J(y) = {E[G(y)]} - E[G(y)] gauss )]] 2 (IV)

[0138] In equation (IV), E(·) is the function for finding the mean, G(·) is a non-quadratic even function, and y gauss y is a zero-mean Gaussian random variable with the same variance as y, where y is the separated signal.

[0139] In step (6), in order to train the automatic blind modulation mode identification model for radiation source signals in the next step, the input signal sequence is converted into the form of In-Phase / Quadrature (IQ) signal or Amplitude / Phase (AP) signal, and the data set required for training, validating and testing the model is generated in the form of a dual-channel sequence, where the I-channel is the in-phase component, which is regarded as the modulation signal in cos(2πf c The projection of f onto t) c The carrier frequency is Q, and the Q-path is the quadrature component, which can be considered as the modulating signal at -sin(2πf) / 2π. c The projection on t), Figure 4 This diagram illustrates some common modulation schemes for IQ signals, and shows a schematic diagram of how IQ signals generate modulated signals. Figure 5 As shown, the mathematical relationship between the modulated signal and the IQ signal is given by equation (V):

[0140] x(t)=I(t)cos(2πf c t)+Q(t)sin(2πf c t)(V)

[0141] In equation (V), x(t) is the modulation signal, and I(t) and Q(t) are the I-channel signal and the Q-channel signal, respectively;

[0142] The mathematical relationship between the AP signal and the IQ signal is shown in equations (VI) and (VII):

[0143]

[0144] P = arctan(Q / I) (VII)

[0145] In equations (VI) and (VII), I and Q are the I-channel signal and the Q-channel signal, respectively.

[0146] In step (7), the automatic blind modulation mode identification model can blindly identify the modulation type without prior knowledge of system parameters. Deep learning can "learn" features from the original data and automatically perform classification by using a classifier layer. The trained model can then identify the modulation mode of each radiation source signal obtained after separating the unknown mixed radiation source signal. The automatic blind modulation mode identification model of radiation source signal is a multi-stream model that uses, but is not limited to, a combination of a convolutional neural network (CNN) and a long short-term memory network (LSTM) to intelligently identify the modulation mode. The automatic blind modulation mode identification model of radiation source signal includes two branches with the same structure, such as... Figure 6 As shown, the input of one branch is IQ format data signal, and the input of the other branch is AP format data signal. Each branch is composed of three CNNs of different sizes and two LSTMs with different numbers of hidden layers to extract signal features. Then, the features learned in the two branches are paired and interacted, flattened and connected to three fully connected layers. Finally, a Softmax layer is added to the end to map the features to the classification categories.

[0147] Furthermore, a training method combining center loss and softmax loss as the loss function can be used to train the automatic blind modulation mode recognition model for radiation source signals. This results in smaller intra-class distances, enhanced intra-class discriminative power, and a more robust judgment result for samples with large intra-class variations. Figure 7 This represents the two-dimensional distribution of features after deep learning for different classes of data when only Softmax Loss is used. Figure 8 This represents the two-dimensional distribution of features after deep learning for different classes of data when using both Center Loss and Softmax Loss. Figure 9 This demonstrates that the automatic modulation recognition model using center loss achieves higher accuracy in identifying the modulation scheme of signal data compared to not using center loss. The center loss function... As shown in equation (VIII):

[0148]

[0149] In equation (VIII), m represents the batch size. Let x represent the k-th class center of the feature. iThis represents the features output from the last fully connected layer before the Softmax layer;

[0150] Relative to x i gradient and update equation The calculations are shown in equations (IX) and (X):

[0151]

[0152]

[0153] If the condition is met, δ(condition) = 1; otherwise, δ(condition) = 0.

[0154] The joint supervision formula of Center Loss and Softmax Loss used to train the automatic blind modulation mode recognition model of radiation source signals to obtain discriminative and separation features. As shown in equation (XI):

[0155]

[0156] In formula (XI), It is Softmax Loss. It is Center Loss, and the hyperparameter λ is used to balance the two loss functions.

[0157] In step (8), due to the limited selection of information sequence features in channel coding, a channel coding feature extraction module based on data randomness theory is proposed. This module extracts channel coding codeword features from several aspects, including code weight, run length, depth, and autocorrelation. The feature vectors for various coding methods include depth spectrum feature parameters, code weight feature parameters, code weight similarity feature parameters, run length feature parameters, longest run correlation feature parameters, and sequence autocorrelation feature parameters. Each sample data corresponds to one feature vector.

[0158] The process for calculating depth spectral feature parameters is as follows:

[0159] Define a codeword c = (c1, c2, ..., c L The differential operation of ) is shown in equation (XII):

[0160]

[0161] In equation (XII), D is used. L-1 (c) indicates that the codeword c has undergone L-1 differentiation operations; the coding depth C is defined after the differentiation operations of the codeword are defined. d Satisfying D i(c) = [0 L-i The smallest non-negative integer i is the depth of codeword c, denoted as C. d Where L is the code length of codeword c, and if such i does not exist, then C d =L, let C be a set of codewords of length L, and use C d (i) represents the number of codewords of depth i in set C, and the depth spectral feature parameter D of the signal is obtained. p As shown in equation (XIII):

[0162]

[0163] The calculation of code weight feature parameters is as follows: The number of non-zero elements in a random codeword of length L varies within the range [0, L]. Let the uniform distribution probability of the code weight of the random codeword be P, and the code weight distribution probability of the encoded codeword be p. i Where i∈[0,L], the code weight feature parameter W for different codes is obtained by statistical averaging. t As shown in equation (XIV):

[0164]

[0165] The process for calculating the code weight similarity feature parameters is as follows: the encoded codeword length is L, and the code weight probability distribution of the encoded codewords is W={p0,p1,…p L}, where the elements are the probabilities of a codeword appearing with code weights of 0, 1, ..., L, and the probability p′ of a random codeword appearing with code weights of 0, 1, ..., L. i Represented using combinations, as shown in equation (XV):

[0166]

[0167] The probability distribution of code weights of random codewords is obtained as w = {p′0, p′1, ..., p′}. L Define the code overlap similarity feature parameter W. d As shown in equation (XVI):

[0168]

[0169] And ρ W,w Let w represent the Pearson correlation coefficient between W and w, so we can extend it to obtain equation (XVII):

[0170]

[0171] The process for calculating run characteristic parameters is as follows:

[0172] T i Let a be the total number of runs contained in a codeword a of length L, where a = (a1, a2, ..., a...). L), then T i As shown in equation (XVIII):

[0173]

[0174] The run length characteristic parameter T was obtained through statistical calculation. r As shown in equation (XIX):

[0175]

[0176] The process for calculating the longest run correlation characteristic parameters is as follows:

[0177] The longest run length in a sequence of length L ranges from [0, L]. Let the probability distribution of the longest run in a random sequence of this length be given by... Divide the channel-coded codewords into N partitions, each partition having a length of L. Calculate and count the longest run length in each partition, and then calculate the probability distribution of the longest run length using these counts, denoted as [Probability of Longest Run Path Distribution]. The correlation between the longest run length distribution of coded codewords and the longest run length distribution of random sequences is calculated using the Pearson correlation coefficient, thus obtaining the longest run correlation characteristic parameter L. t As shown in equation (XX):

[0178]

[0179] The process for calculating the autocorrelation feature parameters of a sequence is as follows:

[0180] The codeword is divided into N partitions, each partition being L in length, and the code block c = (c1, c2, ..., c...). L ), where d represents the logical shift delay, and D(d) represents the number of distinct elements in the test sequence and the shifted sequence, as shown in equation (XXI):

[0181]

[0182] Define the time delay, which is set to 2 here, to obtain the sequence autocorrelation feature parameter AT. c As shown in equation (XXII):

[0183]

[0184] In step (9), a radiation source channel coding type identification model is trained using the constructed feature vector dataset, employing, but not limited to, typical convolutional neural network structures, such as... Figure 10As shown, the radiation source channel coding type identification model uses, but is not limited to, three convolutional layers of different sizes, two fully connected layers, and one softmax layer. Random samples are randomly selected and labeled at different signal-to-noise ratios, serving as the training and testing sets for the neural network. The trained model can then identify the channel coding type of each radiation source signal obtained after separating unknown mixed radiation source signals. Furthermore, a combination of center loss and softmax loss can be used as the loss function training method to reduce intra-class distance, enhance intra-class discriminative power, and provide a more robust judgment result for samples with large intra-class variations.

[0185] Example 3

[0186] A computer device includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the deep learning-based hybrid electromagnetic threat identification method in a highly adversarial environment as described in Embodiment 1 or 2.

[0187] Example 4

[0188] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the deep learning-based hybrid electromagnetic threat identification method under strong adversarial environment described in Embodiment 1 or 2.

[0189] Example 5

[0190] A deep learning-based hybrid electromagnetic threat identification system for high-intensity adversarial environments, such as Figure 11 As shown, the system includes:

[0191] Central equipment: Used for configuration management, equipment status management, database management, alarm information management, standard identification, and visualization of the hybrid electromagnetic threat identification system;

[0192] Front-end monitoring equipment: used to collect mixed electromagnetic signals in the area in real time, perform preprocessing, sampling, and transformation of sample signals, and send the preliminary analyzed data to the central equipment;

[0193] Power supply equipment: Used to power the central equipment and front-end monitoring equipment. It can be powered remotely via AC power or using proprietary power cable resources.

[0194] The central equipment includes application servers and workstations;

[0195] The application server is used to deploy the trained signal reconstruction model, the automatic blind modulation mode identification model of the radiation source signal, and the channel coding type identification model to the high-performance application server to achieve high-speed inference of the task and obtain real-time identification results;

[0196] The workstation is used to: ensure the normal operation of the system, configure the operating environment for the system, monitor the operating status of the equipment, issue warning information for failed or damaged equipment, connect the central equipment and the front-end monitoring equipment using PCIe (Peripheral Component Interconnect Express) cable, visualize information such as the system identification results, store the signals collected in real time in the database for later reference, and issue warning information in a timely manner for the identified threat signals.

[0197] Front-end monitoring equipment includes radio frequency circuits, analog-to-digital conversion circuits, and digital circuits;

[0198] The radio frequency (RF) circuit is used for: after the signal is received by the antenna, it passes through the bandpass filter for frequency band selection and the low-noise amplifier for amplification. The RF signal is then down-converted to the intermediate frequency (IF) by the mixer. In the IF band, the signal is further subjected to channel selection and amplification to collect electromagnetic waves in the target frequency band. The monitoring system should be configured with multi-channel RF circuits according to the application scenario to meet the requirements for the number of observed signals.

[0199] Analog-to-digital conversion circuits include analog-to-digital converters (ADCs) and their matching circuits, used to convert analog signals into digital signals;

[0200] Digital circuits include complex programmable logic devices, field-programmable gate arrays, and digital signal processors;

[0201] Complex Programmable Logic Devices (CPLDs) are used for: starting up small systems within front-end equipment; combining Field Programmable Gate Arrays (FPGAs) and Digital Signal Processors (DSPs) to implement FFT algorithms; performing time-frequency domain transformation on acquired sample signals; and processing data for communication interfaces with central equipment.

[0202] The above embodiments are merely exemplary embodiments of the present invention and are not intended to limit the present invention. The scope of protection of the present invention is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to the present invention within its spirit and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of the present invention.

[0203] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for identifying hybrid electromagnetic threat patterns based on deep learning under strong adversarial environments, characterized in that, Includes the following steps: (1) Construct a wide-area spatial electromagnetic signal database and obtain the complete dataset required for training deep learning models from the wide-area spatial electromagnetic signal database; (2) Use the complete dataset obtained in step (1) to train a signal reconstruction model, reconstruct the interfered signal to obtain the original signal sent by the transmitter, and the difference between the output value of the signal reconstruction model and the input is the possible interference, illegal electromagnetic signal and noise, and the difference is used as the object of subsequent detection and identification. (3) Perform blind detection on the difference mentioned in step (2) to obtain the number of radiation sources; (4) Perform preprocessing operations on the signal data, i.e., the difference obtained in step (2), including zero-mean processing and spherical processing. (5) Based on the number of radiation sources obtained in step (3), perform blind source separation of mixed radiation source signals for signals that are simultaneously mixed in time, space and frequency without prior knowledge. (6) For the input signal sequence, i.e. the signal sequence obtained after blind source separation in step (5), convert it into the form of In-Phase / Quadrature (IQ) signal or Amplitude / Phase (AP) signal, and generate the data set required for training, verifying and testing the automatic blind modulation mode identification model of radiation source signal in the form of dual-channel sequence. (7) Use the dataset constructed in step (6) to train an automatic blind modulation mode identification model for radiation source signals, and perform automatic blind modulation mode identification on each radiation source signal obtained after separating the unknown mixed radiation source signals; (8) Construct a channel coding feature extraction module based on data randomness, extract channel coding codeword features from codeword sequence code weight, run length, depth and autocorrelation, and construct feature vectors for various coding methods; (9) Use the constructed feature vector as a dataset to train a radiation source channel coding type identification model, and identify the channel coding type of each radiation source signal obtained after separating the unknown mixed radiation source signal.

2. The method for identifying hybrid electromagnetic threat patterns based on deep learning in a highly adversarial environment according to claim 1, characterized in that, In step (2), the signal reconstruction model is a denoising autoencoder. The encoder is constructed by stacking three layers of long short-term memory network LSTM, and the output dimension of each layer gradually decreases. The decoder is also composed of stacked three layers of long short-term memory network LSTM and a fully connected layer, and the output dimension of each layer of LSTM gradually increases. The denoising autoencoder is used to reconstruct and train the uncontaminated ideal signals in the complete dataset obtained in step (1), so that for all signals, the reconstruction output of the signal reconstruction model is the original transmitted signal.

3. The method for identifying hybrid electromagnetic threat patterns based on deep learning in a highly adversarial environment according to claim 1, characterized in that, In step (3), the source number detection algorithm of the Gell circle criterion is used to perform blind detection on the difference in step (2) to obtain the number of radiation sources. The source number detection algorithm of the Gell circle criterion GDE(k) is shown in equation (I): In equation (I), k = 1, 2, ..., M-1, M is the number of observed signals, and ρ k Let ρ be the radius of the k-th Gell-Hill circle. i Let GDE(k) be the radius of the i-th Gell circle and D be the correction factor. The method for determining the Gell circle criterion is to start traversing from k=1 and stop traversing when GDE(k) is less than 0 for the first time. At this time, the estimated number of sources is k-1.

4. The method for identifying hybrid electromagnetic threat patterns based on deep learning in a highly adversarial environment according to claim 1, characterized in that, In step (4), zero-mean processing is achieved by subtracting the average value of the received signal, as shown in equation (II): In equation (II), x(t) is the difference signal in step (2). It is the average value of the difference signal in step (2), w(t) represents the signal after zero mean processing, t represents time, and E(w) represents the mean of w(t); after such processing, E(w) = 0, and the process is also centered. Spherization processing involves using a spherization matrix V to process the received signal after zero-mean processing, as shown in equation (III): z(t)=Vw(t)(III) In equation (III), z(t) represents the signal after spheroidization.

5. The method for identifying hybrid electromagnetic threat patterns based on deep learning in a highly adversarial environment according to claim 1, characterized in that, In step (5), based on the number of radiation sources obtained in step (3), a blind source separation algorithm is used to perform mixed source separation on signals that are simultaneously mixed in time, space, and frequency. A fast independent component analysis algorithm based on negative entropy is used. The larger the negative entropy, the stronger the non-Gaussianity of the signal. Since the probability density function is unknown, the approximate expression of the negative entropy J(y) is expressed as equation (IV): J(y)={E[G(y)]}-E[G(y gauss )]] 2 (IV) In equation (IV), E(·) is the function for finding the mean, G(·) is a non-quadratic even function, and y gauss y is a zero-mean Gaussian random variable with the same variance as y, where y is the separated signal; In step (6), the I-path is the in-phase component, which is considered as the modulation signal in cos(2πf c The projection of f onto t) c The carrier frequency is Q, and the Q-path is the quadrature component, which can be considered as the modulating signal at -sin(2πf) / 2π. c The projection onto t), the mathematical relationship between the modulated signal and the IQ signal, is shown in equation (V): x(t)=I(t)cos(2πf c t)+Q(t)sin(2πf c t)(V) In equation (V), x(t) is the modulation signal, and I(t) and Q(t) are the I-channel signal and the Q-channel signal, respectively; The mathematical relationship between the AP signal and the IQ signal is shown in equations (VI) and (VII): P = arctan(Q / I) (VII) In equations (VI) and (VII), I and Q are the I-channel signal and the Q-channel signal, respectively.

6. The method for identifying hybrid electromagnetic threat patterns based on deep learning in a highly adversarial environment according to claim 1, characterized in that, In step (7), the automatic blind modulation mode identification model for radiation source signals is a multi-stream model jointly constructed by a convolutional neural network (CNN) and a long short-term memory network (LSTM) with feature interaction. The automatic blind modulation mode identification model for radiation source signals includes two branches with the same structure. The input of one branch is an IQ format data signal, and the input of the other branch is an AP format data signal. Each branch is composed of three CNNs of different sizes and two LSTMs with different numbers of hidden layers to extract signal features. Then, the features learned in the two branches are made to interact in pairs, flattened, and connected to three fully connected layers. Finally, a Softmax layer is added to the end to map the features to the classification categories. A training method combining center loss and softmax loss as the loss function is used to train the automatic blind modulation mode identification model of radiation source signals. This results in smaller intra-class distances, enhanced intra-class discriminative power, and a more robust judgment result for samples with large intra-class variations. The center loss function L... C As shown in equation (VIII): In equation (VIII), m represents the batch size. Let x represent the k-th class center of the feature. i This represents the features output from the last fully connected layer before the Softmax layer; L C Relative to x i gradient and update equation The calculations are shown in equations (IX) and (X): If the condition is met, δ(condition) = 1; otherwise, δ(condition) = 0. The joint supervision formula L of CenterLoss and Softmax Loss used to train the automatic blind modulation mode recognition model of radiation source signal to obtain discriminative features and separation features is shown in equation (XI): L=L S +λL C (XI) In equation (XI), L S It is Softmax Loss, L C It is Center Loss, and the hyperparameter λ is used to balance the two loss functions.

7. The method for identifying hybrid electromagnetic threat patterns based on deep learning in a highly adversarial environment according to claim 1, characterized in that, In step (8), the feature vectors of various encoding methods include depth spectrum feature parameters, code weight feature parameters, code weight similarity feature parameters, run length feature parameters, longest run correlation feature parameters, and sequence autocorrelation feature parameters. Each sample data corresponds to one feature vector. The radiation source channel coding type identification model includes three convolutional layers of different sizes, two fully connected layers, and one softmax layer.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the deep learning-based hybrid electromagnetic threat identification method under strong adversarial environment as described in any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the deep learning-based hybrid electromagnetic threat identification method under any of claims 1-7.

10. A hybrid electromagnetic threat identification system based on deep learning under strong adversarial environments, characterized in that: The system includes: Central equipment: Used for configuration management, equipment status management, database management, alarm information management, standard identification, and visualization of the hybrid electromagnetic threat identification system; Front-end monitoring equipment: used to collect mixed electromagnetic signals in the area in real time, perform preprocessing, sampling, and transformation of sample signals, and send the preliminary analyzed data to the central equipment; Power supply equipment: Used to supply power to central equipment and front-end monitoring equipment; The central equipment includes application servers and workstations; The application server is used to: deploy the trained signal reconstruction model, the automatic blind modulation mode identification model of radiation source signal, and the channel coding feature extraction module to the high-performance application server to achieve high-speed inference of the task and obtain real-time identification results; The workstation is used to: ensure the normal operation of the system, configure the operating environment for the system, monitor the operating status of the equipment, issue warning information for failed or damaged equipment, connect the central equipment and the front-end monitoring equipment using PCIe cables, visualize the system identification results, store the signals collected in real time in the database, and issue warning information in a timely manner for the identified threat signals. Front-end monitoring equipment includes radio frequency circuits, analog-to-digital conversion circuits, and digital circuits; The radio frequency circuit is used for: after the signal is received by the antenna, it passes through the bandpass filter for frequency band selection and the low noise amplifier for amplification. The radio frequency signal is first down-converted to the intermediate frequency by the mixer, and then the signal is further channel-selected and amplified in the intermediate frequency band to collect electromagnetic waves in the target frequency band. Analog-to-digital conversion circuits include analog-to-digital converters and their matching circuits, used to convert analog signals into digital signals; Digital circuits include complex programmable logic devices, field-programmable gate arrays, and digital signal processors; Complex programmable logic devices are used for: starting up small systems within front-end equipment; implementing FFT algorithms by combining field-programmable gate arrays and digital signal processors; performing time-frequency domain transformation on acquired sample signals; and processing data at the communication interface with the central equipment.

Images