A communication radiation source open set identification method and system

Abstract

The present invention discloses a method and system for open-set identification of communication radiation sources, and relates to the field of radio frequency signal processing technology. The method comprises: obtaining a radio frequency communication signal to be identified; preprocessing the radio frequency communication signal to be identified to obtain a radio frequency fingerprint feature vector; constructing an open-set identification model based on a classification-reconstruction network and an adversarial training algorithm; inputting the radio frequency fingerprint feature vector into the open-set identification model for identification to obtain an identification result; the identification result is that the radio frequency communication signal to be identified is an open-set signal or the radio frequency communication signal to be identified is a closed-set signal; the open-set signal indicates that the communication radiation source of the radio frequency communication signal to be identified is unknown; the closed-set signal indicates that the communication radiation source of the radio frequency communication signal to be identified is known. The present invention can successfully identify and classify closed-set signals while detecting open-set signals, thereby realizing individual identification of communication radiation sources in open-set scenarios.

Description

Technical Field

[0001] The present invention relates to the technical field of radio frequency signal processing, and in particular to a method and system for identifying an open set of communication radiation sources. Background Art

[0002] Communication emitter identification refers to a technology that identifies individual communication emitters based on the fingerprint characteristics of received radio frequency communication signals. Different emitters have subtle internal hardware differences, which are reflected in their emitted radio frequency signals and exhibit unique identifying characteristics, known as radio frequency fingerprints. This allows for identification of individual emitters. This technology is widely used in both civilian and military fields, particularly in military electronic reconnaissance. Currently, militaries around the world are equipped with a large number of electronic devices with identical production models and signal patterns. This makes traditional electronic reconnaissance based on signal pattern recognition difficult in tasks such as distinguishing individual targets. Furthermore, the inherent non-cooperative nature of communication in battlefield environments further exacerbates this difficulty. Therefore, communication reconnaissance based on communication emitter identification technology has broad application prospects and high military value. It can distinguish between emitters of the same production model and signal pattern, but with different individuals, thereby determining the number and scale of targets in a combat mission, and ultimately obtaining intelligence on enemy troop deployments and battlefield situation.

[0003] With the development of artificial intelligence technology, deep learning has begun to be applied to communication emitter identification. Using neural networks, it can comprehensively and deeply extract the characteristics of emitter signals, improving identification performance, which is a new research direction. In the article "Specific emitter identification against unreliable features interference based on time-series classification network structure" in the IEEE Access journal, issue 8 of 2020, YG Liu proposed a network model consisting of a deep bidirectional long short-term memory (Bidirectional Long Short-Term Memory, BiLSTM) network and a one-dimensional residual convolutional network. This network features dilated convolution and squeezed excitation blocks to extract RF fingerprint features hidden in baseband signals, thereby completing the task of individual communication emitter identification. In the article "Unsupervised specific emitter identification method using radio-frequency fingerprint embedded InfoGAN" published in the 15th issue of 2020 in the journal "IEEE Transactions on Information Forensics and Security", JLGong proposed a specific emitter identification method based on the Information Maximized Generative Adversarial Network (InfoGAN) and embedded radio frequency fingerprint features to address the practical problem that specific emitter identification is often used in non-cooperative communication scenarios. This method uses the bispectrum distribution of the signal as the embedded radio frequency fingerprint feature input into the InfoGAN for training. At the same time, noise that follows the Nakagami-m distribution is selected as the input latent vector of the InfoGAN, making the network training more adaptable to emitter signals with wireless channel characteristics. Experiments have shown that this method can effectively handle specific emitter identification under unsupervised conditions.

[0004] Deep learning-based methods for identifying individual communication emitters leverage the powerful fitting capabilities of neural networks to fully learn and extract the fingerprint features hidden in RF signals, improving identification efficiency. However, these methods are a relatively recent area of ​​research, and existing research primarily focuses on identifying individual emitters in closed-set scenarios. This assumes that the training dataset contains RF signals from all RF emitter categories. However, this assumption is easily violated in the real world, and RF signals of unknown categories (i.e., open-set signals) can easily appear during the identification process. However, relatively little research exists on how to detect open-set signals while correctly identifying signals of known categories (i.e., closed-set signals)—that is, how to identify individual communication emitters in open-set scenarios. Summary of the Invention

[0005] The purpose of the present invention is to provide a method and system for open set identification of communication radiation sources, so as to realize individual identification of communication radiation sources in an open set scenario.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A method for identifying an open set of communication radiation sources, the method comprising:

[0008] Acquiring a radio frequency communication signal to be identified;

[0009] Preprocessing the radio frequency communication signal to be identified to obtain a radio frequency fingerprint feature vector;

[0010] Construct an open set recognition model based on classification-reconstruction network and adversarial training algorithm;

[0011] The radio frequency fingerprint feature vector is input into the open set recognition model for recognition to obtain a recognition result; the recognition result is that the radio frequency communication signal to be identified is an open set signal or the radio frequency communication signal to be identified is a closed set signal; the open set signal indicates that the communication radiation source of the radio frequency communication signal to be identified is unknown; the closed set signal indicates that the communication radiation source of the radio frequency communication signal to be identified is known.

[0012] Optionally, preprocessing the radio frequency communication signal to be identified to obtain a radio frequency fingerprint feature vector specifically includes:

[0013] Performing bispectral analysis on the radio frequency communication signal to be identified to obtain a bispectral distribution vector;

[0014] A Radon transform is performed on the bispectral distribution vector to obtain a radio frequency fingerprint feature vector.

[0015] Optionally, the constructing of an open set recognition model based on a classification-reconstruction network and an adversarial training algorithm specifically includes:

[0016] Constructing a classification-reconstruction network; the classification-reconstruction network includes: an encoder, a decoder and a classifier; the encoder is connected to the classifier and the decoder respectively; the encoder and the classifier constitute a classification network; the encoder and the decoder constitute a reconstruction network;

[0017] Acquiring training samples and preprocessing the training samples to obtain closed-set signal feature vectors;

[0018] Inputting the closed set signal feature vector into the classification-reconstruction network, performing supervised closed set training on the classification network and performing reconstruction training on the reconstruction network, respectively, to obtain a trained classification-reconstruction network;

[0019] Generate an open-set signal feature vector based on the trained classification-reconstruction network simulation;

[0020] An open set recognition network is constructed, and the closed set signal feature vector and the open set signal feature vector are input into the open set recognition network, and an adversarial training algorithm is used for training to obtain an open set recognition model.

[0021] Optionally, the simulating generation of an open-set signal feature vector based on the trained classification-reconstruction network specifically includes:

[0022] Inputting the closed set signal feature vector into the trained classification-reconstruction network to obtain a reconstructed signal feature vector;

[0023] The closed-set signal eigenvector is excluded from the reconstructed signal eigenvector to obtain the open-set signal eigenvector.

[0024] Optionally, inputting the closed-set signal feature vector into the trained classification-reconstruction network to obtain a reconstructed signal feature vector specifically includes:

[0025] Encoding the closed set signal feature vector to obtain a latent vector;

[0026] Performing a linear combination on the latent vectors to obtain a combined latent vector;

[0027] The combined latent vector is decoded to obtain a reconstructed signal feature vector.

[0028] The present invention also provides a communication radiation source open set identification system, which is applied to the above method and includes:

[0029] A signal acquisition module, used to acquire the radio frequency communication signal to be identified;

[0030] A signal preprocessing module, configured to preprocess the radio frequency communication signal to be identified to obtain a radio frequency fingerprint feature vector;

[0031] Model building module, used to build open set recognition models based on classification-reconstruction networks and adversarial training algorithms;

[0032] A signal recognition module is used to input the radio frequency fingerprint feature vector into the open set recognition model for recognition to obtain a recognition result; the recognition result is that the radio frequency communication signal to be identified is an open set signal or the radio frequency communication signal to be identified is a closed set signal; the open set signal indicates that the communication radiation source of the radio frequency communication signal to be identified is unknown; the closed set signal indicates that the communication radiation source of the radio frequency communication signal to be identified is known.

[0033] Optionally, the signal preprocessing module specifically includes:

[0034] A bispectral analysis and processing submodule, configured to perform bispectral analysis on the radio frequency communication signal to be identified to obtain a bispectral distribution vector;

[0035] The Radon transform processing submodule is used to perform Radon transform processing on the bispectral distribution vector to obtain a radio frequency fingerprint feature vector.

[0036] Optionally, the model building module specifically includes:

[0037] A classification-reconstruction network construction submodule is used to construct a classification-reconstruction network; the classification-reconstruction network includes: an encoder, a decoder and a classifier; the encoder is connected to the classifier and the decoder respectively; the encoder and the classifier constitute a classification network; the encoder and the decoder constitute a reconstruction network;

[0038] A closed set signal feature vector determination submodule is used to obtain training samples and preprocess the training samples to obtain a closed set signal feature vector;

[0039] a classification-reconstruction network training submodule, configured to input the closed-set signal feature vector into the classification-reconstruction network, perform supervised closed-set training on the classification network and perform reconstruction training on the reconstruction network, respectively, to obtain a trained classification-reconstruction network;

[0040] An open-set signal feature vector generation submodule, configured to simulate and generate an open-set signal feature vector based on the trained classification-reconstruction network;

[0041] The open set recognition model construction submodule is used to construct an open set recognition network, and input the closed set signal feature vector and the open set signal feature vector into the open set recognition network, and use an adversarial training algorithm to train to obtain an open set recognition model.

[0042] Optionally, the open-set signal eigenvector generating submodule specifically includes:

[0043] a reconstructed signal feature vector generating unit, configured to input the closed set signal feature vector into the trained classification-reconstruction network to obtain a reconstructed signal feature vector;

[0044] The open-set signal eigenvector determining unit is configured to exclude the closed-set signal eigenvector from the reconstructed signal eigenvector to obtain the open-set signal eigenvector.

[0045] Optionally, the reconstructed signal feature vector generating unit specifically includes:

[0046] an encoding subunit, configured to encode the closed set signal feature vector to obtain a latent vector;

[0047] A linear combination subunit, configured to linearly combine the latent vectors to obtain a combined latent vector;

[0048] The decoding subunit is used to decode the combined latent vector to obtain a reconstructed signal feature vector.

[0049] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0050] The present invention constructs an open-set recognition model based on a classification-reconstruction network and an adversarial training algorithm. By training the classification network and the reconstruction network separately, the trained classification-reconstruction network can be used to simulate and generate an open-set signal feature vector based on the closed-set signal feature vector. Then, the closed-set signal feature vector and the simulated open-set signal feature vector are input into the open-set recognition network together, and the adversarial training algorithm is used for training to obtain an open-set recognition model. The open-set recognition model can successfully recognize and classify closed-set signals while detecting open-set signals, and the trained open-set recognition model has better noise resistance performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0052] Figure 1 A flow chart of the open set identification method for communication radiation sources provided by the present invention;

[0053] Figure 2 A network structure diagram of the classification-reconstruction network provided by the present invention;

[0054] Figure 3 This is the traditional automatic encoder network structure diagram;

[0055] Figure 4 The network structure diagram of the automatic encoder provided by the present invention;

[0056] Figure 5 A comparison chart of the open set recognition accuracy rate versus open set degree between the method provided by the present invention and the existing algorithm;

[0057] Figure 6 This is a scatter plot of the deep features extracted by the open set recognition network after the t-SNE algorithm reduces the dimension to a 2D plane;

[0058] Figure 7 A comparison of the open set recognition accuracy of the method provided by the present invention and the existing algorithm versus signal-to-noise ratio curves when the open set signal categories are 5 and 10;

[0059] Figure 8 This is a module diagram of the communication radiation source open set identification system provided by the present invention. DETAILED DESCRIPTION

[0060] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0061] The purpose of the present invention is to provide a method and system for open set identification of communication radiation sources, so as to realize individual identification of communication radiation sources in an open set scenario.

[0062] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0063] Figure 1 The flow chart of the communication radiation source open set identification method provided by the present invention is as follows: Figure 1 As shown, the method includes:

[0064] Step 101: Acquire a radio frequency communication signal to be identified.

[0065] Step 102: Preprocess the radio frequency communication signal to be identified to obtain a radio frequency fingerprint feature vector.

[0066] Step 103: Construct an open set recognition model based on the classification-reconstruction network and the adversarial training algorithm.

[0067] Step 104: Input the RF fingerprint feature vector into the open-set recognition model for recognition, obtaining a recognition result. The recognition result indicates that the RF communication signal to be identified is an open-set signal or a closed-set signal. The open-set signal indicates that the communication radiation source of the RF communication signal to be identified is unknown; the closed-set signal indicates that the communication radiation source of the RF communication signal to be identified is known. Furthermore, when the RF communication signal to be identified is a closed-set signal, the recognition result specifically indicates the communication radiation source category corresponding to the RF communication signal to be identified.

[0068] The above steps are discussed in detail below.

[0069] Step 102: Preprocessing the radio frequency communication signal to be identified to obtain a radio frequency fingerprint feature vector, specifically including:

[0070] Step 1021: Perform bispectral analysis on the radio frequency communication signal to be identified to obtain a bispectral distribution vector. The specific formula is:

[0071]

[0072] Where: B(ω1,ω2) represents the bispectral distribution vector, ω1 and ω2 represent the two-dimensional Fourier transform frequency, C 3x (τ1, τ2) represents the third-order cumulant of the radio frequency communication signal to be identified, and the superscript j represents a complex number.

[0073] Step 1022: Perform Radon transform (i.e., Radon transform) on the bispectral distribution vector to obtain a radio frequency fingerprint feature vector. Specifically, Radon transform is to calculate the linear integral of a two-dimensional function relative to an arbitrary straight line on the two-dimensional plane xOy. The specific formula is:

[0074]

[0075] Where: R(α,ρ) represents the RF fingerprint feature vector, α represents the angle between the line and the coordinate axis, ρ represents the vertical distance between the line and the origin, and δ(·) represents the feature function.

[0076] Step 103: Constructing an open set recognition model based on the classification-reconstruction network and the adversarial training algorithm specifically includes:

[0077] Step 1031: Construct a classification-reconstruction network; the classification-reconstruction network includes: an encoder, a decoder, and a classifier; the encoder is connected to the classifier and the decoder respectively; the encoder and the classifier constitute a classification network; the encoder and the decoder constitute a reconstruction network. The specific network structure of the classification-reconstruction network can be found in Figure 2 .

[0078] Step 1032: Acquire training samples and pre-process the training samples to obtain closed-set signal feature vectors. The method for pre-processing the training samples is the same as the method for pre-processing the radio frequency communication signal to be identified, and will not be described in detail here.

[0079] Step 1033: Input the closed-set signal feature vector into the classification-reconstruction network, perform supervised closed-set training on the classification network and perform reconstruction training on the reconstruction network, to obtain a trained classification-reconstruction network.

[0080] Specifically, the loss function for supervised closed-set training of the classification network is:

[0081]

[0082] Where: Θ E represents the network parameters of the encoder, Θ C represents the network parameters of the classifier, E(·) represents the mapping function corresponding to the encoding network, C(·) represents the mapping function corresponding to the classification network, K represents the category of the closed set signal feature vector, N represents the number of samples contained in each category of the closed set signal feature vector, y j represents the category label corresponding to the j-th closed set signal feature vector, x ij Represents the i-th sample in the j-th closed set signal feature vector.

[0083] The loss function for the reconstruction training of the reconstruction network is:

[0084]

[0085] Where: Θ D represents the network parameters of the decoder, L represents the number of network layers of the encoder or decoder (that is, the number of network layers of the encoder and decoder are equal, both are L layers), Represents the j-th closed set signal feature vector obtained by the l-th intermediate layer network of the encoder, Represents the reconstructed signal feature vector corresponding to the j-th closed set signal feature vector obtained by the l-th intermediate layer network of the decoder.

[0086] Step 1034: Generate an open-set signal feature vector based on the trained classification-reconstruction network simulation.

[0087] The step of simulating and generating an open-set signal feature vector based on the trained classification-reconstruction network specifically includes:

[0088] The closed-set signal feature vector is input into the trained classification-reconstruction network to obtain a reconstructed signal feature vector.

[0089] The closed-set signal eigenvector is excluded from the reconstructed signal eigenvector to obtain the open-set signal eigenvector.

[0090] Inputting the closed set signal feature vector into the trained classification-reconstruction network to obtain the reconstructed signal feature vector specifically includes:

[0091] Encode the closed set signal feature vector to obtain the latent vector where z ij The latent vector extracted for the middle layer The collection of

[0092] Perform linear combination on the latent vectors to obtain the combined latent vector The coefficient s j satisfy

[0093] Decode the combined latent vector to obtain the reconstructed signal feature vector

[0094] The step of excluding the closed-set signal feature vector from the reconstructed signal feature vector to obtain the open-set signal feature vector specifically includes:

[0095] The reconstructed signal feature vector that meets the set conditions is used as the open set signal feature vector.

[0096] The setting conditions are:

[0097]

[0098]

[0099] in: represents the reconstructed signal eigenvector, Represents the mth closed set signal feature vector obtained by the lth intermediate layer network of the encoder, represents the reconstructed signal feature vector obtained by the lth intermediate layer network of the decoder, and ε represents the threshold parameter.

[0100] Step 1035: construct an open set recognition network, input the closed set signal feature vector and the open set signal feature vector into the open set recognition network, and use an adversarial training algorithm to perform training to obtain an open set recognition model.

[0101] Specifically, the open set recognition network is trained with closed set supervision, where the labels of the closed set signals are 1, 2, ..., K, and the label of the open set signals is K+1. The open set recognition network is specifically a K+1 classification network, and specifically a convolutional neural network structure, where K is the category of the closed set signal feature vector.

[0102] The mathematical model trained using the adversarial training algorithm is:

[0103]

[0104] Where: J(·) represents the loss function of the open set recognition network, X represents the input vector, Δx represents the perturbation vector, y represents the category of the input vector, θ represents the parameters of the open set recognition network, represents the expected value of the loss function of the open set recognition network, θ opt represents the parameters of the optimal open set recognition network.

[0105] The present invention is discussed in detail below from four aspects: signal preprocessing, classification-reconstruction network design, simulation generation of open set signal feature vectors, and adversarial training.

[0106] 1. Signal preprocessing

[0107] As a lowest-order high-order spectral analysis (HOSA), bispectral analysis is superior in processing non-Gaussian and non-stationary signals. The process of bispectral analysis of a signal is essentially a two-dimensional Fourier transform of the third-order cumulant of the signal, which can be expressed as follows:

[0108]

[0109] Among them, ω1 and ω2 represent the two-dimensional Fourier transform frequency, C 3x (τ1,τ2) represents the third-order cumulant, which can be expressed as:

[0110] C 3x (τ1,τ2)=E[x(t)x(t+τ1)x(t+τ2)] (2)

[0111] The bispectral distribution can well characterize the RF fingerprint characteristics of communication radiation source signals. However, directly using the two-dimensional bispectral distribution for subsequent processing will introduce a huge amount of computation and reduce processing efficiency. Therefore, the Radon transform is introduced to project the two-dimensional bispectral distribution into a one-dimensional feature vector. The Radon transform calculates the linear integral of a two-dimensional function with respect to an arbitrary line on the two-dimensional plane xOy. For the bispectral distribution B(ω1,ω2), its Radon transform can be expressed as:

[0112]

[0113] Where α represents the angle between the line and the coordinate axis, and ρ represents the vertical distance between the line and the origin. The characteristic function δ is defined as:

[0114]

[0115] To ensure that the integration is performed along the straight line ρ = ω1cosθ + ω2sinθ. Here, when performing the Radon transform, ρ = 0 and α = 3π / 4 are taken, that is, a linear integration is performed with respect to the straight line ω1 = ω2, and the result is normalized.

[0116] 2. Classification-Reconstruction Network Design

[0117] Figure 2 The network structure diagram of the classification-reconstruction network provided by the present invention is as follows: Figure 2 As shown in Figure 1, the classification-reconstruction network consists of three sub-networks: encoder, decoder, and classifier. Using the feature vectors obtained by preprocessing K types of signal samples to train the network specifically involves two steps: (1) supervised closed-set training of the classification network composed of the encoder and classifier; and (2) reconstruction training of the reconstruction network composed of the encoder and decoder (i.e., the autoencoder network).

[0118] The K-type eigenvectors can be expressed as {X1,X2,…,X K}, where X j ={x 1j ,x 2j ,…,x Nj}, that is, each type of feature vector contains N samples. The category label corresponding to the feature vector can be expressed as {y1,y2,…,y K}, where y j Is a one-hot encoding vector. When conducting supervised closed set training on the classification network composed of the encoder and the classifier, it is assumed that the network parameters of the encoder and the classifier are Θ E and Θ C , the loss function of network training can be expressed as:

[0119]

[0120] Among them, E(·) and C(·) represent the mapping functions corresponding to the encoding network and the classification network, respectively.

[0121] When the reconstruction network composed of the encoder and decoder is trained, the input feature vector is learned through the encoder’s feature learning to obtain the latent vector {Z1, Z2, …, Z K}, where Z j ={z 1j ,z 2j ,…,z Nj}, the reconstructed feature vector is obtained through the decoder in The purpose of reconstruction network training is to obtain the best latent vector Z j , so that the decoder can restore the input feature vector as much as possible. The latent vector Z at this time j It can best characterize the essential latent features of the input feature vector.

[0122] The essence of the reconstruction network is an automatic encoding network. Its traditional network structure and the network structure proposed in this invention are as follows: Figure 3 、 4 shown. Figure 3 This is the traditional automatic encoder network structure diagram, such as Figure 3 As shown in Figure 2, the traditional network structure first requires the encoder to process the latent vector through multiple layers of the network, and then reconstruct it through the decoder. However, the processing of each intermediate layer of the network carries the risk of feature loss, which weakens the representation ability of the latent vector. Figure 4 The network structure diagram of the automatic encoder provided by the present invention is as follows: Figure 4 As shown in Figure 1, the present invention improves the traditional autoencoder network structure, allowing each encoder layer to compress and extract latent vectors, which are then reconstructed by each decoder layer. This ensures that the latent vectors retain the greatest possible feature information and optimally represent the essential implicit features of the input vector.

[0123] For the initial feature vector sample The result of the processing of the first intermediate layer network can be expressed as:

[0124]

[0125] where f l (·) and h l (·) represent the feature mapping function and nonlinear dimensionality reduction function of the encoder layer l respectively. The latent vector is extracted through each intermediate layer network of the encoder. Where L is the number of network layers, and then the decoder is used to reconstruct and restore. The processing result of the intermediate layer network of the decoder layer l can be expressed as:

[0126]

[0127] where g l (·)and They represent the feature mapping function and nonlinear dimension-raising function of the decoder layer l respectively.

[0128] During the training and reconstruction of the network, it is assumed that the network parameters of the decoder are Θ D , then the loss function of the reconstruction network can be expressed as:

[0129]

[0130] In summary, the mathematical model of the classification-reconstruction network training process can be expressed as follows:

[0131]

[0132] 3. Simulation Generation of Open-Set Signal Eigenvectors

[0133] The space consisting of all known class samples is called the closed space U c , the space composed of all unknown class samples is called the open set space U o , how to identify the elements in the open space based on the closed space is the problem to be solved in this section. In fact, related open set identification research shows that the elements in the open space and the closed space contain the same characteristic quantities. In other words, the open space and the closed space are both contained in the same vector space U = U c ∪U o , both are subspaces of this vector space. Specifically, the open-set signal feature space, consisting of open-set signal eigenvectors, and the closed-set signal feature space, consisting of closed-set signal eigenvectors, belong to a larger vector space: the signal feature space, consisting of RF signal eigenvectors. Both spaces share the same characteristic elements that characterize RF signals.

[0134] Based on the above analysis, open-set signal feature vectors can be obtained by linearly combining different closed-set signal feature vectors. However, the complexity of the original signal feature vectors causes the closed-set space to exhibit highly nonlinear characteristics, so linear combinations of its elements are bound to produce unsatisfactory results. Considering that the latent vectors obtained by encoding and compressing the signal feature vectors can independently represent the category feature information, linearly combining and decoding the latent vectors may yield ideal open-set signal feature vectors. The specific process is as follows:

[0135] (1) Using the trained classification-reconstruction network to obtain the latent vector of K closed set signal feature vectors where z ij The latent vector extracted for the middle layer The collection of

[0136] (2) Through Z j The linear combination of The coefficient s j satisfy

[0137] (3) The latent vector Z u The reconstructed signal feature vector is obtained by inputting it into the decoder according to formula (7)

[0138] Some reconstructed signal samples may belong to the closed signal space, so we need to try to exclude them. At this time for The following two conditions must be met:

[0139]

[0140]

[0141] Only then can it be determined that it belongs to the open signal feature space.

[0142] 4. Adversarial Training

[0143] The feature vectors of the closed-set signal and the feature vectors of the simulated open-set signal are now fed into the open-set recognition network for closed-set supervised training. The labels of the closed-set signal are 1, 2, ..., K, and the label of the open-set signal is K+1. In practice, due to the subtle differences in the characteristics of the radiation source signal, it is easily affected by noise and misidentification occurs. To address this, an adversarial training algorithm is used to improve the noise resistance of the classification network. For the original input vector {X1, X2, ..., X K+1}, and its corresponding category label is {y1,y2,…,y K+1}, at this time artificially add a perturbation vector (adversarial sample) Δx j , that is, the input vector is X j +Δx j When , the classification network output is forced to be y j , so that the trained classification network can obtain correct classification results in the presence of noise interference. The ideal Δx should maximize the network loss value, and the network trained based on this is the most robust to noise. In summary, the mathematical model of adversarial training can be expressed by formula (12):

[0144]

[0145] Among them, θ is the parameter of the open set recognition network, and J(·) is the loss function of the open set recognition network. opt The network loss value should be maximized, so the optimal subtle perturbation can be constructed through the gradient principle:

[0146] Δx=▽ X J(X,y;θ)(13)

[0147] Among them X J represents the gradient of the loss value J(X, y; θ) with respect to the input vector X. It is the direction in which J(X, y; θ) rises fastest. At this time, Δx can maximize J(X+Δx, y; θ). To ensure that Δx is a subtle perturbation, equation (13) must be normalized:

[0148]

[0149] The method provided by the present invention is further verified by a specific example below.

[0150] (1) Five USRP devices of the same model were selected as individual communication radiation sources, and one USRP was selected as the receiving device, and both were connected to a PC. The PC defined the transmission and reception of the communication system through GNURadio, and finally received and obtained five types of closed-set RF signal data. In addition, the RF signals generated by 10 USRPs of the same model were collected as 10 types of open-set RF signal data. The operating frequency of the transmitter signal was set to 2.4GHz, and the receiver sampled the received signal at a sampling frequency of 16MHz. At the same time, the signal modulation mode was set to QPSK, and the bandwidth was 1.2MHz. The data collected for each type of signal was framed. For the closed-set signal, 20,000 frames of signal frames were obtained, and 80% were randomly selected as training samples, 10% as verification samples, and 10% as test samples. Each frame contained 256 sampling data points. For the open-set signal, 2,000 frames of signal frames were obtained, and all of them were used as test samples.

[0151] (2) Perform bispectral analysis on each frame of signal data to obtain a two-dimensional bispectral distribution, and then project the two-dimensional bispectral distribution into a one-dimensional feature vector through Radon transform to complete signal preprocessing.

[0152] (3) The feature vector is fed into the classification-reconstruction network, and its loss function is calculated according to equations (5), (8), and (9), and network training is carried out. After the training is completed, the latent vector that can optimally represent the implicit features of various closed-set signals is obtained. The latent vector is then linearly combined and reconstructed to simulate the open-set signal feature vector.

[0153] (4) Finally, the feature vectors of the known closed-set signal and the open-set signal are fed into the open-set recognition network for supervised training, and an adversarial training algorithm is used to improve the network's noise resistance. The trained open-set recognition network can detect open-set signal samples and successfully identify and classify closed-set signals.

[0154] Figure 5 This is a comparison of the open set recognition accuracy of the method provided by the present invention and the existing algorithm as a function of the degree of openness. The existing algorithm is the algorithm proposed in the journal "System Engineering and Electronics Technology" No. 41, 2019, "Open Set Recognition of Signal Modulation Methods Based on Generative Adversarial Networks". Here, we first define "openness" to describe the degree of openness of open set recognition. The specific definition is shown in formula (15):

[0155]

[0156] Among them C TR 、CTE Represent the set of data categories in the training set and the test set respectively. Then the accuracy is used as the evaluation index of the open set recognition performance, and the Accuracy-Openness relationship curve of the algorithm is drawn. TR Fixed to 5, C TE Take 5, 6, 7, 8, 9, and 10 respectively. Figure 5 It can be seen that as the degree of openness increases, the open set recognition accuracy of the present invention remains at a high level. Even when the number of open set signal classes increases to 10, the recognition accuracy of the algorithm can still reach more than 90%, showing good open set recognition performance. At the same time, compared with the existing open set recognition algorithm, the recognition accuracy of the algorithm under different open set degrees achieves the highest value, and is improved by an average of 5% to 15% compared with the existing algorithm. In addition, as the degree of open set increases, the recognition accuracy of the present invention is more significantly superior to that of the existing algorithm. This shows that the present invention can better adapt to the task of individual identification of communication radiation sources in open set scenarios.

[0157] In order to more intuitively observe the open set recognition performance of the algorithm in this paper, the t-distributed Stochastic Neighbor Embedding (t-SNE) algorithm is used to reduce the dimension and display the deep features extracted by the open set recognition network. Figure 6 This is a scatter plot of the deep features extracted by the open set recognition network after being reduced to a 2D plane using the t-SNE algorithm. Figure 6 The figure shows the scatter plot of the deep features extracted by the open set recognition network after the t-SNE algorithm reduces the dimension to a 2-dimensional plane when the number of open set signal categories is 5. The horizontal and vertical axes in the figure are the two different dimensions of the features after dimensionality reduction. Figure 6 It can be concluded that for the five types of closed-set signals, their deep features after dimensionality reduction are distributed as five different clusters in the two-dimensional scatter plot, and show the characteristics of intra-class aggregation and inter-class distinction. At the same time, the deep features of the open-set signal after dimensionality reduction are distributed as a large cluster in the two-dimensional scatter plot, and are separated from the other five clusters. This shows that the trained open-set recognition network can effectively distinguish different types of closed-set signals and has good closed-set signal classification and recognition capabilities. On the other hand, the algorithm proposed in the present invention generates relatively ideal simulated open-set signal samples. After being sent to the open-set recognition network for training, the network can successfully distinguish open-set signals from other types of closed-set signals and has good open-set signal detection and recognition capabilities.

[0158] Figure 7 This is a comparison chart of the open set recognition accuracy rate versus signal-to-noise ratio curves of the method provided by the present invention and the existing algorithm when the open set signal categories are 5 and 10. Figure 7In the experiment, the collected raw signal data was sent to MATLAB to add Gaussian white noise, and the signal-to-noise ratio was set to 0dB, 3dB, ..., 24dB. During the experiment, the number of open set signal categories was set to 5 and 10 respectively, and the robustness of the algorithm to noise was tested under two different open set degrees. Figure 7 It can be concluded that when the open set signal category is 5, the open set recognition accuracy of the present invention reaches over 80% at a 3dB signal-to-noise ratio, and over 90% at a 9dB signal-to-noise ratio. This shows that the method provided by the present invention can achieve a relatively high open set recognition accuracy in a low signal-to-noise ratio environment. When the open set signal category is 10, the open set recognition accuracy of the present invention reaches over 80% at a 6dB signal-to-noise ratio, and over 90% at a 15dB signal-to-noise ratio. As the number of open set signal categories increases, open set recognition becomes more difficult, causing the signal-to-noise ratio corresponding to the same recognition accuracy to shift backward. However, in this case, the open set recognition accuracy of the present invention reaches over 75% at a 3dB signal-to-noise ratio, and over 85% at a 9dB signal-to-noise ratio, still achieving relatively high open set recognition performance. At the same time, compared with existing algorithms, the present invention achieves the highest recognition accuracy under different signal-to-noise ratio environments. Especially in low signal-to-noise ratio environments, the open-set recognition performance of existing algorithms deteriorates to varying degrees, but the recognition accuracy of the present invention does not decrease significantly, remaining at a relatively high level of 70%. This demonstrates that the recognition performance of the method provided by the present invention is robust to noise.

[0159] The present invention also provides a communication radiation source open set identification system, which corresponds to the above method. Figure 8 The module diagram of the communication radiation source open set identification system provided by the present invention is as follows: Figure 8 As shown, the system includes:

[0160] The signal acquisition module 801 is used to acquire the radio frequency communication signal to be identified.

[0161] The signal preprocessing module 802 is configured to preprocess the radio frequency communication signal to be identified to obtain a radio frequency fingerprint feature vector.

[0162] The model construction module 803 is used to construct an open set recognition model based on the classification-reconstruction network and the adversarial training algorithm.

[0163] The signal recognition module 804 is used to input the radio frequency fingerprint feature vector into the open set recognition model for recognition to obtain a recognition result; the recognition result is that the radio frequency communication signal to be identified is an open set signal or the radio frequency communication signal to be identified is a closed set signal; the open set signal indicates that the communication radiation source of the radio frequency communication signal to be identified is unknown; the closed set signal indicates that the communication radiation source of the radio frequency communication signal to be identified is known.

[0164] Furthermore, the signal preprocessing module 802 specifically includes:

[0165] The bispectral analysis and processing submodule is used to perform bispectral analysis on the radio frequency communication signal to be identified to obtain a bispectral distribution vector.

[0166] The Radon transform processing submodule is used to perform Radon transform processing on the bispectral distribution vector to obtain a radio frequency fingerprint feature vector.

[0167] Furthermore, the model building module 803 specifically includes:

[0168] The classification-reconstruction network construction submodule is used to construct a classification-reconstruction network; the classification-reconstruction network includes: an encoder, a decoder and a classifier; the encoder is connected to the classifier and the decoder respectively; the encoder and the classifier constitute a classification network; the encoder and the decoder constitute a reconstruction network.

[0169] The closed set signal feature vector determination submodule is used to obtain training samples and preprocess the training samples to obtain the closed set signal feature vector.

[0170] The classification-reconstruction network training submodule is used to input the closed set signal feature vector into the classification-reconstruction network, perform supervised closed set training on the classification network and perform reconstruction training on the reconstruction network, and obtain a trained classification-reconstruction network.

[0171] The open-set signal feature vector generation submodule is used to simulate and generate the open-set signal feature vector according to the trained classification-reconstruction network.

[0172] The open set recognition model construction submodule is used to construct an open set recognition network, and input the closed set signal feature vector and the open set signal feature vector into the open set recognition network, and use an adversarial training algorithm to train to obtain an open set recognition model.

[0173] Furthermore, the open set signal feature vector generation submodule specifically includes:

[0174] The reconstructed signal feature vector generating unit is used to input the closed set signal feature vector into the trained classification-reconstruction network to obtain the reconstructed signal feature vector.

[0175] The open-set signal eigenvector determining unit is configured to exclude the closed-set signal eigenvector from the reconstructed signal eigenvector to obtain the open-set signal eigenvector.

[0176] Furthermore, the reconstructed signal feature vector generating unit specifically includes:

[0177] The encoding subunit is used to encode the closed set signal feature vector to obtain a latent vector.

[0178] The linear combination subunit is used to perform linear combination on the latent vectors to obtain a combined latent vector.

[0179] The decoding subunit is used to decode the combined latent vector to obtain a reconstructed signal feature vector.

[0180] The present invention discloses a method and system for open-set identification of communication radiation sources based on a classification-reconstruction network and an adversarial training algorithm. The method first performs bispectral analysis on the signal and performs Radon projection transformation to obtain a feature vector that characterizes the radio frequency fingerprint. Then, a classification-reconstruction network is proposed, which uses the feature vector of the closed-set signal to simulate and generate the feature vector of the open-set signal. Finally, both are sent to the open-set identification network for supervised training, and it is proposed to use an adversarial training algorithm to improve the noise resistance of the open-set identification network and enhance its robustness to noise. After the open-set identification network is trained, it can detect open-set signal samples and successfully identify and classify closed-set signals. The present invention can be applied to non-cooperative communication reconnaissance systems, and can also be used in corresponding software radio and other systems.

[0181] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0182] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A method for identifying open sets of communication radiation sources, characterized in that: The method comprises: Acquiring a radio frequency communication signal to be identified; Preprocessing the radio frequency communication signal to be identified to obtain a radio frequency fingerprint feature vector; An open set recognition model is constructed based on the classification-reconstruction network and adversarial training algorithm, specifically including: Constructing a classification-reconstruction network; the classification-reconstruction network includes: an encoder, a decoder and a classifier; the encoder is connected to the classifier and the decoder respectively; the encoder and the classifier constitute a classification network; the encoder and the decoder constitute a reconstruction network; Acquiring training samples and preprocessing the training samples to obtain closed-set signal feature vectors; Inputting the closed set signal feature vector into the classification-reconstruction network, performing supervised closed set training on the classification network and performing reconstruction training on the reconstruction network, respectively, to obtain a trained classification-reconstruction network; Generate an open-set signal feature vector based on the trained classification-reconstruction network simulation, specifically including: Inputting the closed set signal feature vector into the trained classification-reconstruction network to obtain a reconstructed signal feature vector; Excluding the closed-set signal eigenvector from the reconstructed signal eigenvector to obtain the open-set signal eigenvector specifically includes: The reconstructed signal eigenvector that meets the set conditions is used as the open set signal eigenvector, and the set conditions are: in: represents the reconstructed signal feature vector, Represents the mth closed set signal feature vector obtained by the lth intermediate layer network of the encoder, represents the reconstructed signal feature vector obtained by the lth intermediate layer network of the decoder, ε represents the threshold parameter, E(·) represents the mapping function corresponding to the encoding network, C(·) represents the mapping function corresponding to the classification network, N represents the number of samples contained in each type of closed set signal feature vector, y j represents the category label corresponding to the j-th closed set signal feature vector, L represents the number of network layers of the encoder or decoder, that is, the number of network layers of the encoder and decoder are equal, both are L layers; Constructing an open set recognition network, inputting the closed set signal feature vector and the open set signal feature vector into the open set recognition network, and training the network using an adversarial training algorithm to obtain an open set recognition model; The radio frequency fingerprint feature vector is input into the open set recognition model for recognition to obtain a recognition result; the recognition result is that the radio frequency communication signal to be identified is an open set signal or the radio frequency communication signal to be identified is a closed set signal; the open set signal indicates that the communication radiation source of the radio frequency communication signal to be identified is unknown; the closed set signal indicates that the communication radiation source of the radio frequency communication signal to be identified is known.

2. The communication radiation source open set identification method according to claim 1, characterized in that: The preprocessing of the radio frequency communication signal to be identified to obtain a radio frequency fingerprint feature vector specifically includes: Performing bispectral analysis on the radio frequency communication signal to be identified to obtain a bispectral distribution vector; A Radon transform is performed on the bispectral distribution vector to obtain a radio frequency fingerprint feature vector.

3. The method for identifying an open set of communication radiation sources according to claim 1, wherein: Inputting the closed-set signal feature vector into the trained classification-reconstruction network to obtain a reconstructed signal feature vector specifically includes: Encoding the closed set signal feature vector to obtain a latent vector; Performing a linear combination on the latent vectors to obtain a combined latent vector; The combined latent vector is decoded to obtain a reconstructed signal feature vector.

4. A communication radiation source open set identification system, characterized in that: The system comprises: A signal acquisition module, used to acquire the radio frequency communication signal to be identified; A signal preprocessing module, configured to preprocess the radio frequency communication signal to be identified to obtain a radio frequency fingerprint feature vector; The model building module is used to build an open set recognition model based on the classification-reconstruction network and the adversarial training algorithm. Specifically, it includes: A classification-reconstruction network construction submodule is used to construct a classification-reconstruction network; the classification-reconstruction network includes: an encoder, a decoder and a classifier; the encoder is connected to the classifier and the decoder respectively; the encoder and the classifier constitute a classification network; the encoder and the decoder constitute a reconstruction network; A closed set signal feature vector determination submodule is used to obtain training samples and preprocess the training samples to obtain a closed set signal feature vector; a classification-reconstruction network training submodule, configured to input the closed-set signal feature vector into the classification-reconstruction network, perform supervised closed-set training on the classification network and perform reconstruction training on the reconstruction network, respectively, to obtain a trained classification-reconstruction network; The open-set signal feature vector generation submodule is used to simulate and generate the open-set signal feature vector according to the trained classification-reconstruction network, specifically including: a reconstructed signal feature vector generating unit, configured to input the closed set signal feature vector into the trained classification-reconstruction network to obtain a reconstructed signal feature vector; The open set signal eigenvector determining unit is configured to exclude the closed set signal eigenvector from the reconstructed signal eigenvector to obtain the open set signal eigenvector, specifically comprising: The reconstructed signal eigenvector that meets the set conditions is used as the open set signal eigenvector, and the set conditions are: in: represents the reconstructed signal feature vector, Represents the mth closed set signal feature vector obtained by the lth intermediate layer network of the encoder, represents the reconstructed signal feature vector obtained by the lth intermediate layer network of the decoder, ε represents the threshold parameter, E(·) represents the mapping function corresponding to the encoding network, C(·) represents the mapping function corresponding to the classification network, N represents the number of samples contained in each type of closed set signal feature vector, y j represents the category label corresponding to the j-th closed set signal feature vector, L represents the number of network layers of the encoder or decoder, that is, the number of network layers of the encoder and decoder are equal, both are L layers; An open set recognition model construction submodule is used to construct an open set recognition network, and input the closed set signal feature vector and the open set signal feature vector into the open set recognition network, and use an adversarial training algorithm to train to obtain an open set recognition model; A signal recognition module is used to input the radio frequency fingerprint feature vector into the open set recognition model for recognition to obtain a recognition result; the recognition result is that the radio frequency communication signal to be identified is an open set signal or the radio frequency communication signal to be identified is a closed set signal; the open set signal indicates that the communication radiation source of the radio frequency communication signal to be identified is unknown; the closed set signal indicates that the communication radiation source of the radio frequency communication signal to be identified is known.

5. The communication radiation source open set identification system according to claim 4, characterized in that: The signal preprocessing module specifically includes: A bispectral analysis and processing submodule, configured to perform bispectral analysis on the radio frequency communication signal to be identified to obtain a bispectral distribution vector; The Radon transform processing submodule is used to perform Radon transform processing on the bispectral distribution vector to obtain a radio frequency fingerprint feature vector.

6. The communication radiation source open set identification system according to claim 4, characterized in that: The reconstructed signal feature vector generating unit specifically includes: an encoding subunit, configured to encode the closed set signal feature vector to obtain a latent vector; A linear combination subunit, configured to linearly combine the latent vectors to obtain a combined latent vector; The decoding subunit is used to decode the combined latent vector to obtain a reconstructed signal feature vector.

Images