CN114757224A

CN114757224A - Specific radiation source identification method based on continuous learning and combined feature extraction

Info

Publication number: CN114757224A
Application number: CN202210316468.6A
Authority: CN
Inventors: 闫文君; 谭凯文; 凌青; 于柯远; 朱子强; 张立民; 王程昱; 段可欣
Original assignee: School Of Aeronautical Combat Service Naval Aeronautical University Of Pla
Current assignee: School Of Aeronautical Combat Service Naval Aeronautical University Of Pla
Priority date: 2022-03-28
Filing date: 2022-03-28
Publication date: 2022-07-15

Abstract

The invention discloses a specific radiation source identification method based on continuous learning and combined feature extraction, wherein the method comprises the following steps: acquiring signals of a plurality of radiation sources and processing the signals of the plurality of radiation sources; inputting the signals of the plurality of radiation sources after signal processing into a plurality of trained continuous increment depth limit learning machines, and taking the plurality of trained continuous increment depth limit learning machines as a classifier to output a decision; and (3) fusing output decisions of a single continuous increment depth limit learning machine by using a voting algorithm, and selecting a class with the highest confidence coefficient as a classification result to identify a specific radiation source. The method has high identification precision under a small amount of samples, can realize continuous supervised identification of the collected samples, effectively meets the requirement of dynamic update of a database, has good compatibility on different modulation modes, carrier frequencies and transceiving distances, and can effectively identify a plurality of emitting electrodes.

Description

Specific radiation source identification method based on continuous learning and combined feature extraction

Technical Field

The invention relates to the technical field of communication, in particular to a specific radiation source identification method and device based on continuous learning and combined feature extraction.

Background

Specific Emitter Identification (SEI) refers to a technology for identifying a single emitter by using inherent defects of a physical layer of hardware equipment, and is widely applied to the fields of spectrum management and control, cognitive radio, self-organizing networks and the like. In a real channel, the additional nonlinear distortion of the ji intercepted signal is often not reproducible, so that a scheme for determining the device tag by using a Radio Frequency Fingerprint (RFF) characteristic is feasible. Supervised SEI based on RFF extraction tends to be divided into two phases: the first stage is feature extraction based on transient signals or steady-state signals, and the second stage is to construct a classifier to train and judge the characteristics of the front section. The transient signal is mainly generated at the moment of sudden change of the equipment state, and the characteristics are obviously distinguished. A Bayesian Transient Detector (BTD) is provided, power increase points of received signals are estimated, and matching of a plurality of Wi-Fi signal sources is realized; transient signals are described by using feature combinations such as Fractal Dimension (FD), entropy and kurtosis. However, transient signals have short duration and high interception difficulty, so transient detection faces various challenges in practical application.

Whereas SEI techniques based on steady-state signals have been validated in various wireless communication scenarios. The Hilbert transform has been proven to be an effective class of nonlinear, non-stationary signal analysis methods: extracting time-frequency energy distribution after Hilbert-Huang transformation, and finishing classification by adopting a Support Vector Machine (SVM); an SEI algorithm based on Energy Entropy (EE) and Hilbert moment analysis is provided, and Hilbert-like spectrums are separated by using a correlation coefficient and a Fisher discrimination coefficient, and are verified in single-hop and relay scenes; the Hilbert two-dimensional spectrum is used as a signal representation and sent to a Deep Residual Network (DRN) to extract potential visual features, and a good effect is achieved on a simulation data set for describing power amplifier distortion by using Taylor series; performing Variable Mode Decomposition (VMD) on the received signal to obtain different spectral characteristics, thereby effectively solving the problem of modal aliasing; in the literature, VMD is used to decompose bluetooth signals into band-limited modes and to reconstruct the modal components, classifying the higher order statistics using linear SVM. In addition, it is proposed to use Power Spectral Density (PSD) and Adjacent Channel Power Ratio (ACPR) as RFFs and perform dimensionality reduction using Principal Component Analysis (PCA); nonlinear dynamic characteristics based on Multi-dimensional Approximate Entropy (MAE) are extracted from the preamble signals, and the influence of modulation information on classification precision is reduced; extracting Unintentional Phase Modulation On Pulse (UPMOP) from the observation signal, fitting by using a Bessel curve, and sending into a long-short term neural network for identification; a fingerprint feature extraction method comprising statistical features, carrier frequency and wavelet packet transformation is provided, and an SVM (support vector machine) based on grid search is designed for classification; korean cleaning extracts a difference box dimension and a multi-fractal dimension by utilizing a fractal theory, constructs a feature vector based on a 3D-Hibert energy spectrum, and classifies by adopting an SVM; a Multi-Level Sparse Representation (MLSR) -based SEI method is provided, deep-layer and shallow-layer features are extracted from signals for classification, and Identification of an Automatic Identification System (AIS) of a ship is realized; the method is characterized in that Short Time Fourier Transform (STFT) is used for carrying out feature preprocessing on signals, and a sparse automatic encoder is used for carrying out unsupervised clustering on features, but the contradiction of Time-frequency resolution exists all the Time, and cross terms are difficult to suppress.

In recent years, with the development of 4G and 5G communication technologies, the number of user accesses and the number of access base stations are rapidly increased, and the magnitude of data transmitted through a wireless channel is rapidly increased, so a Deep Learning (DL) scheme based on data-driven modeling is widely applied to SEI. The advantage of DL is that it can automatically extract useful information representation from a large amount of data, fully mining the underlying regularity of the sample distribution. An I/Q signal SEI algorithm based on network compression is provided, sparse regularization, a quantization mask and a near-end gradient structure are embedded into a Complex-Valued Neural network (CVNN), and knowledge distillation is utilized to improve network performance; a compensation parameter extraction algorithm is designed for Zigbee equipment, an effective region in a received signal is selected in a self-adaptive mode according to a signal-to-noise ratio, and the effective region is input into a Multi-Sampling Convolutional Neural Network (MSCNN) for identification; embedding a Long Short-Term Memory (LSTM) into a Recurrent Neural Network (RNN) to complete the characteristic identification of an emitter, and still ensuring good identification precision under low signal-to-noise ratio; in the literature, diversity is accomplished with multiple receivers using neural networks to fuse skewness and kurtosis values of Empirical Mode Decomposition (EMD), Intrinsic time scale Decomposition (ITD), and VMD. However, the existing DL-based SEI model is usually built on a data set with sufficient samples and complete labels, and in an actual non-cooperative communication scenario, the sample size is often limited; and the database is in the middle of dynamic change, and the training of the DL model is usually based on the single learning of the existing sample, once the training set changes, the training needs to be carried out again, and the parameter self-updating capability of the model is poor.

Disclosure of Invention

The present invention is directed to solving, at least in part, one of the technical problems in the related art.

Therefore, a first object of the present invention is to provide a method for identifying a specific radiation source based on continuous learning and joint feature extraction, which extracts a Hilbert spectrum projection and a high-order spectrum after a Variational Modal Decomposition (VMD) from an intercepted signal respectively, and uses the projection and the high-order spectrum as a radio frequency fingerprint for classification after dimensionality reduction; in an Extreme Learning Machine (ELM), a sparse self-coding structure is adopted to carry out unsupervised training on a plurality of hidden layers, and a parameter search strategy is utilized to determine the optimal number of the hidden layers and the number of hidden nodes, so that continuous online matching of a plurality of batches of marked samples is realized. Experimental results show that the invention can show good compatibility to different modulation modes, carrier frequencies and transmitting-receiving distances, and can effectively identify a plurality of emitting electrodes.

Another object of the present invention is to provide a specific radiation source identification device based on continuous learning and joint feature extraction.

In order to achieve the above object, the present invention provides a method for identifying a specific radiation source based on continuous learning and joint feature extraction, including the following steps:

acquiring signals of a plurality of radiation sources and processing the signals of the plurality of radiation sources; inputting the signals of the plurality of radiation sources after signal processing into a plurality of trained continuous increment depth limit learning machines, and taking the plurality of trained continuous increment depth limit learning machines as a classifier to output a decision; and (3) fusing output decisions of a single continuous increment depth limit learning machine by using a voting algorithm, and selecting a class with the highest confidence as a classification result to identify a specific radiation source.

According to the specific radiation source identification method based on continuous learning and combined feature extraction, firstly, VMD (virtual machine description) spectrum gray level histogram vectors and bispectrum matrix diagonal values are constructed from signals to be identified, CIDELM (common information and computational complexity) is constructed to carry out dynamic weight updating on continuous samples, model calculation cost is reduced, and supervision identification on multiple batches of samples is achieved. The method can realize the online real-time identification of the USRPs, the identification accuracy can meet the actual requirement, and the method is not influenced by factors such as modulation modes, carrier frequencies and the like, and has strong robustness.

In order to achieve the above object, another aspect of the present invention provides a specific radiation source identification apparatus based on continuous learning and joint feature extraction, including:

the signal acquisition module is used for acquiring signals of a plurality of radiation sources and processing the signals of the plurality of radiation sources; the signal input module is used for inputting the signals of the plurality of radiation sources after signal processing into the trained plurality of continuous increment depth limit learning machines and taking the trained plurality of continuous increment depth limit learning machines as a classifier to output a decision; and the fusion output module is used for fusing the output decisions of the single continuous increment depth limit learning machine by using a voting algorithm and selecting the class with the highest confidence as a classification result to identify a specific radiation source.

The specific radiation source identification device based on continuous learning and combined feature extraction can realize online real-time identification of a plurality of USRPs, the identification accuracy can meet the actual requirement, and the device is not influenced by factors such as modulation modes, carrier frequencies and the like and has strong robustness.

Additional aspects and advantages of the invention will be set forth in part in the description which follows and, in part, will be obvious from the description, or may be learned by practice of the invention.

Drawings

The above and/or additional aspects and advantages of the present invention will become apparent and readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings of which:

FIG. 1 is a flow chart of a method for identifying a specific radiation source based on continuous learning and joint feature extraction according to an embodiment of the invention;

FIG. 2 is a diagram of an SEI framework based on joint feature extraction according to an embodiment of the present invention;

FIG. 3 is a spectrogram of Hilbert3D after VMD processing according to an embodiment of the present invention;

FIG. 4 is a time-frequency domain projection diagram according to an embodiment of the present invention;

FIG. 5 is a spectral time domain projection diagram according to an embodiment of the present invention;

FIG. 6 is a projection diagram of the amplitude-frequency domain according to an embodiment of the present invention;

FIG. 7 is a schematic diagram of a DELM architecture according to an embodiment of the present invention;

Fig. 8 is a schematic diagram illustrating an influence of the number k of radiation sources being 3 on bispectrum feature identification according to an embodiment of the present invention;

fig. 9 is a schematic diagram illustrating an influence of the number k of radiation sources being 4 on bispectral feature identification according to an embodiment of the invention;

fig. 10 is a schematic diagram illustrating an influence of the number k of radiation sources being 5 on bispectral feature identification according to an embodiment of the invention;

fig. 11 is a schematic diagram illustrating an influence of the number k of radiation sources being 6 on bispectral feature identification according to an embodiment of the invention;

fig. 12 is a diagram of the recognition result of VMD spectral gray scale vector for which the number of sources k is 3 according to an embodiment of the present invention;

fig. 13 is a diagram of the recognition result of VMD spectrum gray scale vector when the number k of radiation sources is 4 according to the embodiment of the present invention;

fig. 14 is a diagram of the recognition result of VMD spectrum gray scale vector when the number k of radiation sources is 5 according to the embodiment of the present invention;

fig. 15 is a diagram of the recognition result of VMD spectrum gray scale vector when the number k of radiation sources is 6 according to the embodiment of the present invention;

FIG. 16 is a diagram of an integrated CIDELM architecture according to an embodiment of the present invention;

FIG. 17 is a diagram illustrating integrated CIDELM recognition capabilities according to an embodiment of the present invention;

fig. 18 is a graph showing the recognition performance of the number k of sources 3 according to the embodiment of the present invention as a function of the transmitting/receiving distance;

Fig. 19 is a graph showing the recognition performance of the number k of radiation sources equal to 4 as a function of the transmission/reception distance according to the embodiment of the present invention;

fig. 20 is a graph showing the recognition performance of the number k of radiation sources equal to 5 as a function of the transmission/reception distance according to the embodiment of the present invention;

fig. 21 is a graph showing the recognition performance of the number k of radiation sources equal to 6 as a function of the transmission/reception distance according to the embodiment of the present invention;

FIG. 22 is a comparison of recognition effects of different methods according to an embodiment of the present invention;

fig. 23 is a schematic structural diagram of a specific radiation source identification device based on continuous learning and joint feature extraction according to an embodiment of the present invention.

Detailed Description

It should be noted that the embodiments and features of the embodiments in the present application may be combined with each other without conflict. The present invention will be described in detail below with reference to the accompanying drawings in conjunction with embodiments.

In order to make the technical solutions of the present invention better understood, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

The following describes a specific radiation source identification method and apparatus based on continuous learning and joint feature extraction according to an embodiment of the present invention with reference to the accompanying drawings, and first, the specific radiation source identification method based on continuous learning and joint feature extraction according to an embodiment of the present invention will be described with reference to the accompanying drawings.

Fig. 1 is a flowchart of a specific radiation source identification method based on continuous learning and joint feature extraction according to an embodiment of the present invention.

As shown in fig. 1, the method for identifying a specific radiation source based on continuous learning and joint feature extraction includes the following steps:

step S1, acquiring signals of a plurality of radiation sources and processing the signals of the plurality of radiation sources;

step S2, inputting the signals of the plurality of radiation sources after signal processing into a plurality of trained continuous increment depth limit learning machines, and taking the plurality of trained continuous increment depth limit learning machines as a classifier to output a decision;

and step S3, fusing output decisions of the single continuous increment depth limit learning machine by using a voting algorithm, and selecting the class with the highest confidence as a classification result to identify a specific radiation source.

The embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

It can be understood that the SEI scheme commonly used at present is to perform feature extraction on the preprocessed signal, and send the signal to a classifier for recognition. The SEI framework based on joint feature extraction and incremental learning is shown in fig. 2. The real intercepted signal from the receiver is preprocessed to extract the radio frequency fingerprint characteristics, and the purpose of fingerprint extraction is to amplify the individual difference information among different radiation sources. According to the invention, the Hilbert spectrum and the high-order spectrum diagonal matrix processed by the VMD are used as the combined characteristic of the radiation source signal, and the purpose is to make up for the defect of single-scale identification. For a continuous feature data stream from K emitters, it is divided into a training set and a test set, and multiple improved depth-limit learning machines are trained simultaneously. And finally, fusing the output decisions of the single models by using a voting algorithm, and selecting the class with the highest confidence coefficient as a final classification result.

Specifically, radiation source signal modeling is performed. Supervised radiation source recognition can be simplified into a K-class decision problem, with the input to the system being a discrete sequence and the output being the discrimination probability for the sample. Interception signal r from the k-th radiation source^(k)(t) is expressed as:

Where s (t) is a transmission signal, h (t) is an equivalent channel impulse response, n (t) is channel additive noise, and the in-phase and quadrature components output after USRP sampling can be represented as:

further, metamorphic modal decomposition. The original signal is reconstructed by utilizing the Intrinsic Mode function component, and is decomposed into an Intrinsic Mode Functions (IMF) with a sparse characteristic, so that the endpoint effect and the Mode confusion phenomenon generated by EMD can be effectively inhibited. The variational modal decomposition decomposes an original signal into a plurality of AM-FM signals:

in the formula u_k(t) is the kth modal component, A_k(t) is the instantaneous amplitude of the wave,

the method is characterized in that a signal phase is formed, the core idea is to construct and solve a variational problem, an original signal is decomposed into a plurality of inherent mode functions, and a constraint variational expression is as follows:

{u_k(t)}＝{u₁(t),u₂(t),...,u_K(t) } is the set of modal components, { Φ_k(t)}＝{f₁(t),f₂(t),...,f_k(t) } is the set of modal center frequencies, where

δ (t) is the impulse response function, and S (t) is the signal to be decomposed. The VMD converts the modal decomposition problem into an unconstrained variant solution, introduces a secondary penalty factor alpha and a Lagrange multiplier lambda (t) to obtain an optimal solution, and an augmented Lagrange expression is as follows:

iterating through updates using an alternating direction multiplier algorithm

And λ ⁿ⁺¹To solve the saddle point in the formula (6) and further obtain the optimal solution set { u } in the frequency domain_k(ω)}，{f_k(ω) } and { λ_k(ω) }, the update formula in the positive frequency condition is:

wherein tau is the update coefficient of Lagrange operator,

for fourier transform, update methods (6) - (8) are repeated until they satisfy the convergence condition:

ε is the discrimination accuracy, set to 10 in this text^-6When the convergence condition is satisfied, the iteration is stopped, and the discrete signal sequence s (m) is represented by k modes as:

hilbert transform is carried out on the obtained k modes, an analytic signal z (m) is constructed, and an instantaneous frequency f is calculated_i(m) and instantaneous amplitude A_i(m)：

In the formula

VMD-HiThe lbert spectrum is expressed as:

in the formula

Representing the real part, herein transforming the projected VMD-Hilbert spectrogram into a grayscale image, the energy value h (i, j) at the (i, j) th time-frequency point in the spectrogram can be transformed into a corresponding grayscale value G (i, j):

l represents the number of bits of the gray scale map,

expressing rounding downwards, the method uses the statistical characteristics based on the gray level image as the radio frequency fingerprint, firstly constructs a gray level histogram which can accurately reflect the frequency of each gray level pixel point in the image and the gray level relation, and maps the distribution condition of the time-frequency energy in the VMD-Hilbert spectrum into a two-dimensional coordinate system. Fig. 3, 4, 5 and 6 are interception signal VMD-Hilbert projections from USRP-2922. Fig. 3 is a Hilbert3D spectrum after VMD processing, fig. 4 is a time-frequency domain projection, fig. 5 is a spectral time domain projection, and fig. 6 is a magnitude-frequency domain projection.

Further, the high-order cumulant can effectively inhibit additive Gaussian color noise, describe the deviation degree of the signal relative to Gaussian white noise, and sample sequence r^(l)The k-th order cumulative amount of (m) is defined as:

C_kx(τ₁,τ₂,...,τ_k-1)＝cum{x(n),x(n+τ₁),...,x(n+τ_k-1) (15)

when it satisfies the absolute condition:

the spectrum of the k-th order cumulant is defined as a k-1 order Fourier transform of the k-th order cumulant:

when k is 3, the spectrum of the higher order cumulant is called bispectrum:

the invention adopts a sliding window method to estimate the high-order cumulant, and the steps are shown in the table 1:

TABLE 1 sliding window method high-order spectral estimation procedure

Further, the IDELM-based samples are continuously learned. Among them, the Extreme Learning Machine (ELM) usually has only one hidden layer, and it randomly generates input weights and hidden node parameters, and the output weights are calculated by analyzing and calculating a generalized inverse matrix. ELM determines the optimal solution by a single step Least Squares Error (LSE) without gradient backpropagation. The ELM with L hidden nodes is represented as:

wherein G (-) is an activation function, w_i＝[w_i,1,w_i,2,...,w_i,n]^T∈RⁿIs the weight between the input layer and the i-th hidden node, beta_iAs output weights, b_iWeight bias for the ith hidden node, w_i·x_jDenotes w_iAnd x_jThe solution of equation (19) can be represented by a matrix as:

T＝Hβ (20)

in the formula:

Where H is the output matrix, β is the output weight, T is the desired output, and the loss function of ELM is:

the training process of the network is equivalent to solving the LSE solution of the linear system H beta ═ T

In the formula

Norm is minimum and unique, T ═ T₁,...,t_N]，Η⁺Moore-Penrose generalized inverse matrix, H.matrixing. When H is present^TWhen H is not singular, there is H⁺＝(H^TH)^-1H^TOtherwise H⁺＝H^T(HH^T)^-1。

Further, sorting is based on the characteristics of successive increments DELM. In a real scene, captured radiation source signals are always Continuous data streams, therefore, a Continuous Learning mechanism is introduced on the basis of a depth limit Learning Machine, a Continuous increment depth limit Learning Machine (CIDELM) is designed, parameters of the CIDELM are continuously updated according to the input sequence of a sample, the classification efficiency is improved, and computing resources are saved. In the first stage, a plurality of feedforward hidden layers are connected in series to serve as a self-encoder to carry out sparse representation on input samples, and hidden information of signal characteristics is mined; and in the second stage, the single-layer ELM is used for carrying out supervised regression, and a final classification result is output.

As shown in FIG. 7, the first hidden layer output H of the Deep Extreme Learning Machine (DELM)₁Expressed as:

H₁＝G(w_1,2·H+B₁) (24)

in the formula w_1,2Representing a weight matrix between the one and two hidden layers, H₁Representing the output of an input through a first hidden layer, B₁Denotes the bias of the first hidden layer with initial weight β₁＝H₁ ⁺T₁The desired output matrix of the second hidden layer can be represented as H_2E＝Tβ₁ ⁺Defining an augmentation matrix w_H2＝[B₂,w_1,2]In the formula B₂Is the input bias of the second hidden layer, then w_H2Can be calculated from the following formula:

in the formula M₂＝[1 H₁]^TWhere 1 is the column vector of the full scalar 1, G^-1(. H) is the inverse of the activation function G (·), the actual output of the second hidden layer being H₂＝G(w_H2·M₂) Output weight β of the second hidden layer₂＝H₂ ⁺T₂The expected output H of the ith hidden layer is calculated inductively_iEActual output H_iAnd an output weight, the DELM output with P hidden layers is represented as:

f_P,L(x)＝H_Pβ_P (26)

the invention considers that the marking signal sample for training is sent to DELM to carry out unsupervised sequence training layer by layer, and the k +1 part of input sample is set as

The number of samples in each group is N_k+1Then outputs the weight β^(k+1)The update formula of (2) is:

in the formula

The ith hidden layer output corresponding to time step k is:

in training CIDELM, the initial output matrix for each hidden layer is first calculated from equation (28) above using the first set of samples as inputs

Then inputting sample data of K-1 time steps in sequence, and outputting H when grouping pair K is 0₀And initial weight

Updating, and finally calculating the weight of the model:

further, the experimental results and analysis of the present invention are as follows:

under experimental conditions, the currently common SEI signal simulation method simulates the nonlinear distortion difference between different radiation source individuals through Taylor polynomials, and the USRP based on GNU Radio is used as a transceiver. The signal is transmitted in a real channel of a laboratory, the algorithm is verified by changing the receiving and transmitting distance, the modulation mode and the carrier frequency, and the connection mode is an I/Q double-path. Six USRP-2922 produced in the same batch are used as signal generators, one USRP-B210 is used as receiving equipment, and logarithmic spiral antennas are selected as transmitting antennas to increase the acting distance. BPSK and BFSK are selected as signal modulation modes, carrier frequency is set to be 500MHz and 1GHz, and bandwidth is 100 KHz. And in each modulation mode, the number of sampling points of each carrier frequency is 106, the down-sampling rate of the USRP-B210 is 1MHz, and a k-fold verification method is adopted to randomly divide the samples into a training set and a test set. The method comprises the steps of calculating two types of characteristics of the same signal sample by using the fingerprint extraction method, setting the number of conversion points to be 1000, normalizing the calculated characteristic vectors, eliminating singular values and then sequentially inputting the singular values into CIDELM.

The single-feature CIDELM parameter settings, it has been demonstrated in the foregoing that CIDELM identification performance is affected by initial weight, initial bias, number of hidden layers, and number of hidden nodes, whereas the initial weight and initial bias of the model are typically generated randomly. Therefore, the invention adopts a parameter optimization strategy to adjust the number of the hidden layers and the number of the hidden nodes so as to obtain a parameter combination with the best performance. Fig. 8, fig. 9, fig. 10, and fig. 11 show the influence of two types of hyper-parameters on bispectral feature recognition, and the default number of nodes of the input hidden layer and the output hidden layer in the graph is 2000. In fig. 8, the number of radiation sources k is 3, fig. 9 the number of radiation sources k is 4, fig. 10 the number of radiation sources k is 5, and fig. 11 the number of radiation sources k is 6.

As can be seen from fig. 8, 9, 10, and 11, when the number k of radiation sources is changed, the single-feature real-time identification model based on CIDELM can achieve a high accuracy by matching and optimizing two-dimensional parameters, where the optimal identification rate reaches 97% when k is 3, 94% when k is 4, 91% when k is 5, and 88% when k is 6. When the number of CIDELM hidden layers is 2-4, the number of hidden nodes is 500 and 1000, and the higher identification accuracy is shown, because the hidden layer model under the condition is compact, the layer sparse coding effect on the input characteristics is better, the probability that the layer output matrix is not full-rank is caused by the increase of the number of the hidden layers and the hidden nodes, and the calculated MP inverse matrix errors are accumulated layer by layer at the moment. When the function approximation problem of small-batch data is solved, by setting proper hidden nodes and hidden layer numbers, the CIDELM can reduce the search space of parameters and improve the identification accuracy and the identification speed; in addition, the CIDELM processes samples in batches, parameters of the model are updated on line in a continuous training mode, and the adaptability of the algorithm to the data scale is obviously enhanced. Fig. 12, fig. 13, fig. 14, and fig. 15 show the recognition results of the VMD spectral gray vector, and it can be seen from the figures that the accuracy of recognizing the VMD spectral gray vector by using cidel can reach more than 91% by setting the number of hidden layers and the number of hidden nodes. Fig. 12 shows that the number of radiation sources k is 3, fig. 13 shows that the number of radiation sources k is 4(c), fig. 14 shows that the number of radiation sources k is 5, and fig. 15 shows that the number of radiation sources k is 6.

Further, based on the joint feature discrimination of the integrated CIDELM, because the initial parameters of the CIDELM are randomly generated, the output of the network is unstable, and the reliability of the model is influenced. Therefore, the invention constructs an integrated CIDELM algorithm and combines the output results of a limited number of similar network structures. Through a mode of synchronous training of a plurality of models, two types of characteristics from the same section of signals are respectively sent to K CIDELMs for training and testing, the result of each CIDELM is voted through a majority voting Algorithm (Boyer-Moore Algorithm, BMA), and the final judgment result is the radiation source type with the vote exceeding half. The result of integrating CIDELM is superior to the classification effect of a single model, and the stability and reliability of the method are stronger. FIG. 16 is a structure of integration CIDELM.

Fig. 17 is an algorithm comparison of the integrated ciedel on a data set, and the present invention verifies the robustness and compatibility of the algorithm by changing the modulation mode, the transceiving distance and the carrier frequency, and the transceiving distance is set to 4 meters. As can be seen from fig. 17, when the number of radiation sources is within 5, the recognition accuracy of the combination of different carrier frequencies and modulation modes can reach over 90%, and the change of the recognition accuracy is less affected by the modulation mode and the carrier frequency; when the number of the radiation sources is more than 6, the identification accuracy can still reach more than 86%, and the fact that the combined feature extraction scheme provided by the invention can show good compatibility on the established data set is proved.

Fig. 18, fig. 19, fig. 20, and fig. 21 are performance performances of the algorithm in different parameter combination data sets, and when the transceiving distance changes, the identification performance slowly decreases, which may be because in a laboratory environment, other devices work to affect a wireless transmission environment, and channel noise increases, but the overall identification accuracy can reach more than 80%, which indicates that the algorithm of the present invention can meet the requirement of real-time high-precision identification.

And comparing the selection method I, the selection method II and the selection method III with the method for generating the mandarin orange. In the first method, HHT time entropy is selected as RFF, and KNN is used for classification; selecting a Hilbert spectrum as an RFF (radio frequency spectrum), and classifying by using a DRN (Dr-noise network); and thirdly, performing dimensionality reduction mapping on the MAE of the signal by using a PCA algorithm. The comparison results are shown in fig. 22, and the number of radiation sources was set to 6. The invention uses the dual-mode feature extraction based on time-frequency domain analysis, and provides the improved ELM with better recognition performance for the one-dimensional vector, and the average recognition accuracy is about 2-4% higher than that of the method.

Further, the algorithm computation complexity is analyzed, the computation amount of the algorithm mainly comes from the computation complexity generated in the front-end dual-mode radio frequency fingerprint extraction and the back-end classifier iterative training, and the iteration time and the average identification time of the algorithm are selected to balance the time complexity of the algorithm. The CPU of the operating system of the experimental computer is Intel (R) core (TM) i7-9750H, the size of the operating memory is 16GB, the GPU is NVIDIA GeForce RTX 3080, and the time complexity analysis of the algorithm is given in Table 2.

TABLE 2 Algorithm time complexity analysis

As can be seen from Table 2, the SEI method of dual-mode feature extraction and online incremental learning provided by the invention has a faster model training speed and a shorter recognition time. Since the back propagation of the gradient is not needed to update the global parameters, the classification method based on CIDELM has higher real-time performance, and compared with the huge training cost of a DL model, the algorithm of the invention has lower computational complexity.

The conclusion is drawn, and the SEI method based on double-feature sorting and continuous increment depth limit learning is provided by the invention aiming at the problems that the performance of an SEI algorithm under the condition of a small sample is influenced by the sample scale, the generalization is low, and the model training cost is high due to the update of a dynamic sample. Firstly, VMD spectrum gray level histogram vectors and bispectrum matrix diagonal values are constructed from signals to be identified, and CIDELM is constructed to carry out dynamic weight updating on continuous samples, so that model calculation cost is reduced, and supervision identification on multiple batches of samples is realized. The experimental results show that: the algorithm provided by the invention can realize online real-time identification of a plurality of USRPs, the identification accuracy can meet the actual requirement, and the algorithm is not influenced by factors such as modulation modes, carrier frequencies and the like and has strong robustness.

According to the specific radiation source identification method based on continuous learning and combined feature extraction, firstly, a VMD (vector-vector discrete cosine transformation) spectrum gray level histogram vector and a bispectrum matrix diagonal value are constructed from a signal to be identified, and a CIDELM (Carrier induced breakdown modulus) is constructed to update dynamic weights of continuous samples, so that model calculation cost is reduced, and supervision identification of multiple batches of samples is realized.

In order to implement the foregoing embodiment, as shown in fig. 23, there is also provided a specific radiation source identification apparatus 10 based on continuous learning and joint feature extraction in this embodiment, where the apparatus 10 includes: the system comprises a signal acquisition module 100, a signal input module 200 and a fusion output module 300.

A signal acquisition module 100, configured to acquire signals of a plurality of radiation sources and perform signal processing on the signals of the plurality of radiation sources;

a signal input module 200, configured to input the signals of the multiple radiation sources after signal processing into the trained multiple continuous increment depth limit learning machines, and use the trained multiple continuous increment depth limit learning machines as a classifier to output a decision;

and the fusion output module 300 is configured to fuse the output decisions of the single continuous increment depth limit learning machine by using a voting algorithm, and select a class with the highest confidence as a classification result to identify a specific radiation source.

Further, the training module is used for training a plurality of continuous increment depth limit learning machines based on signals of a plurality of sample radiation sources.

According to the specific radiation source identification device based on continuous learning and combined feature extraction, disclosed by the embodiment of the invention, online real-time identification of a plurality of USRPs can be realized, the identification accuracy can meet the actual requirement, the device is not influenced by factors such as a modulation mode, carrier frequency and the like, and the robustness is strong.

It should be noted that the foregoing explanation on the embodiment of the specific radiation source identification method based on continuous learning and joint feature extraction is also applicable to the specific radiation source identification apparatus based on continuous learning and joint feature extraction in this embodiment, and details are not repeated here.

It should be understood that the above-described embodiments are exemplary and should not be construed as limiting the present invention, and those skilled in the art may make variations, modifications, substitutions and alterations to the above-described embodiments while remaining within the scope of the present invention.

Claims

1. A specific radiation source identification method based on continuous learning and combined feature extraction is characterized by comprising the following steps:

acquiring signals of a plurality of radiation sources and processing the signals of the plurality of radiation sources;

Inputting the signals of the plurality of radiation sources after signal processing into a plurality of trained continuous increment depth limit learning machines, and taking the trained continuous increment depth limit learning machines as a classifier to output a decision;

and (3) fusing output decisions of a single continuous increment depth limit learning machine by using a voting algorithm, and selecting a class with the highest confidence coefficient as a classification result to identify a specific radiation source.

2. The method of claim 1, wherein before inputting the continuous feature data stream and the diagonal matrix into a trained plurality of continuous incremental depth limit learning machines, further comprising: training the plurality of successive incremental depth limit learning machines based on signals of a plurality of sample radiation sources.

3. The method of claim 2, wherein training the plurality of successive incremental depth-limit learning machines based on signals of a plurality of sample radiation sources comprises:

intercepting signals of a plurality of sample radiation sources by using a universal software radio peripheral platform (USRP), preprocessing the signals of the plurality of sample radiation sources and extracting radio frequency fingerprint characteristics;

utilizing VMD to process the radio frequency fingerprint characteristics to obtain Hilbert time-frequency energy spectrum, performing projection dimensionality reduction, and converting the Hilbert time-frequency energy spectrum into a gray level vector to obtain a continuous characteristic data stream; analyzing the radio frequency fingerprint characteristics by using a high-order spectral vector to obtain a diagonal matrix; wherein the continuous stream of characteristic data of the signals of the plurality of sample radiation sources comprises a training set and a test set;

And inputting the continuous characteristic data stream and the diagonal matrix into a plurality of continuous increment depth limit learning machines for training to obtain the trained continuous increment depth limit learning machines.

4. The method of claim 1, wherein modeling the signal of the radiation source comprises:

wherein s (t) is a transmission signal, h (t) is an equivalent channel impulse response, n (t) is channel additive noise, and the inphase and quadrature components output after USRP sampling are represented as:

5. the method of claim 1, wherein performing a variational modal decomposition of a signal of the radiation source comprises:

decomposing the signal of the radiation source into a plurality of AM-FM signals:

wherein u_k(t) is the kth modal component, A_k(t) is the instantaneous amplitude of the wave,

decomposing the signal of the radiation source into a plurality of natural mode functions for the signal phase, wherein the constraint variational expression is as follows:

wherein, { u_k(t)}＝{u₁(t),u₂(t),...,u_K(t) is the set of modal components, { Φ }_k(t)}＝{f₁(t),f₂(t),...,f_k(t) } is the set of modal center frequencies, where

δ (t) is the impulse response function, S (t) is the signal to be decomposed; introducing a secondary penalty factor alpha and a Lagrange multiplier lambda (t) to obtain an optimal solution, wherein the expression of the augmented Lagrange is as follows:

Iterating through updates using an alternating direction multiplier algorithm

And λⁿ⁺¹The saddle point in the formula (6) is solved to obtain the optimal solution set { u } in the frequency domain_k(ω)}，{f_k(ω) } and { λ_k(ω) }, the update formula in the positive frequency condition is:

wherein tau is the update coefficient of Lagrange operator,

epsilon is the discrimination accuracy, iteration is stopped when a convergence condition is satisfied, and the discrete signal sequence S (m) is expressed by k modes as:

In the formula

The VMD-Hilbert spectrum is represented as:

in the formula (I), the compound is shown in the specification,

representing a real part, converting the projected VMD-Hilbert spectrogram into a gray level image, and converting an energy value H (i, j) of an (i, j) th time-frequency point in the spectrogram into a corresponding gray level value G (i, j):

where, l represents the number of bits of the gray scale map,

indicating a rounding down.

6. The method of claim 3, wherein the higher order spectral vector comprises:

sample sequence r^(l)The k-order cumulant of (m) is defined as:

C_kx(τ₁,τ₂,...,τ_k-1)＝cum{x(n),x(n+τ₁),...,x(n+τ_k-1)(15)

when it satisfies the absolute condition:

When k is 3, the spectrum of higher order cumulants is called bispectrum:

7. the method of claim 6, wherein the extreme learning machine ELM with L hidden nodes is represented as:

wherein G (. cndot.) is an activation function, w_i＝[w_i,1,w_i,2,...,w_i,n]^T∈RⁿAs a weight between the input layer and the i-th hidden node, β_iAs output weight, b_iWeight bias for the ith hidden node, w_i·x_jDenotes w_iAnd x_jThe solution of equation (19) is represented by a matrix:

T＝Hβ(20)

in the formula:

the training process of the network is equivalent to solving the LSE solution of the linear system H beta-T

In the formula (I), the compound is shown in the specification,

norm is minimum and unique, T ═ T₁,...,t_N]，Η⁺Moore-Penrose generalized inverse matrix as H^TWhen H is not singular, there is H⁺＝(H^TH)^-1H^TOtherwise H⁺＝H^T(HH^T)^-1。

8. The method of claim 7, wherein the first hidden layer output H of the deep extreme learning machine DELM₁Expressed as:

H₁＝G(w_1,2·H+B₁)(24)

in the formula w_1,2Representing a weight matrix between the first and second hidden layers, H₁Representing the output of an input through a first hidden layer, B₁Denotes the bias of the first hidden layer with initial weight β₁＝H₁ ⁺T₁The desired output matrix of the second hidden layer is denoted as H_2E＝Tβ₁ ⁺Defining an augmentation matrix w _H2＝[B₂,w_1,2]In the formula B₂For the input bias of the second hidden layer, then w_H2Calculated from the following formula:

f_P,L(x)＝H_Pβ_P(26)

let the k +1 th input sample be

Number of samples in each group is N_k+1Then outputs the weight β^(k+1)The update formula of (2) is:

in the formula

The ith hidden layer output corresponding to time step k is:

when training the CIDELM, the initial output matrix of each hidden layer is calculated by equation (28) with the first group of samples as input

Updating, and finally calculating the weight of the model:

9. a specific radiation source identification device based on continuous learning and joint feature extraction, comprising:

the signal acquisition module is used for acquiring signals of a plurality of radiation sources and processing the signals of the plurality of radiation sources;

The signal input module is used for inputting the signals of the plurality of radiation sources after signal processing into a plurality of trained continuous increment depth limit learning machines and taking the trained continuous increment depth limit learning machines as a classifier to output a decision;

and the fusion output module is used for fusing the output decisions of the single continuous increment depth limit learning machine by using a voting algorithm and selecting the class with the highest confidence coefficient as a classification result to identify a specific radiation source.

10. The apparatus of claim 9, further comprising a training module to train the plurality of successive incremental depth limit learning machines based on signals of a plurality of sample radiation sources.