A specific radiation source identification method based on continuous learning and joint feature extraction

Abstract

The application discloses a specific radiation source identification method based on continuous learning and joint feature extraction, and comprises the following steps: acquiring signals of multiple radiation sources and performing signal processing on the signals of the multiple radiation sources; inputting the signals of the multiple radiation sources processed into multiple trained continuous incremental deep extreme learning machines, and taking the multiple trained continuous incremental deep extreme learning machines as classifiers to output decisions; using a voting algorithm to fuse the output decisions of single continuous incremental deep extreme learning machines, selecting a class with the highest confidence as a classification result to identify a specific radiation source. The application still has high identification precision under a small amount of samples, can realize continuous supervised identification of collected samples, effectively meets the demand of database dynamic updating, has good compatibility for different modulation modes, carrier frequencies and transmission-reception distances, and can effectively identify multiple emission poles.

Description

Technical Field

[0001] This invention relates to the field of communication technology, and in particular to a method and apparatus for identifying specific radiation sources based on continuous learning and joint feature extraction. Background Technology

[0002] Specific emitter identification (SEI) is a technique that uses inherent defects in the physical layer of hardware devices to identify individual emitters. It is widely used in spectrum management, cognitive radio, and self-organizing networks. In real-world channels, the nonlinear distortion added to the intercepted signal is often unreplicable; therefore, using radio frequency fingerprint (RFF) characteristics to determine device tags is feasible. Supervised SEI based on RFF extraction typically consists of two stages: the first stage is feature extraction based on transient or steady-state signals, and the second stage is building a classifier to train and discriminate the features from the first stage. Transient signals are mainly generated at the moment of abrupt changes in device state, and their features are easily distinguishable. A Bayesian Transient Detector (BTD) is proposed to estimate the power increase point of the received signal, enabling matching of multiple Wi-Fi signal sources; it uses a combination of features such as fractal dimension (FD), entropy, and kurtosis to describe the transient signal. However, transient signals have short durations and are difficult to intercept, thus transient detection faces various challenges in practical applications.

[0003] The SEI technology based on steady-state signals has been validated in various wireless communication scenarios. The Hilbert transform has proven to be an effective method for analyzing nonlinear and non-stationary signals: extracting the time-frequency energy distribution after the Hilbert-Huang transform and using Support Vector Machines (SVM) for classification; proposing an SEI algorithm based on Energy Entropy (EE) and Hilbert moment analysis, and using correlation coefficients and Fisher discriminant coefficients to separate Hilbert-like spectra, validated in single-hop and relay scenarios; using the Hilbert two-dimensional spectrum as a signal representation and feeding it into a Deep Residual Network (DRN) to extract latent visual features, achieving good results on a simulation dataset describing power amplifier distortion using Taylor series; performing Variational Mode Decomposition (VMD) on the received signal to obtain different spectral features, effectively solving the mode aliasing problem; in the literature, VMD has been used to decompose Bluetooth signals into band-limited modes and reconstruct the mode components, using linear SVM to classify higher-order statistics.Furthermore, this paper proposes using power spectral density (PSD) and adjacent channel power ratio (ACPR) as the RFF and performing dimensionality reduction using principal component analysis (PCA). It also proposes extracting nonlinear dynamic features based on multi-dimensional approximate entropy (MAE) from the preamble signal to reduce the impact of modulation information on classification accuracy. Unintentional phase modulation on pulse (UPMOP) features are extracted from the observed signal, fitted with Bézier curves, and then fed into a long short-term neural network for identification. A fingerprint feature extraction method incorporating statistical features, carrier frequency, and wavelet packet transform is proposed, and a grid-search-based SVM is designed for classification. Han Jie utilizes fractal theory to extract the difference box dimension and multifractal dimension, constructing a feature vector based on 3D-Hibert energy spectrum, and using SVM for classification. Finally, a multi-level sparse representation method is proposed. The SEI method of Sparse Representation (MLSR) extracts deep and shallow features from the signal for classification, realizing the identification of Automatic Identification System (AIS). It uses Short Time Fourier Transform (STFT) to preprocess the signal features and uses sparse autoencoder to perform unsupervised clustering of features. However, the contradiction between time and frequency resolution always exists, and the cross term is difficult to suppress.

[0004] In recent years, with the development of 4G and 5G communication technologies, the number of user accesses and the number of access base stations have increased rapidly, leading to a rapid increase in the amount of data transmitted through wireless channels. Therefore, deep learning (DL) solutions based on data-driven modeling are widely used in SEI (Search Engine Intelligence). The advantage of DL lies in its ability to automatically extract useful information representations from large amounts of data and fully explore the potential patterns in sample distribution. This paper proposes a network compression-based SEI algorithm for I / Q signals, embedding sparse regularization, quantization masking, and near-end gradient structures into complex-valued neural networks (CVNNs), and utilizing knowledge distillation to improve network performance. A compensation parameter extraction algorithm is designed for Zigbee devices, adaptively selecting effective regions in the received signal based on the signal-to-noise ratio (SNR) and inputting them into a multi-sampling convolutional neural network (MSCNN) for recognition. Long short-term memory (LSTM) structures are embedded into recurrent neural networks (RNNs) to complete emitter characteristic recognition, ensuring good recognition accuracy even at low SNR. In the literature, neural networks are used to fuse the skewness and kurtosis values ​​of Empirical Mode Decomposition (EMD), Intrinsic Timescale Decomposition (ITD), and VMD, and diversity is achieved using multiple receivers. However, existing DL-based SEI models are usually built on datasets with sufficient samples and complete labels. In actual non-cooperative communication scenarios, the sample size is often limited, and the database is constantly changing. The training of DL models is usually based on a single learning of existing samples. Once the training set changes, retraining is required, and the model's parameter self-updating ability is poor. Summary of the Invention

[0005] The present invention aims to at least partially solve one of the technical problems in the related art.

[0006] Therefore, the first objective of this invention is to propose a specific radiation source identification method based on continuous learning and joint feature extraction. This method extracts the Hilbert spectral projection and higher-order spectrum after variational mode decomposition (VMD) from the intercepted signal, and uses these as radio frequency fingerprints for classification after dimensionality reduction. In an Extreme Learning Machine (ELM), a sparse autoencoder structure is used to perform unsupervised training on multiple hidden layers, and a parameter search strategy is employed to determine the optimal number of hidden layers and hidden nodes, enabling continuous online matching of multiple batches of labeled samples. Experimental results show that this invention exhibits good compatibility with different modulation schemes, carrier frequencies, and transmission / reception distances, and can effectively identify multiple emitters.

[0007] Another objective of this invention is to propose a specific radiation source identification device based on continuous learning and joint feature extraction.

[0008] To achieve the above objectives, this invention proposes a method for identifying specific radiation sources based on continuous learning and joint feature extraction, comprising the following steps:

[0009] Signals from multiple radiation sources are acquired and processed. The processed signals from the multiple radiation sources are then input into multiple trained continuous incremental deep learning machines (LILMs), which are used as classifiers to output decisions. A voting algorithm is used to fuse the output decisions of a single LILM, and the class with the highest confidence is selected as the classification result to identify a specific radiation source.

[0010] The specific radiation source identification method based on continuous learning and joint feature extraction in this invention first constructs the VMD spectral grayscale linear vector and the diagonal values ​​of the bispectral matrix from the signal to be identified, and then constructs CIDELM to dynamically update the weights of continuous samples, reducing the computational cost of the model and achieving supervised identification of multiple batches of samples. This invention can achieve online real-time identification of multiple USRPs, with an identification accuracy that meets practical needs, and is not affected by factors such as modulation method and carrier frequency, exhibiting strong robustness.

[0011] To achieve the above objectives, another aspect of the present invention proposes a specific radiation source identification device based on continuous learning and joint feature extraction, comprising:

[0012] The signal acquisition module is used to acquire signals from multiple radiation sources and perform signal processing on these signals. The signal input module is used to input the processed signals from multiple radiation sources into multiple trained continuous incremental deep learning machines (MLMs), and use these MLMs as classifiers to output decisions. The fusion output module is used to fuse the output decisions of a single MLM using a voting algorithm, and select the class with the highest confidence as the classification result to identify a specific radiation source.

[0013] The specific radiation source identification device based on continuous learning and joint feature extraction in this invention can achieve online real-time identification of multiple USRPs. The identification accuracy can meet the actual needs and is not affected by factors such as modulation method and carrier frequency, and has strong robustness.

[0014] 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. Attached Figure Description

[0015] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0016] Figure 1 This 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;

[0017] Figure 2 This is a schematic diagram of the SEI framework based on joint feature extraction according to an embodiment of the present invention;

[0018] Figure 3 The image shows the Hilbert 3D spectrum after VMD processing according to an embodiment of the present invention.

[0019] Figure 4 This is a time-frequency domain projection diagram according to an embodiment of the present invention;

[0020] Figure 5 This is a spectral time-domain projection diagram according to an embodiment of the present invention;

[0021] Figure 6 This is an amplitude-frequency domain projection diagram according to an embodiment of the present invention;

[0022] Figure 7 This is a schematic diagram of the DELM structure according to an embodiment of the present invention;

[0023] Figure 8 This is a schematic diagram illustrating the effect of the number of radiation sources k=3 on bispectral feature recognition according to an embodiment of the present invention;

[0024] Figure 9 This is a schematic diagram illustrating the impact of the number of radiation sources k=4 on bispectral feature recognition according to an embodiment of the present invention;

[0025] Figure 10 This is a schematic diagram illustrating the effect of the number of radiation sources k=5 on bispectral feature recognition according to an embodiment of the present invention;

[0026] Figure 11 This is a schematic diagram illustrating the impact of the number of radiation sources k=6 on bispectral feature recognition according to an embodiment of the present invention;

[0027] Figure 12 This is a diagram showing the recognition results of VMD spectral grayscale vectors when the number of radiation sources k=3 according to an embodiment of the present invention.

[0028] Figure 13 The image shows the recognition results of the VMD spectral grayscale vector when the number of radiation sources k=4 is according to an embodiment of the present invention.

[0029] Figure 14 The image shows the recognition results of the VMD spectral grayscale vector when the number of radiation sources k=5 is according to an embodiment of the present invention.

[0030] Figure 15 The image shows the recognition results of the VMD spectral grayscale vector when the number of radiation sources k=6 is according to an embodiment of the present invention.

[0031] Figure 16 This is a schematic diagram of the integrated CIDELM structure according to an embodiment of the present invention;

[0032] Figure 17 This is a schematic diagram illustrating the integrated CIDELM recognition performance according to an embodiment of the present invention;

[0033] Figure 18 This is a graph showing the recognition performance as a function of transmission and reception distance for a given number of sources k=3 according to an embodiment of the present invention.

[0034] Figure 19 This is a graph showing the change in recognition performance with transmission and reception distance for a radiation source number k=4 according to an embodiment of the present invention;

[0035] Figure 20 This is a graph showing the change in recognition performance as a function of transmission and reception distance for a radiation source number k=5 according to an embodiment of the present invention.

[0036] Figure 21 This is a graph showing the change in recognition performance with transmission and reception distance for a radiation source number k=6 according to an embodiment of the present invention;

[0037] Figure 22 This is a schematic diagram comparing the recognition effects of different methods according to embodiments of the present invention;

[0038] Figure 23 This is a schematic 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 Implementation

[0039] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0040] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0041] The following describes, with reference to the accompanying drawings, a method and apparatus for identifying specific radiation sources based on continuous learning and joint feature extraction according to embodiments of the present invention. First, the method for identifying specific radiation sources based on continuous learning and joint feature extraction according to embodiments of the present invention will be described with reference to the accompanying drawings.

[0042] Figure 1 This 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.

[0043] like Figure 1 As shown, this method for identifying specific radiation sources based on continuous learning and joint feature extraction includes the following steps:

[0044] Step S1: Acquire signals from multiple radiation sources and perform signal processing on the signals from the multiple radiation sources;

[0045] Step S2: Input the signals from multiple radiation sources after signal processing into multiple trained continuous incremental deep limit learning machines, and use the multiple trained continuous incremental deep limit learning machines as classifiers to output decisions.

[0046] Step S3: Use a voting algorithm to fuse the output decisions of a single continuous incremental deep extreme learning machine, and select the class with the highest confidence as the classification result to identify a specific radiation source.

[0047] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0048] Understandably, commonly used SEI schemes currently involve extracting features from preprocessed signals and then feeding them into a classifier for recognition. SEI frameworks based on joint feature extraction and incremental learning include... Figure 2As shown, the actual intercepted signal from the receiver is preprocessed to extract radio frequency fingerprint features. The purpose of fingerprint extraction is to amplify the individual differences between different radiation sources. This invention uses the Hilbert spectrum and higher-order spectral diagonal matrix processed by VMD as combined features of the radiation source signal, aiming to compensate for the shortcomings of single-scale identification. For a continuous feature data stream from K emitters, it is divided into training and test sets, and multiple improved deep extreme learning machines are trained simultaneously. Finally, a voting algorithm is used to fuse the output decisions of individual models, and the class with the highest confidence is selected as the final classification result.

[0049] Specifically, radiation source signal modeling is performed. Supervised radiation source identification can be simplified to a K-class decision problem, where the system input is a discrete sequence and the output is the discrimination probability for each sample. The intercepted signal r from the k-th radiation source... (k) (t) is represented as:

[0050]

[0051] In the formula, s(t) is the transmitted signal, h(t) is the equivalent channel impulse response, and n(t) is the channel additive noise. The in-phase and quadrature components output after USRP sampling can be expressed as:

[0052]

[0053] Furthermore, variational mode decomposition (VMD) reconstructs the original signal using intrinsic mode function (IMF) components, decomposing the original signal into sparse intrinsic mode functions (IMFs). This effectively suppresses the endpoint effects and mode confusion caused by EMD. VMD decomposes the original signal into several AM-FM signals:

[0054]

[0055] In the formula u k (t) represents the k-th modal component, A k (t) represents the instantaneous amplitude. For the signal phase, the core idea is to construct and solve a variational problem, decomposing the original signal into multiple intrinsic mode functions, with the constraint variational expression as follows:

[0056]

[0057] {u k (t)}={u1(t),u2(t),...,u K {Φ(t)} is the set of modal components, {Φ k (t)}={f1(t),f2(t),...,fk Let (t)} be the set of modal center frequencies, where δ(t) is the impulse response function, and S(t) is the signal to be decomposed. VMD transforms the mode decomposition problem into an unconstrained variational solution, introducing a quadratic penalty factor α and a Lagrange multiplier λ(t) to obtain the optimal solution. The augmented Lagrange expression is:

[0058]

[0059] The alternating direction multiplier algorithm is used for update iteration. and λ n+1 To solve for the saddle point in equation (6), and thus obtain the optimal solution set {u} in the frequency domain. k (ω)},{f k (ω)} and {λ k The update formula for (ω)} under positive frequency conditions is:

[0060]

[0061]

[0062]

[0063] In the formula, τ is the update coefficient of the Lagrange operator. For Fourier transform, repeat the update equations (6)-(8) until they meet the convergence condition:

[0064]

[0065] ε represents the discrimination precision, which is set to 10 in this paper. -6 The iteration stops when the convergence condition is met. The discrete signal sequence S(m) is represented by k modes as follows:

[0066]

[0067] Perform Hilbert transform on the obtained k modes to construct the analytic signal z(m) and calculate the instantaneous frequency f. i (m) and instantaneous amplitude A i (m):

[0068]

[0069]

[0070] In the formula The VMD-Hilbert spectrum is represented as follows:

[0071]

[0072] In the formula Representing the real part, this paper converts the projected VMD-Hilbert spectrum into a grayscale image. The energy value H(i,j) at the (i,j)th time-frequency point within the spectrum can then be converted into the corresponding grayscale value G(i,j).

[0073]

[0074] l represents the number of bits in the grayscale image. The term "round down" is used to propose a method that utilizes the statistical characteristics of grayscale images as radio frequency fingerprints. First, a grayscale histogram is constructed, which can accurately reflect the frequency of occurrence of each grayscale pixel in the image and the grayscale relationship. The distribution of time-frequency energy in the VMD-Hilbert spectrum is then mapped to a two-dimensional coordinate system. Figure 3 , Figure 4 , Figure 5 and Figure 6 VMD-Hilbert projection of the intercepted signal from USRP-2922. Figure 3 For the Hilbert 3D spectrum after VMD processing, Figure 4 For time-frequency domain projection, Figure 5 For spectral time-domain projection, Figure 6 It is a projection of the amplitude frequency domain.

[0075] Furthermore, higher-order cumulants can effectively suppress additive Gaussian colored noise and describe the degree of deviation of the signal from Gaussian white noise. The sampling sequence r... (l) The k-th order cumulant of (m) is defined as:

[0076] C kx (τ1,τ2,...,τ k-1 )=cum{x(n),x(n+τ1),...,x(n+τ k-1 (15)

[0077] When it satisfies the condition of absolute summability:

[0078]

[0079] The spectrum of the k-th order cumulant is defined as the (k-1)-th order Fourier transform of the k-th order cumulant:

[0080]

[0081] When k=3, the spectrum of higher-order cumulants is called a bispectrum:

[0082]

[0083] This invention uses the sliding window method to estimate higher-order cumulants, and the steps are shown in Table 1:

[0084] Table 1. Sliding Window Method for Higher-Order Spectrum Estimation

[0085]

[0086] Furthermore, continuous learning based on samples is performed using IDELM. The Extreme Learning Machine (ELM) typically has only one hidden layer, which randomly generates input weights and hidden node parameters, while the output weights are obtained analytically by calculating the generalized inverse matrix. ELM determines the optimal solution through single-step least squares error (LSE), without requiring gradient backpropagation. An ELM with L hidden nodes is represented as:

[0087]

[0088] In the formula, G(·) is the activation function, and w i =[w i,1 ,w i,2 ,...,w i,n ] T ∈R n β represents the weights between the input layer and the i-th hidden node. i For the output weights, b i w is the weight bias for the i-th hidden node. i ·x j Indicates w i and x j The inner product of the two, the solution of equation (19) can be expressed by matrix:

[0089] T=Hβ (20)

[0090] In the formula:

[0091]

[0092] In the formula, H is the output matrix, β is the output weight, T is the expected output, and the loss function of ELM is:

[0093]

[0094] The training process of a network is equivalent to solving the LSE solution of the linear system Hβ=T.

[0095]

[0096] In the formula The norm is minimal and unique, T = [t1,...,t] N ], H + Let H be the Moore–Penrose generalized inverse of matrix H. TWhen H is not singular, then H + =(H T H) -1 H T Otherwise H + =H T (HH T ) -1 .

[0097] Furthermore, feature sorting is based on Continuous Incremental Deep Extreme Learning Machine (CIDELM). Considering that in real-world scenarios, the intercepted radiation source signals are often continuous data streams, this paper introduces a continuous learning mechanism on the basis of Deep Extreme Learning Machine (DELM) to design a Continuous Incremental Deep Extreme Learning Machine (CIDELM). The parameters of DELM are continuously updated according to the input order of the samples, which improves classification efficiency and saves computational resources. The training of CIDELM is divided into two stages: unsupervised feature representation of the hidden layers and supervised label classification of the output layer. In the first stage, this invention uses multiple feedforward hidden layers in series as an autoencoder to perform sparse representation of the input samples and mine the hidden information of the signal features. In the second stage, a single-layer ELM is used for supervised regression to output the final classification result.

[0098] Depend on Figure 7 It can be seen that the output H1 of the first hidden layer of the Deep Extreme Learning Machine (DELM) is expressed as:

[0099] H1=G(w 1,2 ·H+B1) (24)

[0100] In the formula w 1,2 This represents the weight matrix between the first and second hidden layers. H1 represents the output of the input after passing through the first hidden layer, and B1 represents the bias of the first hidden layer. The initial weights of the first hidden layer are β1 = H1. + T1, the expected output matrix of the second hidden layer can be expressed as H 2E =Tβ1 + Define the augmented matrix w H2 =[B2,w 1,2 In the formula, B2 is the input bias of the second hidden layer, then w H2 It can be calculated using the following formula:

[0101]

[0102] In the formula, M2 = [1 H1] T In the formula, 1 is a column vector of all scalar 1, and G -1(·) is the inverse function of the activation function G(·), and the actual output of the second hidden layer is H2 = G(w H2 ·M2), the output weights of the second hidden layer β2=H2 + T2, inductively calculate the expected output H of the i-th hidden layer. iE Actual output H i And the output weights, the DELM output with P hidden layers is represented as:

[0103] f P,L (x)=H P β P (26)

[0104] This invention considers feeding labeled signal samples used for training into a DELM algorithm to perform unsupervised sequence training layer by layer. Let the (k+1)th input sample be... The number of samples in each group is N. k+1 Then the output weight β (k+1) The update formula is:

[0105]

[0106] In the formula The output of the i-th hidden layer corresponding to time step k is:

[0107]

[0108] When training CIDELM, the first set of samples is used as input, and the initial output matrix of each hidden layer is calculated by equation (28) above. Then, the sample data for K-1 time steps are input sequentially, and the output H0 and initial weights at k=0 are grouped together. The model is updated, and the final weights are calculated:

[0109]

[0110] Furthermore, the experimental results and analysis of this invention are as follows:

[0111] Experimental conditions: Currently, commonly used SEI signal simulation methods simulate the nonlinear distortion differences between individual radiation sources using Taylor polynomials. This invention uses a USRP based on GNU Radio as the transceiver. The signal is transmitted in a real laboratory channel. The algorithm is verified by changing the transmission and reception distance, modulation scheme, and carrier frequency. The connection style is I / Q dual-channel. Six USRP-2922 units manufactured in the same batch are used as signal generators, and one USRP-B210 is used as the receiving device. A logarithmic spiral antenna is selected as the transmitting antenna to increase the effective range. The signal modulation schemes are BPSK and BFSK, the carrier frequencies are set to 500MHz and 1GHz, and the bandwidth is 100KHz. The number of downsampling points for each modulation scheme and each carrier frequency is 106. The downsampling rate of the USRP-B210 is 1MHz. The k-fold verification method is used to randomly divide the samples into training and test sets. The proposed fingerprint extraction method is used to calculate two types of features for the same signal sample. The number of transformation points is set to 1000. The calculated feature vectors are normalized and singular values ​​are removed before being sequentially input into CIDELM.

[0112] As demonstrated above, the performance of CIDELM is affected by the initial weights, initial biases, number of hidden layers, and number of hidden nodes. However, the initial weights and initial biases are typically generated randomly. Therefore, this invention employs a parameter optimization strategy to adjust the number of hidden layers and hidden nodes to obtain the optimal parameter combination. Figure 8 , Figure 9 , Figure 10 , Figure 11 The effects of two types of hyperparameters on bispectral feature recognition are presented. The number of nodes in the input and output hidden layers is assumed to be 2000 by default. Figure 8 For the number of radiation sources k=3, Figure 9 The number of radiation sources is k=4. Figure 10 The number of radiation sources is k=5. Figure 11 The number of radiation sources is k = 6.

[0113] from Figure 8 , Figure 9 , Figure 10 , Figure 11As can be seen, the CIDELM-based single-feature real-time recognition model achieves high accuracy by optimizing the matching of two-dimensional parameters when the number of radiation sources k changes. The optimal recognition rate is 97% when k=3, 94% when k=4, 91% when k=5, and 88% when k=6. When the number of hidden layers in CIDELM is 2-4, and the number of hidden nodes is 500 and 1000, it exhibits high recognition accuracy. This is because the hidden layer model is more compact in this case, resulting in better sparse encoding of the input features. Increasing the number of hidden layers and hidden nodes increases the probability of the layer output matrix not being of full rank, leading to the accumulation of errors in the calculated MP inverse matrix layer by layer. When solving the function approximation problem for small batches of data, CIDELM can reduce the parameter search space and improve recognition accuracy and speed by setting appropriate hidden nodes and hidden layers. Moreover, CIDELM processes samples in batches and updates the model parameters online through continuous training, significantly enhancing the algorithm's adaptability to data scale. Figure 12 , Figure 13 , Figure 14 , Figure 15 The recognition results of VMD spectral grayscale vectors are presented. As can be seen from the figure, by setting the number of hidden layers and the number of hidden nodes, the accuracy of CIDELM in recognizing VMD spectral grayscale vectors can reach over 91%. Figure 12 For the number of radiation sources k=3, Figure 13 The number of radiation sources is k = 4 (c). Figure 14 Number of radiation sources k = 5 Figure 15 The number of radiation sources is k = 6.

[0114] Furthermore, joint feature discrimination based on ensemble CIDELM suffers from unstable network outputs due to the randomly generated initial parameters of CIDELM, affecting model reliability. To address this, this invention constructs an ensemble CIDELM algorithm that combines the outputs of a limited number of similar network structures. Through simultaneous training of multiple models, two types of features from the same signal segment are fed into K CIDELMs for training and testing. A majority voting algorithm (Boyer-Moore Algorithm, BMA) is used to vote on the results of each CIDELM, ultimately determining the radiation source category based on the number of votes received. The ensemble CIDELM outperforms the classification performance of a single model, exhibiting greater stability and reliability. Figure 16 The structure is for integrating CIDELM.

[0115] Figure 17To integrate CIDELM's algorithm comparison on the dataset, this invention verifies the algorithm's robustness and compatibility by changing the modulation scheme, transmission / reception distance, and carrier frequency. The transmission / reception distance is set to 4 meters. Figure 17 As can be seen, when the number of radiation sources is less than 5, the recognition accuracy for combinations of different carrier frequencies and modulation methods can reach more than 90%, and the change in recognition accuracy is less affected by the modulation method and carrier frequency; when the number of radiation sources is greater than 6, the recognition accuracy can still reach more than 86%, proving that the joint feature extraction scheme proposed in this invention can show good compatibility on the established dataset.

[0116] Figure 18 , Figure 19 , Figure 20 , Figure 21 The algorithm's performance was shown on datasets with different parameter combinations. When the transmit and receive distance changed, the recognition performance slowly decreased. This may be because in a laboratory environment, the operation of other devices affected the wireless transmission environment and increased channel noise. However, the overall recognition accuracy could still reach over 80%, indicating that the algorithm of this invention can meet the requirements of real-time high-precision recognition.

[0117] Methods 1, 2, and 3 were selected and compared with the proposed method. Method 1 used the Hilbert time entropy as the RFF and employed KNN for classification; Method 2 used the Hilbert spectrum as the RFF and employed DRN for classification; Method 3 used the PCA algorithm to perform dimensionality reduction mapping on the signal's MAE. The comparison results are as follows: Figure 22 As shown, the number of radiation sources is set to 6. This invention uses dual-mode feature extraction based on time-frequency domain analysis, and proposes an improved ELM that has better recognition performance for one-dimensional vectors, with an average recognition accuracy approximately 2% to 4% higher than the methods described above.

[0118] Furthermore, the computational complexity of the algorithm is analyzed. The computational complexity of the algorithm in this invention mainly comes from the front-end dual-mode RF fingerprint extraction and the iterative training of the back-end classifier. The iteration time and average recognition time of the algorithm are selected to measure the time complexity of the algorithm. The operating system of the experimental computer is an Intel(R) Core(TM) i7-9750H CPU with 16GB of RAM and an NVIDIA GeForce RTX 3080 GPU. Table 2 shows the time complexity analysis of the algorithm.

[0119] Table 2 Algorithm Time Complexity Analysis

[0120]

[0121] As shown in Table 2, the SEI method with dual-mode feature extraction and online incremental learning proposed in this invention has a faster model training speed and a shorter recognition time. Since it does not rely on backpropagation of gradients to update global parameters, the CIDELM-based classification method has high real-time performance. Compared to the huge training cost of DL models, the computational complexity of the algorithm in this invention is lower.

[0122] Therefore, this invention proposes a new SEI method based on dual-feature sorting and continuous incremental deep limit learning to address the issues of low generalization due to sample size limitations and high model training costs caused by dynamic sample updates in SEI algorithms under small sample conditions. First, VMD spectral grayscale vectors and bispectral matrix diagonals are constructed from the signal to be identified. Then, CIDELM is constructed to dynamically update the weights of continuous samples, reducing the computational cost of the model and enabling supervised identification of multiple batches of samples. Experimental results show that the proposed algorithm can achieve online real-time identification of multiple USRPs, with an accuracy rate that meets practical requirements. Furthermore, it is unaffected by factors such as modulation scheme and carrier frequency, exhibiting strong robustness.

[0123] The specific radiation source identification method based on continuous learning and joint feature extraction in this invention first constructs the VMD spectral grayscale linear vector and bispectral matrix diagonal value from the signal to be identified, and then constructs CIDELM to dynamically update the weights of continuous samples, reducing the computational cost of the model and realizing supervised identification of multiple batches of samples.

[0124] To achieve the above embodiments, such as Figure 23 As shown, this embodiment also provides a specific radiation source identification device 10 based on continuous learning and joint feature extraction. The device 10 includes: a signal acquisition module 100, a signal input module 200, and a fusion output module 300.

[0125] The signal acquisition module 100 is used to acquire signals from multiple radiation sources and perform signal processing on the signals from the multiple radiation sources.

[0126] The signal input module 200 is used to input the signals from multiple radiation sources after signal processing into multiple trained continuous incremental depth limit learning machines, and to use the multiple trained continuous incremental depth limit learning machines as classifiers to output decisions.

[0127] The fusion output module 300 is used to fuse the output decisions of a single continuous incremental deep extreme learning machine using a voting algorithm, and select the class with the highest confidence as the classification result to identify a specific radiation source.

[0128] Furthermore, it also includes a training module for training multiple consecutive incremental deep limit learning machines based on signals from multiple sample radiation sources.

[0129] The specific radiation source identification device based on continuous learning and joint feature extraction according to embodiments of the present invention can realize online real-time identification of multiple USRPs, the identification accuracy can meet the actual needs, and it is not affected by factors such as modulation method and carrier frequency, and has strong robustness.

[0130] It should be noted that the foregoing explanation of the specific radiation source identification method embodiment based on continuous learning and joint feature extraction also applies to the specific radiation source identification device based on continuous learning and joint feature extraction in this embodiment, and will not be repeated here.

[0131] It is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for identifying specific radiation sources based on continuous learning and joint feature extraction, characterized in that, Includes the following steps: Acquire signals from multiple radiation sources and perform signal processing on the signals from the multiple radiation sources; The signals from the multiple radiation sources after signal processing are input into multiple trained continuous incremental deep limit learning machines, and the multiple trained continuous incremental deep limit learning machines are used as classifiers to output decisions. The output decisions of a single continuous incremental deep extreme learning machine are fused using a voting algorithm, and the class with the highest confidence is selected as the classification result to identify a specific radiation source. Before inputting the processed signals from the plurality of radiation sources into the trained plurality of continuous incremental depth limit learning machines, the method further includes: training the plurality of continuous incremental depth limit learning machines based on the signals from the plurality of sample radiation sources. The training of the multiple continuous incremental deep extreme learning machines based on signals from multiple sample radiation sources includes: The signals of multiple sample radiation sources were intercepted using the Universal Software Radio Peripheral Platform (USRP), and the signals of the multiple sample radiation sources were preprocessed and radio frequency fingerprint features were extracted. The Hilbert time-frequency energy spectrum obtained by processing the radio frequency fingerprint features using VMD is projected and reduced in dimension, and then converted into a grayscale vector to obtain a continuous feature data stream; and a diagonal matrix is ​​obtained by analyzing the radio frequency fingerprint features using high-order spectral vectors; wherein, the continuous feature data stream of the signals from the multiple sample radiation sources includes a training set and a test set; The continuous feature data stream and the diagonal matrix are input into multiple continuous incremental depth limit learning machines for training, so as to obtain the multiple trained continuous incremental depth limit learning machines. The design incorporates a continuously incremental deep extreme learning machine (DELM), which continuously updates the DELM parameters according to the input order of the samples. The training of CIDELM is divided into two stages: unsupervised feature representation of the hidden layers and supervised label classification of the output layers. In the first stage, multiple feedforward hidden layers are concatenated as an autoencoder to perform sparse representation of the input samples and mine the hidden information of the signal features. In the second stage, a single-layer ELM is used for supervised regression to output the final classification result. The output H1 of the first hidden layer of the Deep Extreme Learning Machine (DELM) is represented as: H1=G(in 1,2 H+B1)(24) In the formula, H is the output matrix, and w 1,2 This represents the weight matrix between the first and second hidden layers. H1 represents the output of the input after passing through the first hidden layer, and B1 represents the bias of the first hidden layer. The initial weights of the first hidden layer are β1 = H1. + T1, the expected output matrix of the second hidden layer is represented as H. 2E =Tβ1 + T represents the desired output, and H represents the desired output. + Let w be the Moore–Penrose generalized inverse of matrix H. H2 =[B2,w 1,2 In the formula, B2 is the input bias of the second hidden layer, then w H2 Calculated by the following formula: In the formula, M2 = [1 H1] T In the formula, 1 is a column vector of all scalar 1, and G -1 (·) is the inverse function of the activation function G(·), and the actual output of the second hidden layer is H2 = G(w H2 ·M2), the output weights of the second hidden layer β2=H2 + T2, inductively calculate the expected output H of the i-th hidden layer. iE Actual output H i And the output weights, the DELM output with P hidden layers is represented as: f P,L (x)=H P β P (26) Let the (k+1)th input sample be D. k+1 =(x i ,t i ), i= The number of samples in each group is N. k+1 Then the output weight β (k+1) The update formula is: In the formula The output of the i-th hidden layer corresponding to time step k is: When training the CIDELM (Continuously Incremental Deep Learning Machine), the first set of samples is used as input, and the initial output matrix of each hidden layer is calculated by equation (28). Then, the sample data for K-1 time steps are input sequentially, and the output H0 and initial weights at k=0 are grouped together. The model is updated, and the final weights are calculated: 。 2. A specific radiation source identification device based on continuous learning and joint feature extraction using the method described in claim 1, characterized in that, include: The signal acquisition module is used to acquire signals from multiple radiation sources and perform signal processing on the signals from the multiple radiation sources; The signal input module is used to input the processed signals from the multiple radiation sources into a trained multi-continuous incremental deep learning machine, and to use the trained multi-continuous incremental deep learning machine as a classifier to output a decision. The fusion output module is used to fuse the output decisions of a single continuous incremental deep extreme learning machine using a voting algorithm, and select the class with the highest confidence as the classification result to identify a specific radiation source. It also includes a training module for training the multiple consecutive incremental deep limit learning machines based on signals from multiple sample radiation sources.

Images