Multi-feature radiation source individual identification method

Abstract

The invention relates to a multi-feature radiation source individual identification method, belongs to the technical field of radio frequency radiation source identification and deep learning, and solves the technical problems that single feature information of a UHF (Ultra High Frequency) waveband is insufficient to comprehensively represent subtle feature differences and radio frequency fingerprint features are difficult to distinguish radiation source individuals in a low signal-to-noise ratio environment. Comprising the following steps: acquiring communication signal data emitted by a to-be-identified radiation source, and preprocessing to obtain a plurality of corresponding fused images; and inputting the plurality of fusion images into a trained AttenSE-ResNet50 model, performing classification identification to obtain a classification result probability vector of each fusion image, and performing voting based on each classification result probability vector to obtain an identification result of the radiation source device individual. And accurate identification of the radiation source equipment in a complex electromagnetic environment is realized.

Description

Technical Field

[0001] The present invention relates to the field of radio frequency radiation source identification and deep learning technology, and in particular to a method for identifying individual radiation sources with multiple characteristics. Background Art

[0002] In recent years, with the rapid development of wireless communication technology, an increasing number of communication devices have been connected to the Internet. However, the open nature of wireless channels also presents various security threats. Wireless device identification technology leverages the hardware differences of each communication device to determine the source of transmitted signals based on their distinct characteristics, enabling targeted monitoring and tracking of critical communication devices and electronic media. This technology can effectively distinguish between different radiating devices of the same manufacturer, batch, and signal specifications, making it crucial for distinguishing counterfeit devices and unidentified radiating sources.

[0003] RF fingerprint feature extraction is a key component of wireless device identification technology. This feature extraction method extracts unique fingerprint information from communication signals, enabling identification of different radiating devices. However, challenges remain in the development of RF fingerprint feature extraction methods. First, due to the complexity and interference of wireless communication signals, mitigating the impact of environmental noise and interference on RF fingerprint extraction accuracy remains a challenge. Furthermore, minimizing algorithm complexity while maintaining extraction effectiveness is a challenge. Furthermore, feature extraction algorithms must balance real-time performance, robustness, and complexity.

[0004] Traditional feature engineering methods rely heavily on feature extraction algorithms, requiring extensive knowledge and experience in signal processing. Furthermore, some hardware features are interrelated and difficult to extract individually. Therefore, more effective algorithms are needed to identify radiating devices. In recent years, deep learning has experienced rapid development and has been gradually applied to the identification of individual communication radiating devices, achieving significant results in many scenarios. Deep learning methods can automatically learn to discover patterns in RF signals and extract higher-dimensional, more representative features. Furthermore, various deep learning methods can address the challenges of identification in complex environments. These methods typically use signal time-frequency plots, bispectral plots, or constellation trajectory plots as input, training the model using large-scale training sets to produce classification models with excellent recognition and transfer capabilities. However, recognition methods that rely solely on information in a single transform domain perform differently in different scenarios, channels, and noise conditions. Furthermore, since the selected features in a particular domain cannot be guaranteed to be optimal, suboptimal recognition results are often achieved. Summary of the Invention

[0005] In view of the above analysis, an embodiment of the present invention aims to provide a multi-feature radiation source individual identification method to solve the technical problems of the existing RF radiation source device identification method in that the single feature information in the UHF band is not sufficient to fully represent the subtle feature differences and the RF fingerprint feature is difficult to perform individual identification of the radiation source device in a low signal-to-noise ratio environment.

[0006] The purpose of the present invention is mainly achieved through the following technical solutions:

[0007] The present invention provides a method for identifying an individual radiation source with multiple characteristics, comprising the following steps:

[0008] Acquire communication signal data emitted by the radiation source to be identified, and obtain corresponding multiple fused images after preprocessing;

[0009] The multiple fused images are input into the trained AttenSE-ResNet50 model for classification and recognition to obtain a classification result probability vector for each fused image, and voting is performed based on each classification result probability vector to obtain the identification result of the individual radiation source device.

[0010] Furthermore, the acquiring of communication signal data and preprocessing to obtain corresponding multiple fused images includes:

[0011] After trimming the signal-free portion of the signal sample data, normalizing the amplitude and performing segmentation to obtain corresponding multiple segmented signals;

[0012] Extract the time-frequency graph features and differential bispectral graph features of each segmented signal;

[0013] Visualizing the time-frequency graph features and the differential bispectrum graph features to obtain corresponding time-frequency graphs and differential bispectrum graphs;

[0014] Filling the time-frequency graph and the differential bispectrum graph internally to obtain a corresponding three-channel image;

[0015] The three-channel images corresponding to the time-frequency diagram and the differential bispectrum of each segmented signal are spliced ​​and fused to obtain a three-channel spliced ​​fused image corresponding to each segmented signal;

[0016] The size of the first channel of each of the three-channel stitched fusion images is adjusted using a Resize function to obtain multiple fusion images corresponding to the multiple segmented signals.

[0017] Furthermore, the AttenSE-ResNet50 model includes, in sequence, a convolutional initial processing module, first, second, third, and fourth depth residual units, a classification pre-processing module, and a fully connected classification layer; wherein,

[0018] The convolution initial processing module includes a convolution layer, a batch normalization and ReLU6 nonlinear activation layer, and a maximum pooling layer in sequence, which is used to perform a preliminary convolution operation on the input fusion image, extract the initial features, perform batch normalization and ReLU6 nonlinear transformation, and obtain the basic feature map after maximum pooling downsampling;

[0019] The first depth residual unit includes first, second, and third residual modules with the same structure, which are cascaded in sequence; the second depth residual unit includes fourth, fifth, sixth, and seventh residual modules with the same structure, which are cascaded in sequence; the third depth residual unit includes eighth to thirteenth residual modules with the same structure, which are cascaded in sequence; the fourth depth residual unit includes fourteenth to sixteenth residual modules with the same structure, which are cascaded in sequence; the first to fourth depth residual units are used to obtain the first fingerprint feature of the corresponding radiation source device individual based on the basic feature map;

[0020] The classification pre-processing module includes a global average pooling layer, a flattening layer, and a dropout layer in sequence, and is used to obtain a second fingerprint feature of the corresponding radiation source device based on the first fingerprint feature of the radiation source device;

[0021] The fully connected classification layer is used to output a classification result probability vector based on the second fingerprint feature of the individual radiation source device; and perform a voting decision based on the classification result probability vector to obtain the ID identification of the individual radiation source device that emits the communication signal data.

[0022] Furthermore, the first to sixteenth residual modules each include a first, second and third convolutional blocks; the first and second convolutional blocks are followed by batch normalization and a ReLU6 nonlinear activation layer; the third convolutional block is followed by a batch normalization layer, a two-layer attention module, an SE attention module, a skip mapping connection between the residual module input and the SE module output, and a ReLU6 nonlinear activation; each residual module is used to perform channel transformation and size transformation on the input image.

[0023] Furthermore, the dual-layer attention module includes a channel attention submodule and a spatial attention submodule connected in sequence;

[0024] The channel attention submodule collects comprehensive features of the input feature map through the parallel maximum pooling layer and average pooling layer to obtain two corresponding feature maps, and uses the shared perception layer to assign weight values ​​to the importance of the two input feature maps to obtain two feature vectors; the two feature vectors are added and then activated by Sigmoid to generate a channel attention weight feature; then the channel attention weight feature is multiplied by the input feature map to obtain a channel attention weight vector map;

[0025] The spatial attention submodule performs maximum pooling and average pooling on the channel attention weight vector map, connects the outputs of the two, and performs convolution operation and Sigmoid activation to obtain a spatial attention weight feature vector map;

[0026] Multiply the spatial attention weight feature vector map with the channel attention weight vector map to obtain the enhanced feature expression map output by the double-layer attention module as the input of the SE attention module.

[0027] Furthermore, the SE attention module performs global pooling on each channel of the enhanced feature expression graph, and then sequentially performs the first full connection, ReLU activation, second full connection and Sigmoid activation to generate channel weights in the [0,1] interval;

[0028] The SE attention module includes a global pooling layer, a first fully connected layer, a ReLU activation layer, a second fully connected layer and a Sigmoid activation layer;

[0029] The global average pooling layer is used to compress the spatial information of the input enhanced feature expression map to obtain a channel descriptor;

[0030] The first fully connected layer and the second fully connected layer are used to perform nonlinear transformation on the channel descriptor to generate channel weights;

[0031] The ReLU activation layer is used to introduce nonlinearity; the Sigmoid activation layer is used to map the channel weight to the [0,1] interval to obtain a channel attention weight map;

[0032] The SE attention module performs channel weighting on the enhanced feature expression map, enhancing key feature channels and suppressing redundant feature channels.

[0033] Furthermore, the extracting of the time-frequency graph features and the differential bispectral graph features of each segmented signal includes:

[0034] Perform short-time Fourier transform on each segmented signal to obtain the time-frequency graph features corresponding to the segmented signal;

[0035] Each segmented signal is subjected to n-point differentiation to obtain a differential signal, and the differential signal is subjected to bispectral transformation to obtain differential bispectral features corresponding to the segmented signal.

[0036] Furthermore, the AttenSE-ResNet50 model is trained based on the following sample training set:

[0037] Signal data with BPSK modulation type under different signal-to-noise ratios are collected and preprocessed to obtain signal sample data, which are then combined with corresponding sample labels to form a sample training set; wherein the sample label is the ID identifier of the individual radiation source device.

[0038] Furthermore, the sample training set is divided into a training set, a test set and a validation set according to a preset ratio;

[0039] Set the training batch, training round, learning rate, minimum learning rate, learning rate annealing period and weight decay coefficient;

[0040] Training the AttenSE-ResNet50 model using the training set; using a cross-entropy loss function during training, and dynamically adjusting the learning rate using a cosine annealing learning rate scheduler;

[0041] Continuously adjust and update model parameters through back propagation and gradient descent optimization algorithms;

[0042] Until the cross entropy loss function converges or the training round is completed, the AttenSE-ResNet50 model parameters are saved to obtain a trained AttenSE-ResNet50 model.

[0043] Furthermore, a confidence threshold is set for the probability distribution output by the AttenSE-ResNet50 model;

[0044] If the maximum probability value in the classification result probability vector is higher than the confidence threshold, it is determined that the input communication signal data belongs to the category corresponding to the trained communication signal data; otherwise, it is determined that it may belong to the communication signal data emitted by an untrained radiation source device individual, or the aging of the radiation source device and extreme electromagnetic interference.

[0045] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0046] 1. This invention characterizes the inherent fingerprint characteristics of devices from different perspectives by fusing the local features of the time-frequency graph and the high-order nonlinear features in the differential bispectrum. This solves the problem that single feature information is insufficient to fully characterize the subtle feature differences of devices. Compared with existing methods that rely solely on single transform domain information, it can more accurately identify different radiation source devices. It also avoids the problem that single feature information of RF radiation sources in the UHF band is insufficient to fully characterize the subtle feature differences of devices.

[0047] 2. This invention introduces a dual attention mechanism module and an SE module, which enables the network to selectively enhance the feature responses of key channels and spatial positions, suppress irrelevant features and noise interference, and accurately extract effective features even in low signal-to-noise ratio environments. This effectively solves the technical problem of the difficulty in distinguishing RF fingerprint features in low signal-to-noise ratio environments in existing technologies and improves recognition accuracy.

[0048] 3. The AttenSE-ResNet50 network model structure constructed by the present invention solves the gradient vanishing problem in the deep neural network training process through the four-level first, second, third and fourth depth residual units distributed as (3, 4, 6, 3) and the skip connection design, thereby improving the training efficiency and classification performance of the model. Compared with the traditional complex and difficult-to-train model, the training method uses less sample data.

[0049] 4. This invention uses deep learning methods to automatically learn the patterns in radio frequency signals and abstract higher-dimensional and more representative features. It does not need to rely heavily on manual extraction and signal processing experience like traditional feature engineering methods, thus ensuring the effectiveness and efficiency of feature extraction.

[0050] 5. In a 0dB signal-to-noise ratio environment, the individual radiation source identification method of the present invention significantly improves the classification accuracy compared to the recognition algorithms using bispectrum, time-frequency diagram or differential bispectrum alone.

[0051] In the present invention, the above-mentioned technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of the present invention will be described in the following description, and some advantages will become apparent from the description or be learned through practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The accompanying drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like parts throughout the drawings.

[0053] Figure 1 This is a flow chart of a method for identifying individual radiation sources with multiple characteristics according to an embodiment of the present invention;

[0054] Figure 2 Schematic diagram of the time-frequency characteristics of the measured signal collected by the receiver in a 5dB signal-to-noise ratio environment in an embodiment of the present invention;

[0055] Figure 3 Comparison of differential bispectral contour maps of measured signals collected by two receiver devices 1 and 2 in a 5dB signal-to-noise ratio environment in an embodiment of the present invention;

[0056] Figure 4 This is an image obtained by splicing and fusing the time-frequency graph features and the differential bispectral graph features in an embodiment of the present invention;

[0057] Figure 5 Schematic diagram of the network structure of the AttenSE-ResNe50 model in an embodiment of the present invention;

[0058] Figure 6 Schematic diagram of the improved residual module network structure in an embodiment of the present invention;

[0059] Figure 7 This is a schematic diagram of the network structure of the dual attention module in an embodiment of the present invention;

[0060] Figure 8 Schematic diagram of the network structure of the SE attention module in an embodiment of the present invention;

[0061] Figure 9a Schematic diagram of the recognition performance of four devices under different signal-to-noise ratios of receiver signals in an embodiment of the present invention;

[0062] Figure 9b Schematic diagram of the recognition performance of four devices under different signal-to-noise ratios of mobile phone Bluetooth signals in an embodiment of the present invention;

[0063] Figure 10 Schematic diagram of recognition performance of different networks under different signal-to-noise ratios in an embodiment of the present invention. DETAILED DESCRIPTION

[0064] The preferred embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, and are not used to limit the scope of the present invention.

[0065] The present invention aims to address the technical issues with existing RF radiation source device identification technology: single characteristic information is insufficient to fully represent subtle characteristic differences in the UHF (Ultra High Frequency) band, which covers the 300MHz to 3GHz radio spectrum, and RF fingerprint characteristics are difficult to distinguish in low signal-to-noise ratio environments. This invention proposes a multi-characteristic radiation source individual identification method to enhance the ability to identify individual RF radiation sources.

[0066] The method of the present invention extracts the local features of the time-frequency graph of the acquired communication signal data and the high-order nonlinear features in the differential bispectrum, and performs feature splicing and fusion in the spatial dimension to obtain the corresponding fused image; then the fused image is input into the AttenSE-ResNet50 model constructed and trained in the present invention to perform deep feature extraction, integrate the multi-dimensional feature information of the signal, make full use of the instantaneous frequency changes in the time-frequency domain and the nonlinear characteristics of the high-order spectral domain of the differential bispectrum, and improve the model's sensitivity to signal features and its ability to recognize complex signals. The AttenSE-ResNet50 model introduces a dual attention module and SE

[0067] The Squeeze-and-Excitation (Squeeze-and-Excitation) attention module enables the network to selectively enhance the feature responses of key channels and spatial positions, and suppress irrelevant features and noise interference. The method of the present invention effectively solves the problem of incomplete recognition ability due to insufficient information and a single feature dimension, and also significantly improves the feature differentiation ability in a low signal-to-noise ratio environment, providing stronger robustness and accuracy for the individual identification of radiation source devices. In a 0dB signal-to-noise ratio environment, compared with the recognition algorithm using bispectrum, time-frequency diagram or differential bispectrum alone, the recognition performance of the multi-feature radiation source individual identification method of the present invention has increased by 70.42%, 8.33%, and 5.02%, respectively.

[0068] A specific embodiment of the present invention discloses a method for identifying individual radiation sources with multiple characteristics, such as Figure 1 As shown, the following steps are included:

[0069] Step S1: Acquire communication signal data emitted by the radiation source to be identified, and obtain corresponding multiple fused images after preprocessing;

[0070] Step S2: Input the multiple fused images into the trained AttenSE-ResNet50 model to perform classification and recognition to obtain a classification result probability vector for each fused image, and vote based on each classification result probability vector to obtain an identification result of the individual radiation source device.

[0071] Step S1 includes steps S11-S12.

[0072] Step S11: Acquire communication signal data emitted by the radiation source to be identified.

[0073] Use an ultra-short wave receiver to receive the communication signal data sent by the radiation source device to be identified. The receiver configuration is as follows:

[0074] (1) Receiving Antenna: Select an appropriate receiving antenna to ensure that the receiver's operating frequency band covers the signal frequency band of the individual radiating device to be identified. For UHF signals, select a UHF receiving antenna.

[0075] (2) Receiver settings: Configure the receiver parameters, such as center frequency, sampling rate, bandwidth, etc., to match the characteristics of the radiation source signal.

[0076] The collected signal data is stored in a storage device for subsequent preprocessing and recognition processing.

[0077] Step S12: pre-process the communication signal data emitted by the radiation source to be identified to obtain a corresponding plurality of fused images.

[0078] The acquiring of communication signal data and preprocessing to obtain corresponding multiple fused images includes:

[0079] After trimming the signal-free portion of the signal sample data, normalizing the amplitude and performing segmentation to obtain corresponding multiple segmented signals;

[0080] Extract the time-frequency graph features and differential bispectral graph features of each segmented signal;

[0081] Visualizing the time-frequency graph features and the differential bispectrum graph features to obtain corresponding time-frequency graphs and differential bispectrum graphs;

[0082] Filling the time-frequency graph and the differential bispectrum graph internally to obtain a corresponding three-channel image;

[0083] The three-channel images corresponding to the time-frequency diagram and the differential bispectrum of each segmented signal are spliced ​​and fused to obtain a three-channel spliced ​​fused image corresponding to each segmented signal;

[0084] The size of the first channel of each of the three-channel stitched fusion images is adjusted using a Resize function to obtain multiple fusion images corresponding to the multiple segmented signals.

[0085] in,

[0086] The step of extracting the time-frequency graph features and the differential bispectral graph features of each segmented signal includes:

[0087] Perform short-time Fourier transform on each segmented signal to obtain the time-frequency graph features corresponding to the segmented signal;

[0088] Each segmented signal is subjected to n-point differentiation to obtain a differential signal, and the differential signal is subjected to bispectral transformation to obtain differential bispectral features corresponding to the segmented signal.

[0089] In order to further improve the recognition rate of radiation source devices, the present invention explores the characteristics of different transformation domains, integrates signal feature information of different dimensions, constructs multi-dimensional features that can robustly represent the identity of the device, and extracts subtle differences in these features.

[0090] Perform N operations on the amplitude normalized communication signal data s Point segmentation processing is performed to obtain corresponding multiple segment signals; for example, N s There are 4096 sampling points.

[0091] N is performed on the communication signal data s Point segmentation processing, the number of segmented signals M is calculated as follows,

[0092]

[0093] Among them, the size of M depends on the length L of the communication signal data, the window length N and the frame shift R. Indicates rounding down, which means starting from the beginning of the communication signal data, moving R sampling points each time until the remaining signal length is less than a complete window N, and how many segmented signals there are in total.

[0094] Then, by extracting the local features of the time-frequency diagram of each segmented signal and the high-order nonlinear features in the differential bispectrum, the instantaneous frequency changes in the time-frequency domain and the nonlinear characteristics of the high-order spectral domain of the differential bispectrum are fully utilized to classify and identify individual radiation source devices.

[0095] (1) Perform short-time Fourier transform on each segmented signal to obtain the time-frequency graph features of each segmented signal.

[0096] The Short-Time Fourier Transform (STFT) is a widely used time-frequency analysis algorithm. It divides communication signal data into short time segments and then performs a Fourier transform on each segment, thereby describing the signal in both time and frequency dimensions.

[0097] Since STFT localizes the analysis of communication signal data by using a window function, it can analyze the frequency characteristics of communication signal data within a local time range. This characteristic enables STFT to reflect the spectrum changes of RF communication signal data within a specific time period, thereby helping to extract the local characteristics of the fingerprint of RF communication signal data. The discrete time STFT is as follows:

[0098]

[0099] Among them, S k,mIt is the result of STFT under time frame m and frequency index k, k = 1, 2, ···, N, N is the length of the window function ω[n], R is the frame shift (also called jump step or step size), index m is the time frame number, m = 1, 2, …, M, M is the length of the time frame, that is, how many segmented signals the communication signal data is divided into; s[n+mR] is the nth sampling point of the mth segmented signal of the discrete segmented signal to be analyzed; index n is the sampling point position within the window function, n = 0, 1, 2, ···, N-1; s[n] is the time domain representation of the communication signal data.

[0100] S k,m (m=1,2,…,M) is the time-frequency graph feature of each corresponding segmented signal.

[0101] (2) Perform differential bispectral transformation on each segmented signal to obtain the differential bispectral features corresponding to each segmented signal.

[0102] Differential processing is a method used to enhance signal characteristics. It can amplify and retain some key features in the signal, such as gain imbalance and carrier frequency offset. These features are crucial for RF signal fingerprint recognition and can enhance the identifiability of the features.

[0103] Differential processing also has a noise reduction effect, reducing noise interference in the signal to a certain extent, thereby improving the signal-to-noise ratio. This method can more clearly display features that are originally weak or easily submerged by noise, facilitating subsequent feature extraction and classification.

[0104] For example, the acquired communication signal data x(t) is as follows:

[0105]

[0106] Among them, x I (t) and x Q (t) are the transmitted I and Q signals, respectively, β I and β Q is the I / Q gain imbalance, α I and α Q is the gain imbalance, f t is the carrier frequency of the transmitter; j is the imaginary part.

[0107] The same receiver is used to receive the signal at the receiving end. The receiver carrier frequency is f r Due to technical deviations in the production process, the received signal will also have a frequency offset. The received communication signal data y(t) is as follows:

[0108]

[0109] Among them, y I(t), y Q (t) are the received I-channel and Q-channel signals respectively.

[0110] Perform n-point differential processing on the received communication signal data, and the differential signal d(t) is as follows:

[0111]

[0112] Among them, d I (t) and d Q (t) are the real and imaginary parts of the differential signal, y * (t+n) means that n points of offset are made to the received signal after conjugation, and n is the differential factor; x(t+n) means that n points of offset are made to the communication signal data, and x I (t+n) and x Q (t+n) is the n-point offset of the I and Q paths of the transmitted signal; θ = f r -f t is the carrier frequency offset. For ease of analysis, when n = 1, assuming that the deviation between the two points of the communication signal data is very small, x(t) ≈ x(t+1), then the above formula can be simplified and the real part is taken as follows:

[0113]

[0114] From formula (6), we can see that The term amplifies the fingerprint of I / Q gain imbalance and also includes the fingerprint of carrier frequency offset. The introduction of differential processing further enhances the discernibility of subtle changes in signal features, making these key features more prominent even in noisy environments.

[0115] As a high-order statistical analysis tool, bispectral transform can effectively reveal nonlinear characteristics and phase coupling in signals. It has unique advantages in detecting nonlinear characteristics related to I / Q gain imbalance and carrier frequency offset. Furthermore, bispectral transform has the ability to suppress Gaussian white noise, enabling it to maintain high feature extraction accuracy even in high-noise environments, thereby improving the accuracy and reliability of signal analysis. Therefore, it is widely used in high-order spectral feature extraction. Therefore, performing bispectral estimation on the differentially processed signal can obtain differential bispectral features.

[0116] This section explains how differential processing improves signal features. Later, it is combined with bispectral features. The differentially processed signal undergoes bispectral transformation. Because differential processing amplifies signal features, the bispectral image has higher resolution in terms of diagonal line spacing, texture, and edge features, and can be better identified using deep learning.

[0117] The bispectrum is defined as follows:

[0118] Assume that the k-th order high-order cumulant c kx (τ1,…,τ k-1 ) is absolutely summable, and the k-1 order discrete Fourier transform of the k-order cumulant of the signal is as follows:

[0119]

[0120] Among them, ω1,…,ω k-1 Corresponding to the frequency component after Fourier transform; S kx (ω1,…,ω k-1 ) represents the k-1-dimensional discrete Fourier transform of the k-order cumulant of the segmented signal x; τ1,…,τ k-1 Indicates the offset of the segmented signal at different time points; c kτ (τ1,…,τ k-1 ) represents the segmented signal at time points τ1,…,τ k-1 The joint statistical properties of .

[0121] When k=3, the third-order spectrum (bispectrum) B x (ω1,ω2) is,

[0122]

[0123] When generating a differential bispectral dataset, direct and indirect methods can be used to estimate the bispectrum. The direct method first estimates the Fourier transform coefficients after segmentation, and then performs a third-order correlation calculation on them to obtain the bispectrum estimate;

[0124] In the indirect rule, the third-order cumulant of the signal is first calculated, and then Fourier transform is performed on it to obtain the bispectral estimation value.

[0125] Since the indirect method can better handle the nonlinear characteristics and phase coupling phenomena in complex signals, and in addition, the indirect method can better resist the influence of noise through the calculation of third-order cumulative amounts, thereby improving the robustness and reliability of recognition, the present invention adopts the indirect method for bispectrum estimation.

[0126] The time-frequency features S corresponding to the segmented signal are generated by the above short-time Fourier transform and differential bispectral transform. k,m and differential bispectral features B x (ω1,ω2).

[0127] In order to effectively utilize these features for classification, clear input feature data is provided to the AttenSE-ResNet50 model, and MATLAB software is used to visualize the time-frequency graph features and differential bispectrum features to obtain the time-frequency graph and differential bispectrum of the signal.

[0128] (3) Each segmented signal is constructed into an image to obtain a three-channel spliced ​​fusion image corresponding to the time-frequency map and the differential bispectrum.

[0129] Splicing and fusing different domain features (time-frequency graph features and differential bispectral graph features) of the same segmented signal can significantly enhance the amount of feature information and provide complementary features, thereby improving the discrimination and generalization capabilities of the model. This method helps to reduce the overfitting phenomenon that may be caused by a single feature by integrating information from different features, thereby improving the robustness and overall performance of the model. Therefore, the present invention adopts a strategy of splicing and fusing different feature data in a height direction to maximize the complementarity between different domain features and enhance the performance of the model in a complex signal environment.

[0130] Specifically, the time-frequency graph features and differential spectrum graph features are processed by filling the spectrum internally to enhance the edge features and obtain a 224×224×1 three-channel image. Then, the two domain feature images of 224×224×1 are spliced ​​and fused in the height domain to obtain a 448×224×1 three-channel spliced ​​fusion image.

[0131] (4) Use the resize function to resize the three-channel spliced ​​fusion image to obtain the fusion image corresponding to each segmented signal.

[0132] To adapt to the subsequent AttenSE-ResNet50 model input feature requirements, for example, the resize function in Python is used to resize the fused image 448×224×1 to 224×224×1 to obtain a fused image suitable for input into the AttenSE-ResNet50 model. This effectively utilizes the complementarity between different features, thereby improving the model's recognition ability and adaptability, and ensuring stable performance in complex environments.

[0133] For example, the resize function reduces the first channel of the three-channel stitched fusion image to half of the original image; in specific applications, each channel of the three-channel stitched fusion image can be specifically adjusted according to specific needs.

[0134] The function of step S1 is to obtain the communication signal data of the radiation source to be identified, and convert it into a fusion image of multi-domain features through preprocessing (including cropping, normalization, segmentation, feature extraction, image construction and size adjustment operations), construct a standardized input to characterize the fingerprint of the radiation source device, and provide an adaptive time-frequency-bispectrum joint feature expression for the classification and recognition of the AttenSE-ResNet50 model.

[0135] Step S2 includes steps S21-S23.

[0136] Step S21: Input the multiple fused images into the trained AttenSE-ResNet50 model to perform classification and recognition to obtain a classification result probability vector for each fused image, and vote based on each classification result probability vector to obtain an identification result of the individual radiation source device.

[0137] Setting a confidence threshold for the probability distribution output by the AttenSE-ResNet50 model;

[0138] If the maximum probability value in the classification result probability vector is higher than the confidence threshold, it is determined that the input communication signal data belongs to the category corresponding to the trained communication signal data; otherwise, it is determined that it may belong to the communication signal data emitted by an untrained radiation source device individual, or the aging of the radiation source device and extreme electromagnetic interference.

[0139] Exemplarily, the confidence threshold is set to 0.7;

[0140] When the maximum probability value is <0.7, the possible reasons are as follows:

[0141] (1) The radiation source to be identified may be a new device that has not been trained;

[0142] (2) Aging of radiation source equipment (e.g., crystal oscillator frequency deviation exceeds ±100ppm);

[0143] (3) Extreme electromagnetic interference (SNR < -10dB).

[0144] The communication signal data emitted by a radiation source to be identified is divided into multiple segmented signals, and each segmented signal is preprocessed into a corresponding fused image. The multiple fused images corresponding to the communication signal data are classified and identified using the AttenSE-ResNet50 model to obtain the corresponding probability vector for each classification result.

[0145] Determine whether the maximum probability value in each classification result probability vector is less than the confidence threshold;

[0146] If not, that is, there is at least one classification result probability vector with a maximum probability value greater than or equal to the confidence threshold, then the voting phase begins;

[0147] If so, all classification results are considered unreliable. It is determined that the input communication signal data may come from an untrained radiation source device, or the radiation source may be aged or subject to extreme electromagnetic interference, resulting in misjudgment and identification, and the process ends.

[0148] The voting process is as follows:

[0149] Collect candidate categories: For each classification result probability vector, if its maximum probability value is greater than or equal to the confidence threshold, the corresponding category is taken as a candidate category.

[0150] Count votes: Count all candidate categories and calculate the number of votes each category receives.

[0151] Final identification results: Based on the voting results, the category with the most votes is selected as the ID of the individual radiation source device to be identified. If multiple categories have the same number of votes and are the most voted, the category with the most stable performance in the validation set is selected or marked as requiring further analysis.

[0152] For example, assuming that the communication signal data to be identified has 5 segmented signals, which correspond to 5 fused images after preprocessing, and the confidence threshold is set to 0.7, the corresponding classification result probability vector is as follows:

[0153] Fusion image 1: [0.8, 0.1, 0.05, 0.05];

[0154] Fusion image 2: [0.75, 0.15, 0.05, 0.05];

[0155] Fusion image 3: [0.6, 0.2, 0.1, 0.1];

[0156] Fusion image 4: [0.7, 0.15, 0.1, 0.05];

[0157] Fused image 5: [0.65, 0.2, 0.1, 0.05].

[0158] Judgment: The maximum probability values ​​for images 1, 2, and 4 are 0.8, 0.75, and 0.7, respectively, all greater than 0.7; the maximum probability values ​​for images 3 and 5 are 0.6 and 0.65, respectively, both less than 0.7. Therefore, not all maximum probability values ​​are less than the threshold.

[0159] Collect candidate categories: The candidate categories of images 1, 2, and 4 are category 1, category 1, and category 1 respectively.

[0160] Voting statistics: Category 1 received 3 votes.

[0161] The final result is determined as follows: Category 1 has the highest number of votes, so the ID of the individual radiation source device to be identified is determined to be Category 1.

[0162] Step S22: Build the AttenSE-ResNet50 model, as shown in Figure 5 shown.

[0163] The AttenSE-ResNet50 model includes a convolutional initial processing module, the first, second, third and fourth depth residual units, a classification pre-processing module and a fully connected classification layer in sequence; wherein,

[0164] The convolution initial processing module includes a convolution layer, a batch normalization and ReLU6 nonlinear activation layer, and a maximum pooling layer in sequence, which is used to perform a preliminary convolution operation on the input fusion image, extract the initial features, perform batch normalization and ReLU6 nonlinear transformation, and obtain the basic feature map after maximum pooling downsampling;

[0165] The first depth residual unit includes first, second, and third residual modules with the same structure, which are cascaded in sequence; the second depth residual unit includes fourth, fifth, sixth, and seventh residual modules with the same structure, which are cascaded in sequence; the third depth residual unit includes eighth to thirteenth residual modules with the same structure, which are cascaded in sequence; the fourth depth residual unit includes fourteenth to sixteenth residual modules with the same structure, which are cascaded in sequence; the first to fourth depth residual units are used to obtain the first fingerprint feature of the corresponding radiation source device individual based on the basic feature map;

[0166] The classification pre-processing module includes a global average pooling layer, a flattening layer, and a dropout layer in sequence, and is used to obtain a second fingerprint feature of the corresponding radiation source device based on the first fingerprint feature of the radiation source device;

[0167] The fully connected classification layer is used to output a classification result probability vector based on the second fingerprint feature of the individual radiation source device; and perform a voting decision based on the classification result probability vector to obtain the ID identification of the individual radiation source device that emits the communication signal data.

[0168] The input of the AttenSE-ResNet50 model is a 224×224×1 fused image;

[0169] The convolution initial processing module is used for preliminary feature extraction and downsampling to reduce the amount of computation. The input is a 224×224×1 fused image.

[0170] After extracting features from the input image through a 7×7 convolutional layer with 64 channels and a stride of 2, BN+ReLU6 is used for normalization and nonlinear activation. Finally, a 3×3 max pooling layer with a stride of 2 is used for further downsampling and enhancing salient features, resulting in an input image of 56×56×64 dimensions suitable for the first residual unit. They are expressed as follows:

[0171] Convolutional layer Conv7×7, 64, s=2; indicates that a large convolution kernel with a size of 7×7 captures the global features of the fused image, the number of output channels is 64, the stride is 2, the spatial size is compressed to retain the identification row information, the output size is halved for downsampling, the output feature of the convolution layer is 112×112×64, followed by batch normalization and ReLU6 nonlinear activation layer and maximum pooling Maxpool3×3, s=2; Maxpool3×3, s=2 indicates that the pooling window size is 3×3, the stride is 2, the output size is halved and reverse downsampling is performed, and there is no padding by default; the output size is a basic feature map of 56×56×64, which is adapted to the input of the first residual unit.

[0172] In order to enable the AttenSE-ResNet50 model to focus on extracting features at different levels and enhancing feature representation capabilities at each stage, the present invention adopts the first, second, third and fourth residual units of the four levels with a residual distribution of (3, 4, 6, 3), all of which are deep residual unit structures, and uses the improved first to sixteenth residual modules to perform layer-by-layer feature extraction on the fingerprint image at each level. Among them, the improved residual module has a jump mapping connection, and the output features are extracted from the input features through a three-layer structure of dimensionality reduction-convolution-dimensionality increase, and after combining the dual attention mechanism and SE module to enhance the feature representation, they are added to the original input features through a shortcut connection to supplement the features lost after the input passes through multiple convolution layers, solving the gradient disappearance problem.

[0173] The first to fourth depth residual units are used for extracting deep fingerprint features of communication signal data.

[0174] The first depth residual unit outputs a 56×56×256 dimension of depth feature, which serves as the input of the second depth residual unit.

[0175] The second depth residual unit outputs a depth feature of 28×28×512 dimensions, which serves as the input of the third depth residual unit.

[0176] The output of the third depth residual unit is a depth feature of 14×14×1024 dimensions, which serves as the input of the fourth depth residual unit.

[0177] The fourth deep residual unit outputs a deep feature of 7×7×2048 dimensions, which serves as the input of the pre-classification processing module.

[0178] The first depth residual unit includes the first, second and third residual modules with the same structure cascaded in sequence; the second depth residual unit includes the fourth, fifth, sixth and seventh residual modules with the same structure cascaded in sequence; the third depth residual unit includes the eighth to thirteenth residual modules with the same structure cascaded in sequence;

[0179] The first depth residual unit includes the first, second and third residual modules with the same structure cascaded in sequence; the first, second and third depth residual units have similar structures. The first to sixteenth residual modules have similar network structures and functions.

[0180] The first to sixteenth residual modules each include a first, second and third convolutional blocks; the first and second convolutional blocks are followed by batch normalization and a ReLU6 nonlinear activation layer; the third convolutional block is followed by a batch normalization layer, a two-layer attention module, an SE attention module, a skip mapping connection between the residual module input and the SE module output, and a ReLU6 nonlinear activation; each residual module is used to perform channel transformation and size transformation on the input image.

[0181] The following is an example of the first residual module:

[0182] (1) The first convolution block has a kernel size of 1×1 and 64 channels;

[0183] (2) The first layer is normalized and ReLU6 nonlinear activation is performed to normalize and nonlinearly activate the feature map after the convolution operation;

[0184] (3) The second convolution block has a kernel size of 3 × 3 and 64 channels;

[0185] (4) The second batch of normalization and Relu6 nonlinear activation are used to normalize and nonlinearly activate the feature maps after the convolution operation;

[0186] (5) The third convolution block has a convolution kernel size of 1×1 and a channel number of 256;

[0187] (6) The third batch is normalized to normalize the feature map after the convolution operation.

[0188] (7) A two-layer attention module, including sequentially connected channel attention and spatial attention submodules; outputting an enhanced feature expression map;

[0189] (8) SE attention module, which obtains channel attention weights based on the enhanced feature expression graph;

[0190] (9) A skip mapping connection from the first residual block input to the SE attention module output, adding the residual block input and the SE attention module output;

[0191] (10) Relu6 nonlinear activation.

[0192] The network structures and functions of the second to sixteenth residual modules are similar to those of the first residual module and will not be repeated here.

[0193] The first to sixteenth are improved residual modules, using two 1×1 convolution kernels and one 3×3 convolution kernel. The two 1×1 convolution kernels reduce the number of output channels, first reducing the dimensionality and then increasing it, effectively reducing the computational complexity of the residual network and the number of network parameters in the basic residual unit, thereby improving the operational efficiency of the deep residual network. The 3×3 convolution kernel reduces the information size and utilizes a smaller receptive field, avoiding the loss of detail in the signal feature image caused by a larger receptive field. While maintaining efficient computation, it effectively captures local spatial information, facilitating the network to accurately extract more precise signal features and improving the accuracy of feature extraction.

[0194] To more effectively extract and represent signal features, while simultaneously addressing the vanishing gradient problem and improving the training and classification performance of deep neural networks, this paper proposes combining a dual attention mechanism module, an SE module, and a residual block as the basic network architecture for signal feature extraction. To better adapt feature extraction to the residual network, the concatenated and fused three-channel input image requires channel transformation and size reduction.

[0195] The attention mechanism can focus on important information and ignore minor information, thereby effectively extracting information. The present invention adopts a dual attention mechanism module including a channel attention submodule and a spatial attention submodule to extract and utilize important feature information, such as Figure 7 As shown, the double-layer attention module network structure.

[0196] The dual-layer attention module includes a channel attention submodule and a spatial attention submodule connected in sequence;

[0197] The channel attention submodule collects comprehensive features of the input feature map through the parallel maximum pooling layer and average pooling layer to obtain two corresponding feature maps, and uses the shared perception layer to assign weight values ​​to the importance of the two input feature maps to obtain two feature vectors; the two feature vectors are added and then activated by Sigmoid to generate a channel attention weight feature; then the channel attention weight feature is multiplied by the input feature map to obtain a channel attention weight vector map;

[0198] The spatial attention submodule performs maximum pooling and average pooling on the channel attention weight vector map, connects the outputs of the two, and performs convolution operation and Sigmoid activation to obtain a spatial attention weight feature vector map;

[0199] Multiply the spatial attention weight feature vector map with the channel attention weight vector map to obtain the enhanced feature expression map output by the double-layer attention module as the input of the SE attention module.

[0200] The channel attention submodule primarily compresses the spatial content of the feature map, comprehensively collects input feature information through maximum pooling and average pooling, assigns weights to the importance of feature information through a shared multilayer perceptron, and concatenates these weights in sequence to produce two feature vectors. The two vectors are added together and activated with a sigmoid activation function to obtain the weighted features of the channel attention mechanism. Multiplying these by the input features yields the new feature information of the channel attention mechanism.

[0201] The spatial attention submodule performs average pooling and maximum pooling on the feature maps of the channel axis and connects them, then performs convolution operation and Sigmoid activation to obtain the weighted feature vector and multiply it with the result of the channel attention mechanism to finally obtain the feature information extracted by the spatial attention mechanism.

[0202] As a channel attention module, the SE attention module can extract and strengthen the channel features of the input multi-channel feature map without changing the size of the input feature map and the network structure, thereby further improving the overall performance of the network. The SE attention module first performs global average pooling on each channel to compress the spatial information (Squeeze); then the vector is sent to two fully connected layers, and channel weights (Excitation) in the range of 0-1 are generated through ReLU and Sigmoid; finally, these weights are used to scale the original input feature map element by element according to the channel, thereby amplifying the information-rich channels and suppressing redundant channels. The entire process hardly increases the amount of computation, but allows the network to adaptively highlight key features and improve representation capabilities. Figure 8 As shown, the network structure of SE attention module.

[0203] The SE attention module performs global pooling on each channel of the enhanced feature expression graph, and then performs the first full connection, ReLU activation, second full connection and Sigmoid activation in sequence to generate channel weights in the [0,1] interval;

[0204] The SE attention module includes a global pooling layer, a first fully connected layer, a ReLU activation layer, a second fully connected layer and a Sigmoid activation layer;

[0205] The global average pooling layer is used to compress the spatial information of the input enhanced feature expression map to obtain a channel descriptor;

[0206] The first fully connected layer and the second fully connected layer are used to perform nonlinear transformation on the channel descriptor to generate channel weights;

[0207] The ReLU activation layer is used to introduce nonlinearity; the Sigmoid activation layer is used to map the channel weight to the [0,1] interval to obtain a channel attention weight map;

[0208] The SE attention module performs channel weighting on the enhanced feature expression map, enhancing key feature channels and suppressing redundant feature channels.

[0209] Step S23: Train the AttenSE-ResNet50 model until the loss function converges or the maximum number of iterations is reached, to obtain a trained AttenSE-ResNet50 model.

[0210] The AttenSE-ResNet50 model is trained based on the following sample training set:

[0211] Signal data with BPSK modulation type under different signal-to-noise ratios are collected and preprocessed to obtain signal sample data, which are then combined with corresponding sample labels to form a sample training set; wherein the sample label is the ID identifier of the individual radiation source device.

[0212] Through real receivers and simulations, communication signal sample data with a modulation type of BPSK (Binary Phase Shift Keying, which transmits binary data by changing the phase of the carrier) under different signal-to-noise ratios are obtained.

[0213] Binary phase shift keying (BPSK) uses the phase change of the carrier to transmit information. Its absolute phase can take two possible values, representing binary information 0 and 1, while its carrier amplitude and frequency do not change.

[0214] The expression of BPSK signal is:

[0215]

[0216] Among them, a n is the baseband information of the nth symbol, with a value of ±1, where +1 and -1 correspond to binary information 0 and 1, T is the symbol period, g(t) is the baseband pulse with a duration of T, and f c is the center carrier frequency of the signal, Indicates the initial phase of the signal, which is 0 or π, and n0(t) is the signal that obeys N(0,σ 2 ) of additive Gaussian white noise; N(0,σ 2 ) has a mean of 0 and a variance of σ 2 of random noise.

[0217] The sample training set is divided into a training set, a test set and a validation set according to a preset ratio;

[0218] Set the training batch, training round, learning rate, minimum learning rate, learning rate annealing period and weight decay coefficient;

[0219] Training the AttenSE-ResNet50 model using the training set; using a cross-entropy loss function during training, and dynamically adjusting the learning rate using a cosine annealing learning rate scheduler;

[0220] Continuously adjust and update model parameters through back propagation and gradient descent optimization algorithms;

[0221] Until the cross entropy loss function converges or the training round is completed, the AttenSE-ResNet50 model parameters are saved to obtain a trained AttenSE-ResNet50 model.

[0222] Exemplarily, the present invention uses the same type of communication signal data transmitted by 4 transmitters of the same model, with 600 communication signal data from each transmitter, totaling 2400 sample data, which are divided according to the ratio of 8:1:1, of which 1920 are training sets, 240 are test sets, and 240 are verification sets.

[0223] In the improved structure of the present invention, the dual attention module ensures that the AttenSE-ResNet50 model can capture effective feature information in different dimensions by extracting channel and spatial attention at the same time, and adaptively adjusts the weights of different regions and channels according to the input. Subsequently, the SE attention module further focuses on the relationship between different channels, and on the basis of retaining the original input features to the greatest extent, learns the dependencies between different channels, and focuses on the key points in the multi-channel data, thereby enhancing the performance of the network. Introducing the SE attention module after applying the dual attention mechanism can flexibly adjust the weights of different channel features, enhance the representation ability of globally important feature channels, help improve the network's noise resistance and ability to capture detailed features, and ultimately improve the overall classification effect.

[0224] In addition, in order to better preserve the original input features, the present invention replaces the activation function from ReLU to ReLU6. Compared with the ReLU function, the ReLU6 function has a smaller computational load. In addition, ReLU6 has a better buffering effect than the ReLU activation function and can prevent gradient disappearance. Therefore, the ReLU6 activation function used forms a deep network in the basic residual unit structure at all levels, which can more effectively preserve the original features of the image while avoiding the gradient disappearance problem and improving the stability and performance of the model. These improvements enable the updated residual block to significantly improve feature extraction and representation capabilities while maintaining computational efficiency, thereby improving overall classification performance.

[0225] The function of step S2 is to use the trained AttenSE-ResNet50 model to pre-process the deep fingerprint features (first and second fingerprint features) of the multiple fused images, and to classify and identify the individual radiation source device based on the second fingerprint feature through the fully connected classification layer. By analyzing the probability vector of the classification result of each image and making voting decisions based on these probability vectors, the ID of the individual radiation source device that emits communication signal data can be accurately identified.

[0226] Example 1:

[0227] 1. The experimental signal is the same signal transmitted by four transmitters of the same model, which is actually collected and received by the same receiver. The signal modulation method is BPSK modulation, the signal sampling rate is 96kHz, and the carrier frequency is 24kHz.

[0228] 2. In order to obtain the time-frequency domain characteristics and differential bispectral high-order spectral domain fingerprint characteristics of the device, the present invention first preprocesses the real signal data of the communication signals received from the four transmitters, cuts off the signal-free part in the signal, normalizes the amplitude, and performs 4096-point segmentation processing;

[0229] Then, the short-time Fourier transform is applied to the segmented signal data, the short-time Fourier window length is selected as 512, the step size is 256, and the FFT length is 512. Figure 2 This is the time-frequency diagram of the measured signal collected by the receiver in a 5dB signal-to-noise ratio environment;

[0230] At the same time, a two-point differential processing is performed on the segmented signal data to amplify the I / Q gain imbalance and carrier frequency offset in the signal, which reflect the nonlinear characteristics of the signal, to obtain the time-frequency diagram characteristics of the data. Then, the third-order cumulants of the 4096-point signal of each segmented signal are calculated. Each segmented signal is divided into 16 segments with 256 points per segment, and its third-order cumulants are calculated. The third-order cumulants of all segments are averaged and then a two-dimensional Fourier transform is performed to obtain the signal differential bispectral characteristics.

[0231] Figure 3 The following is a comparison of the differential bispectral contour maps of the measured signals of transmitter device 1 and transmitter device 2 collected by two receiver devices in a 5dB signal-to-noise ratio environment. The horizontal and vertical axes represent the values ​​of the two frequency components f1 and f2. Each point (f1, f2) in the differential bispectral map reflects the degree of nonlinear coupling between the corresponding frequency components. The depth of the color represents the amplitude of the differential bispectrum. The brighter the area, the stronger the nonlinear coupling. Figure 3 It can be seen that the characteristics of differential dual spectrum amplification I / Q gain imbalance and carrier frequency offset, which reflect the nonlinear characteristics of the signal, can provide key feature basis for device identification through changes in diagonal line spacing, texture, edge distribution and local patterns.

[0232] 3. The obtained time-frequency domain feature maps and differential bispectral features, these fingerprint feature sequences, are processed using the spectrum internal filling method to enhance edge features, resulting in a 224×224×1 three-channel image. The two 224×224×1 images with the same label are then spliced ​​and fused in the height domain to obtain a 448×224×1 three-channel spliced ​​fused image. Next, to adapt to the subsequent network model input requirements, the resize function in Python is used to resize the three-channel spliced ​​fused image to 224×224×1, thus obtaining the final fused image suitable for the input model. Figure 3 It is the image after the time-frequency map and the differential bispectrum map are spliced ​​and fused.

[0233] Each transmitter device uses 600 images as the training set, with a total of 4 transmitter devices. The images are divided into 1920 images for the training set, 240 images for the test set, and 240 images for the validation set in an 8:1:1 ratio.

[0234] 4. Figure 5 This is the network structure of the AttenSE-ResNet50 model in the present invention. First, the 224×224×1 fusion image obtained by splicing and fusion in the image construction process is input into the AttenSE-ResNet50 model; in the model, after the fusion image passes through the 7×7 dimensional input convolution layer, the instance normalization layer (Batch Normalization, BN) and the ReLu6 layer channel change and size change, the signal image is obtained as 112×112×1, and then the input image (56×56×64) is obtained through the maximum pooling layer; then the signal features are extracted through a total of 16 improved residual unit networks distributed as (3, 4, 6, 3) to obtain a 7×7×2048 feature map; finally, the dimensionality reduction processing of the average pooling Flatten layer is performed to obtain the final 1×1×2048 feature data, and the classification of the radiation sources of different communication equipment is completed through the fully connected layer classifier. Finally, the classification performance of the model is evaluated by analyzing the test set and its recognition effect is verified. Figure 6 The improved residual module network structure used for network feature extraction.

[0235] Example 2:

[0236] 1. Dataset: The communication sample signals used in the experiment were actually collected from four transmitters of the same model and received by the same receiver. The signal modulation method was BPSK modulation, with a sampling rate of 96kHz and a carrier frequency of 24kHz. To compare the recognition performance of the algorithm under different signal-to-noise ratios, the noise environment with different signal-to-noise ratios was simulated by adding additive white Gaussian noise using the simulation software Matlab 2019a. The signal-to-noise ratio range of the dataset was {-10dB, -7dB, -5dB, -2dB, 0dB, 5dB, 8dB, 10dB, 15dB, 20dB}, and the image resolution was 224×224×1, with three channels. 600 samples were extracted from each device, for a total of 2400 samples at each signal-to-noise ratio. The dataset was divided into a training set with a ratio of 8:1:1, with 1920 samples, a test set with 240 samples, and a validation set with 240 samples. The image samples were labeled with their corresponding types. The dataset was preprocessed using spectral infill to increase the network's sensitivity to spectral edge features.

[0237] To further validate the algorithm's universality, we used a publicly available Bluetooth signal dataset (https: / / zenodo.org / records / 3876140) collected at a 5GHz sampling rate for testing. By testing data from various sources and conditions, we ensured that the proposed algorithm not only performs well in specific environments but also has broad adaptability and stability, thus verifying its effectiveness and reliability in practical applications.

[0238] 2. Experimental environment and network configuration

[0239] In the experiment, the AttenSE-ResNet50 model was run on a server with an NVIDIA TITAN RTX 3090 GPU. The deep learning framework was Pytorch, the training batch size was 16, and the maximum training epochs was 50. The loss function used the cross-entropy loss function (CrossEntropyLoss), and the loss function was optimized by the adaptive momentum estimation (Adam) optimizer, where the learning rate was set to 1e-4 and the weight decay coefficient was 5e-4. At the same time, the cosine annealing learning rate scheduler (CosineAnnealingLR) was used to dynamically adjust the learning rate to improve the efficiency and effect of model training. The learning rate annealing period (T_max) was set to 10, and the minimum learning rate (eta_min) was 1e -6 .

[0240] 3. Performance simulation

[0241] In addition to the multi-input feature fusion device recognition algorithm based on time-frequency diagram and differential bispectrum proposed in this invention, the experimental algorithm also conducts comparative analysis of recognition performance under different signal-to-noise ratios with the nonlinear feature extraction algorithm based on short-time Fourier transform combined with convolutional neural network and bispectral transform.

[0242] 3-1. Classification and recognition performance of different algorithms

[0243] To verify the effectiveness of the proposed multi-input feature fusion recognition method based on time-frequency diagrams and differential bispectra, a comparative experiment was conducted using a method that trained three signal features separately and a method that trained multiple features simultaneously using the proposed method. The convolutional neural network model used for the single-feature approach was a modified ResNet50, while the multi-feature approach used the proposed multi-input feature fusion network model. In the experiment, the two datasets were processed identically to test their classification and recognition performance. Figure 9a The recognition performance of the four devices under different signal-to-noise ratios of the receiver signal is shown in Figure 2. Figure 9b The recognition performance of four devices under different signal-to-noise ratios of mobile phone Bluetooth signals.

[0244] observe Figure 9a and 9b It can be seen that the multi-domain feature fusion algorithm proposed in the present invention has the highest recognition accuracy, and the recognition rate after fusion is higher than the recognition rate of using the time-frequency diagram, dual spectrum, and differential dual spectrum alone. In addition, the recognition rate of using the differential dual spectrum alone is also higher than that of using the dual spectrum alone. This is because the time-frequency diagram can reflect the time-frequency characteristics of the transmitted signal, the dual spectrum can reflect the amplitude and phase characteristics of the signal transmitted by the analog signal source, and the differential dual spectrum improved by the dual spectrum enhances the signal I / Q gain imbalance and carrier frequency offset characteristics. The single feature has limited characterization ability for the subtle characteristic differences between different radiation sources. The feature fusion of the time-frequency diagram and the better differential dual spectrum can more comprehensively reflect the subtle characteristic differences between different individual devices, improve the recognition accuracy of different individual devices, and thus prove the effectiveness of the algorithm proposed in this article.

[0245] 3-2. Classification and recognition performance of different algorithms

[0246] In order to verify the performance of the improved multi-domain feature fusion network of the present invention, an ablation experiment was conducted. The time-frequency diagram and differential bispectral diagram of the receiver signal were fused and input into the multi-domain feature fusion AttenSE-ResNet50 model of the present invention, the multi-domain fusion network SE+ResNet50 with SE self-attention module, and the ResNet50 multi-domain fusion network after removing SE module and self-attention module for comparative experiments. Figure 10 is the recognition performance of each model under different signal-to-noise ratios.

[0247] from Figure 10 As can be seen, the improved and optimized AttenSE-ResNet50 multi-domain feature fusion algorithm containing the SE module and the multi-attention module of the present invention has significantly improved the recognition accuracy compared to the original model Resnet50 model and the SE+ResNet50 model with only the SE attention module added. When the signal-to-noise ratio is -5dB, the recognition accuracy of the AttenSE-ResNet50 model is 5.37% higher than that of the SE+ResNet50 model with only the SE attention mechanism module added; the recognition accuracy is increased by 9.12% compared to the original model ResNet50. This result shows that the introduction of the SE attention module and the dual-layer attention module effectively enhances the robustness and recognition ability of the AttenSE-ResNet50 model in low signal-to-noise ratio environments, significantly improving the performance of radio frequency identification.

[0248] In summary, the multi-feature radiation source individual identification method according to the embodiment of the present invention has the following beneficial effects:

[0249] 1. This invention characterizes the inherent fingerprint characteristics of devices from different perspectives by fusing the local features of the time-frequency graph and the high-order nonlinear features in the differential bispectrum. This solves the problem that single feature information is insufficient to fully characterize the subtle feature differences of devices. Compared with existing methods that rely solely on single transform domain information, it can more accurately identify different radiation source devices. It also avoids the problem that single feature information of RF radiation sources in the UHF band is insufficient to fully characterize the subtle feature differences of devices.

[0250] 2. This invention introduces a dual attention mechanism module and an SE module, which enables the network to selectively enhance the feature responses of key channels and spatial positions, suppress irrelevant features and noise interference, and accurately extract effective features even in low signal-to-noise ratio environments. This effectively solves the technical problem of the difficulty in distinguishing RF fingerprint features in low signal-to-noise ratio environments in existing technologies and improves recognition accuracy.

[0251] 3. The AttenSE-ResNet50 network model structure constructed by the present invention solves the gradient vanishing problem in the deep neural network training process through the four-level first, second, third and fourth depth residual units distributed as (3, 4, 6, 3) and the skip connection design, thereby improving the training efficiency and classification performance of the model. Compared with the traditional complex and difficult-to-train model, the training method uses less sample data.

[0252] 4. This invention uses deep learning methods to automatically learn the patterns in radio frequency signals and abstract higher-dimensional and more representative features. It does not need to rely heavily on manual extraction and signal processing experience like traditional feature engineering methods, thus ensuring the effectiveness and efficiency of feature extraction.

[0253] 5. In a 0dB signal-to-noise ratio environment, the individual radiation source identification method of the present invention significantly improves the classification accuracy compared to the recognition algorithms using bispectrum, time-frequency diagram or differential bispectrum alone.

[0254] Those skilled in the art will appreciate that all or part of the process steps of the above-described embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, such as a magnetic disk, an optical disk, a read-only memory, or a random access memory.

[0255] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed by the present invention should be covered by the scope of protection of the present invention.

Claims

1. A multi-feature radiation source individual identification method, characterized in that: include: Acquire communication signal data emitted by the radiation source to be identified, and obtain corresponding multiple fused images after preprocessing; The multiple fused images are input into the trained AttenSE-ResNet50 model for classification and recognition to obtain a classification result probability vector for each fused image, and voting is performed based on each classification result probability vector to obtain the identification result of the individual radiation source device.

2. The multi-feature radiation source individual identification method according to claim 1, characterized in that: The acquiring of communication signal data and preprocessing to obtain corresponding multiple fused images includes: After trimming the signal-free portion of the signal sample data, normalizing the amplitude and performing segmentation to obtain corresponding multiple segmented signals; Extract the time-frequency graph features and differential bispectral graph features of each segmented signal; Visualizing the time-frequency graph features and the differential bispectrum graph features to obtain corresponding time-frequency graphs and differential bispectrum graphs; Filling the time-frequency graph and the differential bispectrum graph internally to obtain a corresponding three-channel image; The three-channel images corresponding to the time-frequency diagram and the differential bispectrum of each segmented signal are spliced ​​and fused to obtain a three-channel spliced ​​fused image corresponding to each segmented signal; The size of the first channel of each of the three-channel stitched fusion images is adjusted using a Resize function to obtain multiple fusion images corresponding to the multiple segmented signals.

3. The multi-feature radiation source individual identification method according to claim 2, characterized in that: The AttenSE-ResNet50 model includes a convolutional initial processing module, the first, second, third and fourth depth residual units, a classification pre-processing module and a fully connected classification layer in sequence; wherein, The convolution initial processing module includes a convolution layer, a batch normalization and ReLU6 nonlinear activation layer, and a maximum pooling layer in sequence, which is used to perform a preliminary convolution operation on the input fusion image, extract the initial features, perform batch normalization and ReLU6 nonlinear transformation, and obtain the basic feature map after maximum pooling downsampling; The first depth residual unit includes first, second, and third residual modules with the same structure, which are cascaded in sequence; the second depth residual unit includes fourth, fifth, sixth, and seventh residual modules with the same structure, which are cascaded in sequence; the third depth residual unit includes eighth to thirteenth residual modules with the same structure, which are cascaded in sequence; the fourth depth residual unit includes fourteenth to sixteenth residual modules with the same structure, which are cascaded in sequence; the first to fourth depth residual units are used to obtain the first fingerprint feature of the corresponding radiation source device individual based on the basic feature map; The classification pre-processing module includes a global average pooling layer, a flattening layer, and a dropout layer in sequence, and is used to obtain a second fingerprint feature of the corresponding radiation source device based on the first fingerprint feature of the radiation source device; The fully connected classification layer is used to output a classification result probability vector based on the second fingerprint feature of the individual radiation source device; and perform a voting decision based on the classification result probability vector to obtain the ID identification of the individual radiation source device that emits the communication signal data.

4. The multi-feature radiation source individual identification method according to claim 3, characterized in that: The first to sixteenth residual modules each include a first, second and third convolutional blocks; the first and second convolutional blocks are followed by batch normalization and a ReLU6 nonlinear activation layer; the third convolutional block is followed by a batch normalization layer, a two-layer attention module, an SE attention module, a skip mapping connection between the residual module input and the SE module output, and a ReLU6 nonlinear activation; each residual module is used to perform channel transformation and size transformation on the input image.

5. The multi-feature radiation source individual identification method according to claim 4, characterized in that: The dual-layer attention module includes a channel attention submodule and a spatial attention submodule connected in sequence; The channel attention submodule collects comprehensive features of the input feature map through the parallel maximum pooling layer and average pooling layer to obtain two corresponding feature maps, and uses the shared perception layer to assign weight values ​​to the importance of the two input feature maps to obtain two feature vectors; the two feature vectors are added and then activated by Sigmoid to generate the channel attention weight feature; Then the channel attention weight feature is multiplied by the input feature map to obtain a channel attention weight vector map; The spatial attention submodule performs maximum pooling and average pooling on the channel attention weight vector map, connects the outputs of the two, and performs convolution operation and Sigmoid activation to obtain a spatial attention weight feature vector map; Multiply the spatial attention weight feature vector map with the channel attention weight vector map to obtain the enhanced feature expression map output by the double-layer attention module as the input of the SE attention module.

6. The multi-feature radiation source individual identification method according to claim 5, characterized in that: The SE attention module performs global pooling on each channel of the enhanced feature expression graph, and then performs the first full connection, ReLU activation, second full connection and Sigmoid activation in sequence to generate channel weights in the [0,1] interval; The SE attention module includes a global pooling layer, a first fully connected layer, a ReLU activation layer, a second fully connected layer and a Sigmoid activation layer; The global average pooling layer is used to compress the spatial information of the input enhanced feature expression map to obtain a channel descriptor; The first fully connected layer and the second fully connected layer are used to perform nonlinear transformation on the channel descriptor to generate channel weights; The ReLU activation layer is used to introduce nonlinearity; The Sigmoid activation layer is used to map the channel weight to the [0,1] interval to obtain a channel attention weight map; The SE attention module performs channel weighting on the enhanced feature expression map, enhancing key feature channels and suppressing redundant feature channels.

7. The multi-feature radiation source individual identification method according to claim 2, characterized in that: The step of extracting the time-frequency graph features and the differential bispectral graph features of each segmented signal includes: Perform short-time Fourier transform on each segmented signal to obtain the time-frequency graph features corresponding to the segmented signal; Each segmented signal is subjected to n-point differentiation to obtain a differential signal, and the differential signal is subjected to bispectral transformation to obtain differential bispectral features corresponding to the segmented signal.

8. The multi-feature radiation source individual identification method according to claim 1, characterized in that: The AttenSE-ResNet50 model is trained based on the following sample training set: Signal data with BPSK modulation type under different signal-to-noise ratios are collected and preprocessed to obtain signal sample data, which are then combined with corresponding sample labels to form a sample training set; wherein the sample label is the ID identifier of the individual radiation source device.

9. The multi-feature radiation source individual identification method according to claim 8, characterized in that: The sample training set is divided into a training set, a test set and a validation set according to a preset ratio; Set the training batch, training round, learning rate, minimum learning rate, learning rate annealing period and weight decay coefficient; Training the AttenSE-ResNet50 model using the training set; using a cross-entropy loss function during training, and dynamically adjusting the learning rate using a cosine annealing learning rate scheduler; Continuously adjust and update model parameters through back propagation and gradient descent optimization algorithms; Until the cross entropy loss function converges or the training round is completed, the AttenSE-ResNet50 model parameters are saved to obtain a trained AttenSE-ResNet50 model.

10. The multi-feature radiation source individual identification method according to any one of claims 1 to 9, characterized in that: Setting a confidence threshold for the probability distribution output by the AttenSE-ResNet50 model; If the maximum probability value in the classification result probability vector is higher than the confidence threshold, it is determined that the input communication signal data belongs to the category corresponding to the trained communication signal data; Otherwise, it is determined that the data may be communication signal data emitted by an untrained radiation source device, or the radiation source device may be aged and subjected to extreme electromagnetic interference.

Images