A communication radiation source identification method based on time domain multi-class feature fusion

Abstract

The present application relates to the field of radiation source individual identification, and particularly relates to a communication radiation source identification method based on time domain multi-class feature fusion, the present application uses different methods to convert the captured electromagnetic signals into four different radiation source fingerprint feature representation modes in the time domain, makes different types of neural networks correspond to the radiation source fingerprint feature representation, extracts different radiation source fingerprint features, performs arrangement processing to form four feature matrices with the same data format, stacks the feature matrices, uses a 3D convolution network to perform feature fusion and extraction on the stacked feature matrices, obtains the final fusion features, and uses a full connection layer to realize the identification of the radiation source individuals according to the extracted features. The present application uses a signal time domain feature extraction method, simplifies the signal processing calculation, improves the efficiency of the signal processing stage, and can obtain more comprehensive radiation source fingerprint features.

Description

Technical Field

[0001] This invention relates to the field of individual radiation source identification, and specifically to a communication radiation source identification method based on temporal multi-class feature fusion. Background Technology

[0002] Radiation source identification technology, also known as Specific Radiation Source Identification (SEI), is a key technology in communications reconnaissance and electronic warfare, and a cutting-edge direction in electronic warfare both domestically and internationally. Due to the non-ideal characteristics of semiconductor devices and the various variations that arise during the manufacturing process of radiation source equipment, each device possesses an inherent attribute—a radiation source fingerprint. This inherent attribute cannot be eliminated or forged and is unique. Radiation source identification technology extracts features from captured electromagnetic signals to obtain the radiation source fingerprint within the electromagnetic signals, thereby identifying the individual radiation source equipment.

[0003] Radiation sources are identified by calculating and measuring the captured electromagnetic signals to obtain their characteristic parameters, and then matching these parameters with those in a database. However, this method is limited by a large amount of prior information and manual analysis. By introducing deep learning, and leveraging the data mining capabilities of neural networks, signals can be fed into the neural network to obtain an automatic model for extracting radiation source fingerprint features. The input to the network model is typically signal transform domain information, modulation domain information, or signal image information. When processing signals to obtain network input information, complex transform domain information requires considerable computational time, especially in end-to-end applications. Complex transform processing consumes significant computational resources. The time domain, as the transform domain, offers unique advantages in terms of computational resource conservation. Within a given transform domain, different signal processing methods affect the radiation source fingerprint features represented by the signal. The representation of radiation source fingerprint features exhibits preferences depending on the computational processing method, and single-class processing methods often result in missing radiation source fingerprint features. Therefore, by using various signal processing methods to characterize different types of radiation source fingerprint features within a specified domain, the diversity of radiation source fingerprint feature characterization can be further increased, the completeness of radiation source fingerprint features can be improved, and more comprehensive radiation source fingerprint features can be extracted. Summary of the Invention

[0004] The purpose of this invention is to provide a communication radiation source identification method based on the fusion of multiple time-domain features. This invention adopts a signal time-domain feature extraction method to simplify signal processing calculations and improve the efficiency of the signal processing stage. In addition, according to the calculation methods of different neural networks, the time-domain signal is processed in a corresponding manner, and four different types of radiation source fingerprint features are extracted in the single transform domain of the signal. The radiation source fingerprint features in the signal time domain are fully explored, and the different features are fused to obtain a more comprehensive time-domain radiation source feature.

[0005] To achieve the above-mentioned technical objectives and effects, the present invention is implemented through the following technical solution:

[0006] A method for identifying communication radiation sources based on temporal multi-class feature fusion includes the following steps:

[0007] S1: Set up M individual communication radiation sources and collect signals through signal acquisition equipment;

[0008] The M communication radiation sources are all of the same type (e.g., mobile phones) to identify the differences in radiation source fingerprints between devices of the same type. Signals emitted by the M communication radiation sources are collected using a signal acquisition device and stored as time-domain data, resulting in signal time-domain data s(t).

[0009] S2: The acquired time-domain signal is preprocessed. Energy is detected in the time-domain signal s(t), and a threshold T0 is obtained using simulated noise signal for energy detection. Valid signals are then filtered and retained based on threshold T0. Communication signal transmission has a specific frequency band. The upper limit f1 and lower limit f2 of the communication signal transmission frequency are used as the passband frequencies of the bandpass filter. The signal is then bandpass filtered to obtain the filtered signal s'(t).

[0010] S3: Process signals based on category features;

[0011] S3.1: Normalize the signal s'(t) to compress the signal amplitude, stabilize the time-domain waveform, and obtain the signal S(t) that focuses on the time-domain waveform. Use a drawing tool to display the waveform of the signal S(t) and convert its time-domain signal waveform into image data to obtain the signal time-domain waveform P0.

[0012] S3.2: Compress the amplitude of the signal s'(t), take the difference between the signal amplitude and the mean of the time domain data, divide it by the maximum value of the difference among the absolute values ​​of the time domain data, arrange the results into a one-dimensional array, and obtain the time domain signal array S0(n).

[0013] S3.3: Select an appropriate length n for the time-domain signal length, segment the signal s'(t) into m time-dimension segments, and arrange these segments to obtain the segmented signal matrix X. m,n The segmented signal matrix X m,n Normalization using columns yields the piecewise matrix X′. m,n ;

[0014] S4: Reconstruct the time-domain waveform P0 of the signal (reconstructed to 224*224) and normalize the three channels to form the 'RGB' three-channel original data matrix X. rgb ;

[0015]

[0016] Using a 2D convolutional neural network, the matrix X rgb Feature extraction is performed, and the initial layer uses a three-channel 2D convolution kernel to extract features from matrix X. rgb Perform convolution to extract the individual features F1 of the radiation source;

[0017] S5: Transform the time-domain signal array S0(n) by multiplying all values ​​in the array S0(n) by 255, amplify the normalized array S0(n) values, and use a 1D convolutional neural network to perform 1D convolution on the array S0(n) to extract the individual features F2 of the radiation source.

[0018] S6: The time-domain signal array S0(n) is fed into the LSTM network, and the individual features F3 of the radiation source are extracted through the forward and backward correlation calculation of the neurons.

[0019] S7: Divide the signal into a segmented matrix X' m,n Perform size reconstruction (reconstructing to 224*224) and single-channel normalization to form the single-channel original data matrix X. o Using a 2D convolutional neural network to process the original data matrix X o Perform 2D convolution, starting with a single-channel convolutional layer to extract individual features of the radiation source F4;

[0020] S8: Unify the individual characteristics of the four types of radiation sources to form four types of 2D feature matrices X. F The four types of feature matrices are stacked to form a 3D feature set F. 3D Using 3D convolutional layers on the feature set F 3D Perform 3D convolution to achieve the desired feature set F 3D Perform feature fusion, feature dimensionality reduction, and fusion feature extraction to obtain the fused feature F. f ;

[0021] S9: Fusing features F f Perform dimensional expansion to obtain the fused feature array x f , will fuse feature array x f By connecting to a fully connected network layer, individual radiation sources can be classified.

[0022] Furthermore, in step S2, energy detection is performed on the simulated noise signal to obtain the energy threshold T0:

[0023]

[0024] In the formula, N is the number of samples. x represents the original signal. Denotes the N-point Fourier transform of x, |·|2 This represents the vector norm, used to calculate the energy of a signal x sampled at N points.

[0025] Furthermore, in step S3, the signal s'(t) is segmented and reconstructed into a segmented signal matrix X. m,n In this context, the segment length n and the number of segments m are constrained by the length L of the signal s'(t). s'(t) Their relationship satisfies the following equation:

[0026] L s'(t) =n*m

[0027] n≈m

[0028] Therefore, the piecewise signal matrix X is obtained through reconstruction. m,n Approximately a square matrix, as shown in the following equation:

[0029]

[0030] Furthermore, in steps S4 and S7, a 2D convolutional neural network is used for feature extraction. The parameters of its convolutional layer are {kernel=k,padding=p,stride=s}, and the convolutional kernel is a square matrix A of size k×k. kk For X o Feature extraction using 2D convolution yields the following features:

[0031]

[0032] In the formula H out and W out H represents the size of the output result after convolution. out ×W out Their values ​​are as follows:

[0033]

[0034]

[0035] In the formula, H and W are the original data matrices of a single channel X. o The dimensions are H×W;

[0036] Furthermore, in step S5, a 1D convolutional neural network is used to extract features. The 1D convolutional kernel parameters are {kernel=k,padding=p,stride=s}, and the extracted features F are... i for:

[0037]

[0038] Furthermore, in step S6, LSTM is used for feature extraction to extract feature h. t for:

[0039] f t =σ(W f [h t-1 ,x t ]+b f )

[0040] i t =σ(W i [h t-1 ,x t ]+b i )

[0041]

[0042]

[0043] O t =σ(W o [h t-1 ,x t ]+b o )

[0044] h t =o t *tanh(C t )

[0045] In the formula W f W i W c W o Let b be the weight matrix. f ,b i ,b c ,b o For the corresponding bias term, σ is the activation function sigmoid, tanh is the activation function, and '*' represents the dot product operation; function f t The function i is used to determine which information can be forgotten, controlled by the sigmoid function, which passes through or partially passes through information based on the output of the previous time step; t Used to generate new information that needs updating, the sigmoid function determines which values ​​to use for updating; A tanh layer is used to generate new candidate values; C t Combining the above three steps, discard unnecessary information and add new information; function o t The model's output is determined by obtaining an initial output through a sigmoid layer; h t Use tanh to convert C t The values ​​are scaled to [-1, 1] and then multiplied pairwise with the output obtained from the sigmoid function to obtain the model's output.

[0046] The beneficial effects of this invention are:

[0047] This invention utilizes four different types of time-domain signal processing methods to transform captured electromagnetic signals into four different types of radiation source fingerprint feature representations in the time domain. Different types of neural networks are then associated with these four different types of radiation source fingerprint feature representations. Feature extraction algorithms from these corresponding neural networks extract the fingerprint features of each of the four different types of radiation sources. The feature data is then arranged according to the data format of each category, forming four feature matrices with the same data format. These feature matrices are stacked, and a 3D convolutional network is used to fuse and extract the features from the stacked matrices to obtain the final fused features. Based on the extracted features, a fully connected layer is used to identify individual radiation sources. This invention features fast and simple signal processing, automatic feature extraction, and the ability to obtain more comprehensive radiation source fingerprint features. Specifically, it has the following beneficial effects:

[0048] 1. Efficient Signal Processing Method: This invention uses time-domain signals as the processing object, avoiding complex transform-domain information processing and thus reducing computational resource consumption. Time-domain signals are intuitive and easy to understand. By performing preprocessing operations such as energy detection and bandpass filtering on time-domain signals, signal processing can be performed quickly and easily.

[0049] 2. Automated Feature Extraction: This invention introduces convolutional neural networks and LSTM networks, leveraging their powerful feature extraction capabilities to automatically extract individual features of radiation sources from time-domain signals. Convolutional neural networks can effectively extract spatial features from time-domain signals through local perception and weight sharing, while LSTM networks can capture time-series features from time-domain signals by modeling time-series information.

[0050] 3. Multi-class Feature Fusion: This invention employs various feature processing methods in the time domain, such as time-domain waveforms, time-domain signal arrays, and segmented signal matrices. These methods process the signal from different angles and dimensions, obtaining diverse feature representations. By unifying, stacking, and performing 3D convolution on feature matrices of different categories, feature fusion and extraction are achieved, thereby better uncovering key information in the signal.

[0051] 4. Feature Fusion and Dimensionality Reduction: This invention unifies, stacks, and performs 3D convolution on feature matrices of different categories. The 3D convolutional layer then achieves feature fusion, dimensionality reduction, and feature extraction from the feature set. This method can better extract key information from the signal and improve the classification accuracy of individual radiation sources. Furthermore, by expanding the fused features and using them as input to the fully connected layer, the representational power of the features is further enhanced.

[0052] In summary, this technical solution achieves accurate identification of individual communication radiation sources through efficient signal processing methods, automated feature extraction, multi-class feature fusion, and feature fusion and dimensionality reduction. These methods and technologies have significant application value in the field of individual radiation source identification and provide strong support for research and applications in communication reconnaissance and electronic warfare.

[0053] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0054] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 This is a schematic diagram of radiation source feature extraction.

[0056] Figure 2 This is a schematic diagram of the time-domain signal waveform representation;

[0057] Figure 3 This is a schematic diagram of a 2D convolutional network structure;

[0058] Figure 4 This is a schematic diagram of a 1D convolutional network structure;

[0059] Figure 5 Results of identification of 10 individual radiation sources. Detailed Implementation

[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 are within the scope of protection of the present invention.

[0061] Example 1

[0062] like Figure 1 As shown

[0063] The communication radiation source identification method based on time-domain multi-class feature fusion described in this embodiment includes the following steps:

[0064] S1: Set up M individual communication radiation sources and collect signals through signal acquisition equipment;

[0065] The M communication radiation sources are all of the same type (e.g., mobile phones) to identify the differences in radiation source fingerprints between devices of the same type. Signals emitted by the M communication radiation sources are collected using a signal acquisition device and stored as time-domain data, resulting in signal time-domain data s(t).

[0066] S2: The acquired time-domain signal is preprocessed. Energy is detected in the time-domain signal s(t), and a threshold T0 is obtained using simulated noise signal for energy detection. Valid signals are then filtered and retained based on threshold T0. Communication signal transmission has a specific frequency band. The upper limit f1 and lower limit f2 of the communication signal transmission frequency are used as the passband frequencies of the bandpass filter. The signal is then bandpass filtered to obtain the filtered signal s'(t).

[0067] S3: Process signals based on category features;

[0068] S3.1: Normalize the signal s'(t) to compress the signal amplitude, stabilize the time-domain waveform, and obtain the signal S(t) that focuses on the time-domain waveform. Use a drawing tool to display the waveform of the signal S(t) and convert its time-domain signal waveform into image data to obtain the signal time-domain waveform P0.

[0069] S3.2: Compress the amplitude of the signal s'(t), take the difference between the signal amplitude and the mean of the time domain data, divide it by the maximum value of the difference among the absolute values ​​of the time domain data, arrange the results into a one-dimensional array, and obtain the time domain signal array S0(n).

[0070] S3.3: Select an appropriate length n for the time-domain signal length, segment the signal s'(t) into m time-dimension segments, and arrange these segments to obtain the segmented signal matrix X. m,n The segmented signal matrix X m,n Normalization using columns yields the piecewise matrix X′. m,n ;

[0071] S4: Reconstruct the time-domain waveform P0 of the signal (reconstructed to 224*224) and normalize the three channels to form the 'RGB' three-channel original data matrix X. rgb ;

[0072]

[0073] Using a 2D convolutional neural network, the matrix X rgb Feature extraction is performed, and the initial layer uses a three-channel 2D convolution kernel to extract features from matrix X. rgb Perform convolution to extract the individual features F1 of the radiation source;

[0074] S5: Transform the time-domain signal array S0(n) by multiplying all values ​​in the array S0(n) by 255, amplify the normalized array S0(n) values, and use a 1D convolutional neural network to perform 1D convolution on the array S0(n) to extract the individual features F2 of the radiation source.

[0075] S6: The time-domain signal array S0(n) is fed into the LSTM network, and the individual features F3 of the radiation source are extracted through the forward and backward correlation calculation of the neurons.

[0076] S7: Divide the signal into a segmented matrix X' m,n Perform size reconstruction (reconstructing to 224*224) and single-channel normalization to form the single-channel original data matrix X. o Using a 2D convolutional neural network to process the original data matrix X o Perform 2D convolution, starting with a single-channel convolutional layer to extract individual features of the radiation source F4;

[0077] S8: Unify the individual characteristics of the four types of radiation sources to form four types of 2D feature matrices X. F The four types of feature matrices are stacked to form a 3D feature set F. 3D Using 3D convolutional layers on the feature set F 3D Perform 3D convolution to achieve the desired feature set F 3D Perform feature fusion, feature dimensionality reduction, and fusion feature extraction to obtain the fused feature F. f ;

[0078] S9: Fusing features F f Perform dimensional expansion to obtain the fused feature array x f , will fuse feature array x f By connecting to a fully connected network layer, individual radiation sources can be classified.

[0079] In this embodiment, in step S2, energy detection is performed on the simulated noise signal to obtain the energy threshold T0 as follows:

[0080]

[0081] In the formula, N is the number of samples.

[0082] In this embodiment, in step S3, the signal s'(t) is segmented and reconstructed into a segmented signal matrix X. m,n In this context, the segment length n and the number of segments m are constrained by the length L of the signal s'(t). s'(t) Their relationship satisfies the following equation:

[0083] L s′(t) =n*m

[0084] n≈m

[0085] Therefore, the piecewise signal matrix X is obtained through reconstruction. m,n Approximately a square matrix, as shown in the following equation:

[0086]

[0087] In this embodiment, in steps S4 and S7, a 2D convolutional neural network is used for feature extraction. The parameters of its convolutional layer are {kernel = k, padding = p, stride = s}, and the convolutional kernel is a square matrix A of size k×k. kk For X o Feature extraction using 2D convolution yields the following features:

[0088]

[0089] In the formula H out and W out H represents the size of the output result after convolution. out ×W out Their values ​​are as follows:

[0090]

[0091]

[0092] In the formula, H and W are the original data matrices of a single channel X. o The dimensions are H×W;

[0093] In this embodiment, in step S5, a 1D convolutional neural network is used to extract features. The 1D convolutional kernel parameters are {kernel=k,padding=p,stride=s}, and the extracted features F i for:

[0094]

[0095] In this embodiment, in step S6, LSTM is used for feature extraction to extract feature h. t for:

[0096] f t =σ(W f [h t-1 ,x t ]+b f )

[0097] i t =σ(W i [h t-1 ,x t ]+b i )

[0098]

[0099]

[0100] O t =σ(W o [h t-1 ,x t ]+b o )

[0101] h t =o t *tanh(C t )

[0102] In the formula W f W i W c W o Let b be the weight matrix. f ,b i ,b c ,b o For the corresponding bias term, σ is the activation function sigmoid, tanh is the activation function, and '*' is the dot product operation.

[0103] Example 2

[0104] like Figures 2 to 5 As shown

[0105] To better illustrate the method of the present invention, the following simulation experimental conditions are constructed in this embodiment:

[0106] S1: For 10 radiation source individuals (M=10 in number), their signals are modulated in the same way, resulting in 10 types of radiation source signals of the same type, to simulate radiation source signals of the same type. The 10 types of signals are modulated to ensure that the signal amplitude and frequency are the same, and Gaussian white noise is added to the signals to obtain 10,000 data points at different signal-to-noise ratios, i.e., 1,000 data points for each type of signal. For each signal data point, a data storage period of 3000 seconds is recorded as signal s(t).

[0107] S2: Perform energy detection on the signal s(t) to obtain the signal energy value T. s :

[0108]

[0109] The noise energy value T0 is obtained by energy detection of the simulated noise signal. An energy ratio threshold α = 10 is set. The energy ratio is obtained by comparing the signal energy value with the noise energy value, and then compared with the threshold α to filter the signal. The signal s(t) has a frequency of f0. A bandpass filter is applied to the signal s(t) with f1 = 0.5 × f0 and f2 = 1.5 × f0 to obtain the filtered signal s'(t).

[0110] S3: Process signals based on category features;

[0111] S3.1: Normalize the signal s'(t) to obtain the signal S(t):

[0112]

[0113] Using a plotting tool, the normalized and compressed signal S(t) is displayed as a waveform graph. Detailed characteristics of the time-domain signal waveform are then highlighted, such as… Figure 2 As shown, the waveform is converted into image data to obtain the signal time-domain waveform P0:

[0114]

[0115] P0 is a three-channel image matrix, where P r P g P b This refers to single-channel data in graphical data format.

[0116] S3.2: Perform amplitude compression on the signal s'(t) to obtain the signal S0(t):

[0117]

[0118] S0(t) is a finite-length signal, which is arranged as a one-dimensional array of time-domain signals S0(n).

[0119] S3.3: The time-domain signal length is 3000. Choosing a suitable length n = 50, the signal s'(t) is segmented into m = 60 time-dimension signal segments. These segments are then arranged to obtain the segmented signal matrix X. m,n :

[0120]

[0121] Each signal segment is arranged sequentially according to its time interval, forming a 60*50 two-dimensional matrix. This transforms each time-domain signal data from a one-dimensional time-domain array into a two-dimensional signal matrix, enabling the characterization of the radiation source that highlights the structural features of the time-domain signal. The segmented signal matrix X... m,n Normalization of the columns yields the piecewise matrix X'. m,n :

[0122]

[0123] In the formula X n For the piecewise signal matrix X m,n The nth column.

[0124] S4: Reconstruct the signal time-domain waveform P0 (to 224*224) and normalize the three channels to form the 'RGB' three-channel original data matrix X. rgb ;

[0125]

[0126] Size reconstruction is achieved by interpolating or sampling the pixels in the image data to change the size of the image.

[0127] The processed matrix X rgb As input to a 2D convolutional neural network, features are extracted using the convolutional neural network. The structure of the 2D convolutional neural network is as follows: Figure 3 As shown. The input matrix X is processed by 2D convolution. rgb Feature extraction is performed to obtain feature F1.

[0128]

[0129] in The formula can be used to calculate the size of the feature map extracted after each convolutional layer.

[0130] S5: Based on the time-domain signal array S0(n), its amplitude is proportionally amplified to increase the magnitude of the signal array. This amplified signal is then used as input to a 1D convolutional neural network. The 1D convolution is then used to extract features from the time-domain signal array. The structure of the 1D convolutional neural network is as follows: Figure 4 As shown, by performing 1D convolution on the time-domain signal array, the overall correlated radiation source feature F2 in the time domain of the signal is extracted;

[0131] S6: Similar to S5, the time-domain signal array S0(n) is amplified and then used as the input of the LSTM network. The LSTM network extracts the radiation source features of the input to obtain the time-series radiation source features F3 in the signal time domain.

[0132] S7: Divide the segmented signal matrix X' m,n After processing similarly to S4, but using single-channel data here, a 2D convolutional network is also used for feature extraction to obtain the signal temporal structure-related radiation source features F4.

[0133] S8: For the four different types of time-domain features F1, F2, F3, and F4 extracted from the signal, analyze the data format of the four types of features, unify them into a 2D matrix form through a rearrangement method, and then stack the four types of features to form a 3D feature set F. 3D The final feature extraction is performed using 3D convolution to obtain the temporal fusion feature F of the signal. f ;

[0134] S9: Based on signal time-domain fusion features F f F is analyzed through a fully connected neural network. f Calculations were performed to obtain the differences in characteristics between different radiation sources, and these differences were converted into category features to classify individual radiation sources. The identification results for simulated signal data from 10 individual radiation sources are as follows: Figure 5 As shown, the recognition rate curves during network training represent the results, with a recognition rate of 96%.

[0135] This invention employs neural networks for feature extraction in the time domain of communication signals. First, a 2D convolutional neural network is used to extract waveform features from the signal's time-domain data, obtaining individual features of radiation sources related to the signal's waveform. Then, a 1D convolutional neural network is used to extract features from the signal's time-domain data, obtaining individual features of radiation sources related to the overall numerical value of the signal's time-domain data. Next, the signal's time-domain data is rearranged, and a 2D convolutional neural network is used to extract features, obtaining individual features of radiation sources related to the structure of the signal's time-domain data. Finally, a recurrent neural network is used to extract features from the signal's time-domain data, obtaining individual features of radiation sources related to the time-series correlation. An attention mechanism is used to optimize four different types of features in the time domain. These four types of features are then spatially concatenated to form a feature set. A 3D convolutional neural network is used to fuse and extract features from this feature set. Finally, a fully connected network layer is used to achieve individual radiation source classification and identification.

[0136] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for identifying communication radiation sources based on temporal multi-class feature fusion, characterized in that, Includes the following steps: S1: Set up M individual communication radiation sources and collect signals through signal acquisition equipment; The M communication radiation sources are all of the same type to identify the differences in radiation source fingerprints between devices of the same type. The signals emitted by the M communication radiation sources are collected by the signal acquisition device and stored in the form of time-domain data, and the obtained signal time-domain data is s(t). S2: Preprocess the acquired time-domain signal, converting the time-domain signal... Energy detection is performed, and a threshold is obtained by simultaneously using simulated noise signals for energy detection. Through threshold Valid signals are filtered and retained; the communication signal transmission frequency band is limited to the upper limit of the communication signal transmission frequency. and transmission frequency lower limit The passband frequency of a bandpass filter is used to perform bandpass filtering on a signal to obtain the filtered signal. ; S3: Process the signal based on category features to obtain the signal's time-domain waveform. Time-domain signal array Segmented signal matrix Piecewise matrix ; S4: Plot the signal time-domain waveform The dimensions were reconstructed and the three channels were normalized to a size of 224*224, forming an 'RGB' three-channel original data matrix. ; ； Using 2D convolutional neural networks, for matrices Feature extraction is performed, and the initial layer uses a three-channel 2D convolution kernel to extract features from the matrix. Convolution is performed to extract individual features of the radiation source. ; S5: Convert the time-domain signal array Perform a transformation to make the array Multiply all values ​​by 255, then enlarge and normalize the array. The values ​​are obtained by using a 1D convolutional neural network on the array. Perform 1D convolution to extract individual features of the radiation source. ; S6: Convert the time-domain signal array The data is fed into an LSTM network, and individual features of the radiation source are extracted through the correlation calculation between neurons. ; S7: Divide the signal into a segmented matrix. The dimensions are reconstructed and single-channel normalized to a 224*224 format, forming the original single-channel data matrix. Using a 2D convolutional neural network to process the original data matrix 2D convolution is performed, with a single-channel convolutional layer used as the initial layer to extract individual features of the radiation source. ; S8: Unify the individual characteristics of the four types of radiation sources to form four types of 2D feature matrices. The four types of feature matrices are stacked to form a 3D feature set. Using 3D convolutional layers on the feature set Perform 3D convolution to achieve the processing of feature sets Perform feature fusion, feature dimensionality reduction, and fusion feature extraction to obtain fused features. ; S9: Integrating Features Perform dimensional expansion to obtain the fused feature array. , will merge feature array By connecting to a fully connected network layer, individual radiation sources can be classified.

2. The communication radiation source identification method based on time-domain multi-class feature fusion according to claim 1, characterized in that: In step S2, energy detection is performed on the simulated noise signal to obtain the energy threshold. for: In the formula, N is the number of samples, and x represents the original signal. Represents the N-point Fourier transform of x. This represents the vector norm, used to calculate the energy of a signal x sampled at N points.

3. The communication radiation source identification method based on time-domain multi-class feature fusion according to claim 1, characterized in that: Step S3 specifically includes the following sub-steps: S3.1: Send signal Normalization is performed to compress the signal amplitude, stabilize the time-domain waveform, and obtain a signal that emphasizes the time-domain waveform representation. Use drawing tools to plot the signal The waveform is displayed, and its time-domain signal waveform is converted into image data to obtain the signal time-domain waveform. ; S3.2: Send signal Amplitude compression is performed by subtracting the signal amplitude from the mean of the time-domain data, then dividing by the maximum difference among the absolute values ​​of the time-domain data. The results are then arranged into a one-dimensional array to obtain the time-domain signal array. ; S3.3: Selecting the segment length based on the length of the time-domain signal , will signal Segmentation is performed to obtain Given a signal segment with a time dimension, arrange the signal segments to obtain a segmented signal matrix. ; segment the signal matrix Normalization using columns yields a piecewise matrix. .

4. The communication radiation source identification method based on time-domain multi-class feature fusion according to claim 1, characterized in that: In step S3, the signal Perform segmented arrangement and reconstruction into segmented signal matrix In the middle, the length of the segmented signal and number of segments Constrained by signal length Their relationship satisfies the following equation: Therefore, the piecewise signal matrix is ​​obtained through reconstruction. Approximately a square matrix, as shown in the following equation: 。 5. The communication radiation source identification method based on time-domain multi-class feature fusion according to claim 1, characterized in that: In steps S4 and S7, feature extraction is performed using a 2D convolutional neural network, with the following convolutional layer parameters: The convolution kernel is of size square array ;right Feature extraction using 2D convolution yields the following features: In the formula and Indicates the size of the output result after convolution. Their values ​​are as follows: In the formula and Single-channel original data matrix Size .

6. The communication radiation source identification method based on time-domain multi-class feature fusion according to claim 1, characterized in that: In step S5, a 1D convolutional neural network is used to extract features, and the parameters of the 1D convolutional kernel are as follows: Feature extraction for: 。 7. The communication radiation source identification method based on time-domain multi-class feature fusion according to claim 1, characterized in that: In step S6, LSTM is used for feature extraction to extract features. for: In the formula This is the weight matrix. For the corresponding bias term, Let sigmoid be the activation function and tanh be the activation function. Dot product; function The function determines which information can be forgotten, controlled by the sigmoid function, which passes through or partially passes through information based on the output of the previous time step; Used to generate new information that needs updating, the sigmoid function determines which values ​​to use for updating; The tanh layer is used to generate new candidate values; Combining the above three steps, discard unnecessary information and add new information; function The model's output is determined by obtaining an initial output through a sigmoid layer; Use tanh The values ​​are scaled to [-1, 1] and then multiplied pairwise with the output obtained from the sigmoid function to obtain the model's output.

Images