Radio frequency fingerprint identification method and system based on multi-channel time-frequency joint attention mechanism

Abstract

The invention discloses a radio frequency fingerprint identification method and system based on a multi-channel time-frequency joint attention mechanism. The method comprises the following steps: receiving a radio frequency signal; training the received radio frequency signal, wherein the training comprises the following steps of: carrying out normalization processing on a radio frequency baseband signal; performing short-time Fourier transform on the radio frequency signal after normalization processing to obtain two-dimensional time-frequency representation; taking the two-dimensional time-frequency representation as the input of a dynamic channel attention network with a residual structure, and obtaining a multi-channel two-dimensional time-frequency representation containing channel attention weight weighting; taking the multi-channel two-dimensional time-frequency representation as the input of a time-frequency joint attention network, and obtaining a weighted feature tensor containing a time-frequency joint attention weight; taking the feature tensor as an input of a classification network to obtain a prediction probability vector of each category to which the radio frequency signal belongs; updating network parameters according to the prediction probability vector; and repeatedly training the received radio frequency signal until the network converges. Therefore, the classification accuracy of the classification network in a low signal-to-noise ratio environment is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of radio frequency fingerprint recognition, and in particular to a radio frequency fingerprint recognition method and system based on a multi-channel time-frequency joint attention mechanism. Technical Background

[0002] With the development of Internet of Things (IoT) technology, IoT devices have covered all areas of people's daily life. The ensuing IoT security issues have brought new challenges to IoT device authentication algorithms.

[0003] Radio Frequency Fingerprint Identification (RFFI) is a technology that authenticates IoT devices at the physical layer. This method does not bring the IoT device the additional computing and storage overhead required by the authentication algorithm, and the RF fingerprint comes from the inherent hardware defects of the IoT device itself. This hardware defect is not easy to forge and will cause slight differences in the RF signal sent by the device, which can be used by the server to authenticate the identity of the IoT device.

[0004] However, most existing RF fingerprint recognition methods will experience performance degradation in low signal-to-noise ratio environments. There are two main solutions to this problem. One solution is to design an additional noise reduction module to first reduce the noise of the input RF signal before inputting it into the subsequent classification network, thereby reducing the impact of noise on recognition accuracy; the other solution is to design an additional feature extraction module to extract more complex internal dependencies in the noisy RF signal, enhance the algorithm's robustness to noise, and improve classification accuracy. However, the performance improvement of the corresponding algorithms of the above two solutions in low signal-to-noise ratio environments is still not significant. RF fingerprint recognition technology requires a more powerful noise reduction module or a more effective feature extraction module to solve the problem of performance degradation in low signal-to-noise ratio environments. Summary of the invention

[0005] In order to solve the above technical problems, the present invention proposes a radio frequency fingerprint recognition method and system based on a multi-channel time-frequency joint attention mechanism. The dynamic channel attention mechanism and the time-frequency joint attention mechanism of this method can capture more complex internal dependencies in noisy radio frequency signals and improve the recognition performance of the recognition algorithm in a low signal-to-noise ratio environment.

[0006] In a first aspect, a radio frequency fingerprint recognition method based on a multi-channel time-frequency joint attention mechanism is provided, the method comprising the following steps:

[0007] receiving radio frequency signals;

[0008] Training the received RF signal includes:

[0009] Performing normalization processing on the radio frequency signal;

[0010] Perform short-time Fourier transform on the normalized RF signal to obtain a two-dimensional time-frequency representation;

[0011] The two-dimensional time-frequency representation is used as the input of the dynamic channel attention network with a residual structure to obtain a multi-channel two-dimensional time-frequency representation weighted by channel attention weights;

[0012] The multi-channel two-dimensional time-frequency representation is used as the input of the time-frequency joint attention network to obtain a feature tensor weighted by the time-frequency joint attention weights.

[0013] The feature tensor is used as the input of the classification network to obtain the predicted probability vector of each category to which the RF signal belongs;

[0014] Update network parameters according to the predicted probability vector;

[0015] Repeat the training of the received RF signal until the network converges;

[0016] Output the optimal network parameters for the testing process.

[0017] Preferably, the normalizing the radio frequency signal specifically includes:

[0018] Assume that the received RF complex signal discrete sequence is The maximum value r of the real and imaginary parts of the complex signal sequence is calculated according to the following formula max and the minimum value r min :

[0019] r max =max{Re{r[n]},Im{r[n]}}

[0020] r min =min{Re{r[n]},Im{r[n]}}

[0021] Perform maximum-minimum normalization on the received RF complex signal to obtain a normalized RF complex signal

[0022]

[0023] Preferably, the step of performing short-time Fourier transform on the normalized radio frequency signal to obtain a two-dimensional time-frequency representation specifically includes:

[0024] The one-dimensional N-point RF complex signal sequence r bn [n] Perform short-time Fourier transform to obtain a two-dimensional complex matrix

[0025]

[0026] Among them, S(n,m) is the element corresponding to the nth row and mth column of the two-dimensional complex matrix S, N F is the number of Fourier transform points, N T is the number of time windows for short-time Fourier transform, w[k] is the window length N F Hanning window function, h is the sliding step of the window function, 0≤n≤N F -1, 0≤m≤N T -1;

[0027] Take the two-dimensional complex matrix The real and imaginary parts are spliced ​​into a two-dimensional time-frequency representation of the dual channel As input to the subsequent network.

[0028] Preferably, the two-dimensional time-frequency representation is used as the input of a dynamic channel attention network with a residual structure to obtain a multi-channel two-dimensional time-frequency representation weighted with channel attention weights, specifically including:

[0029] The dual-channel two-dimensional time-frequency representation is transformed by two-dimensional convolution Perform channel expansion to obtain multi-channel two-dimensional time-frequency representation

[0030] X c =Conv 3×3 (X)

[0031] Utilizing X c Do average pooling and standard deviation pooling to get and Using F avg and F std Calculate the dynamic channel attention weight coefficient

[0032] w c =Sigmoid(Conv 1×1 (ReLu(Conv 1×1 (F av g+ Fstd ))))

[0033] Using w c The multi-channel two-dimensional time-frequency representation X c Perform channel weighting, and then use the residual structure to add it to the initial two-channel two-dimensional video representation X to obtain a multi-channel two-dimensional time-frequency representation weighted with channel attention weights

[0034] X cw =X c ×c +Conv 3×3 (X).

[0035] Preferably, the multi-channel two-dimensional time-frequency representation is used as the input of the time-frequency joint attention network to obtain a feature tensor weighted by the time-frequency joint attention weight, specifically including:

[0036] Multi-channel two-dimensional time-frequency representation Perform multi-scale two-dimensional convolution to extract features and get

[0037] X m =ReLu(Conv 1×1 (X cw ))+ReLu(Conv 3×3 (X cw ))+ReLu(Conv 5×5 (X cw ))

[0038] Use N F ×1 and 1×N T The convolution kernel X m Do a two-dimensional convolution to get the frequency attention weight F f and the temporal attention weight F t :

[0039]

[0040] Using frequency attention weight F f and the temporal attention weight F t , combined with the automatic broadcast mechanism and the dynamic gating mechanism, the time-frequency joint attention weight coefficient is calculated

[0041] w tf =Sigmoid(F f ×F t ×AvgPool(FCF f ×F t )))

[0042] Using the time-frequency joint attention weight coefficient w tf Multi-channel two-dimensional time-frequency representation X cw Multiply them together to get the feature tensor X weighted by the time-frequency joint attention weight ctf :

[0043] X ctf =X cw × tf .

[0044] Preferably, the step of using the feature tensor as an input of a classification network to obtain a prediction probability vector of each category to which the radio frequency signal belongs specifically includes:

[0045] The feature tensor weighted with time-frequency joint attention weights Feature extraction and classification are performed through a two-dimensional convolutional network, the two-dimensional convolutional network is AlexNet;

[0046] X ctf After input into AlexNet, the predicted probability vector of the original RF signal belonging to each category is output M is the number of transmitter categories.

[0047] Preferably, updating the network parameters according to the predicted probability vector specifically includes:

[0048] Select cross entropy as the loss function and use Adam optimizer to update parameters; the calculation formula of cross entropy is:

[0049]

[0050] in, is a one-hot encoding vector, if the RF signal transmitter category is m, then p[m] is 1, otherwise it is 0.

[0051] In the second aspect, a radio frequency fingerprint recognition system based on a multi-channel time-frequency joint attention mechanism is provided, and the system includes the following modules:

[0052] A receiving module, used for receiving radio frequency signals;

[0053] The training module is used to train the received radio frequency signal, including:

[0054] A normalization processing module, used for performing maximum-minimum normalization processing on the received radio frequency signal;

[0055] A short-time Fourier transform module is used to transform a one-dimensional radio frequency signal into a two-dimensional time-frequency representation;

[0056] Dynamic channel attention module, which is used to extract dynamic channel attention weight coefficients based on two-dimensional time-frequency representation to obtain multi-channel two-dimensional time-frequency representation weighted with channel attention weights;

[0057] The time-frequency joint attention module is used to extract the time-frequency joint attention weight coefficient based on the multi-channel two-dimensional time-frequency representation and then obtain the feature tensor weighted by the time-frequency joint attention weight;

[0058] A probability prediction module is used to use the feature tensor weighted by the time-frequency joint attention weight as the input of the classification network to obtain the prediction probability vector of each category to which the RF signal belongs;

[0059] An updating module, used for updating network parameters according to the predicted probability vectors of each category to which the radio frequency signal belongs;

[0060] An iterative training module, used to repeatedly train the received RF signal until the network converges;

[0061] The output module is used to output the optimal network parameters for the testing process.

[0062] In a third aspect, a computer-readable storage medium storing one or more programs is provided, wherein the one or more programs include instructions, which, when executed by a computing device, cause the computing device to perform any of the methods described.

[0063] In a fourth aspect, a computing device is provided, including:

[0064] One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs include instructions for executing any of the methods described.

[0065] The beneficial effects of the present invention are:

[0066] 1. Using the dynamic channel attention mechanism, the network can focus more on channels with more important information, so as to highlight effective RF signals and suppress noise interference;

[0067] 2. Use the time-frequency joint attention mechanism to capture the long-range dependencies of RF signals in the time and frequency domains, and process the RF fingerprint features of RF signals to be more discriminative;

[0068] 3. By combining the dynamic channel attention mechanism and the time-frequency joint attention mechanism, the constructed multi-channel time-frequency joint attention network can adaptively calculate the appropriate channel attention according to the signal-to-noise ratio, and enhance the RF fingerprint characteristics of the RF signal according to the time-frequency joint attention, thereby effectively improving the performance of the classification network in a low signal-to-noise ratio environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] In order to more clearly illustrate the technical solutions in the examples of the present invention, the following briefly introduces the drawings required for the description of the examples. Obviously, the drawings described below are only some examples of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative work.

[0070] Attached Figure 1 It is a flow chart of the training process of the radio frequency fingerprint recognition algorithm based on the multi-channel time-frequency joint attention mechanism of the present invention;

[0071] Attached Figure 2 A schematic diagram of the network structure of the radio frequency fingerprint recognition algorithm based on the multi-channel time-frequency joint attention mechanism of the present invention;

[0072] Attached Figure 3 This is a network structure diagram of the dynamic channel attention module in the present invention;

[0073] Attached Figure 4 This is the network structure diagram of the time-frequency joint attention module in the present invention. DETAILED DESCRIPTION

[0074] In order to make the purpose and technical solution of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be described more clearly and completely in combination with the drawings of the embodiments in the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present application.

[0075] Since the present invention is based on a deep neural network, the description of the specific implementation method will be described separately from the two aspects of the training process and the testing process.

[0076] like Figure 1-4 As shown, the present invention provides a radio frequency fingerprint recognition method based on a multi-channel time-frequency joint attention mechanism, comprising the following steps:

[0077] receiving radio frequency signals;

[0078] Training the received RF signal includes:

[0079] Performing normalization processing on the radio frequency signal;

[0080] Perform short-time Fourier transform on the normalized RF signal to obtain a two-dimensional time-frequency representation;

[0081] The two-dimensional time-frequency representation is used as the input of the dynamic channel attention network with a residual structure to obtain a multi-channel two-dimensional time-frequency representation weighted by channel attention weights;

[0082] The multi-channel two-dimensional time-frequency representation is used as the input of the time-frequency joint attention network to obtain a feature tensor weighted by the time-frequency joint attention weights.

[0083] The feature tensor is used as the input of the classification network to obtain the predicted probability vector of each category to which the RF signal belongs;

[0084] Update network parameters according to the predicted probability vector;

[0085] Repeat the training of the received RF signal until the network converges;

[0086] Output the optimal network parameters for the testing process.

[0087] Specifically, the normalization process of the radio frequency signal includes the following steps:

[0088] (1.1) Assume that the received RF complex signal discrete sequence is The maximum value r of the real and imaginary parts of the complex signal sequence is calculated according to the following formula max and the minimum value r min :

[0089] r max =max{Re{r[n]},Im{r[n]}}

[0090] r min =min{Re{r[n]},Im{r[n]}}

[0091] (1.2) Perform maximum-minimum normalization processing on the received RF complex signal to obtain a normalized RF complex signal

[0092]

[0093] Specifically, performing short-time Fourier transform on the normalized radio frequency signal to obtain a two-dimensional time-frequency representation includes the following steps:

[0094] (2.1) The one-dimensional N-point RF complex signal sequence r bn [n] Perform short-time Fourier transform to obtain a two-dimensional complex matrix

[0095]

[0096] Among them, S(n,m) is the element corresponding to the nth row and mth column of the two-dimensional complex matrix S, N F is the number of Fourier transform points, N T is the number of time windows for short-time Fourier transform, w[k] is the window length N F Hanning window function, h is the sliding step of the window function, 0≤n≤N F -1, 0≤m≤N T -1;

[0097] (2.2) Take the two-dimensional complex matrix The real and imaginary parts are spliced ​​into a two-dimensional time-frequency representation of the dual channel As input to the subsequent network.

[0098] Specifically, the two-dimensional time-frequency representation is used as the input of a dynamic channel attention network with a residual structure to obtain a multi-channel two-dimensional time-frequency representation weighted with channel attention weights, including the following steps:

[0099] (3.1) Use two-dimensional convolution to transform the dual-channel two-dimensional time-frequency representation obtained in step (2.2) Perform channel expansion to obtain multi-channel two-dimensional time-frequency representation

[0100] X c =Conv 3×3 (X)

[0101] (3.2) Using X c Do average pooling and standard deviation pooling to get and Using F avg and F std Calculate the dynamic channel attention weight coefficient

[0102] w c =Sigmoid(Conv 1×1 (ReLu(Conv 1×1 (F avg +F std ))))

[0103] (3.3) Using w c The multi-channel two-dimensional time-frequency representation X c Perform channel weighting and then use Figure 3 The residual structure shown is added to the initial two-channel two-dimensional video representation X to obtain a multi-channel two-dimensional time-frequency representation weighted by channel attention weights

[0104] X cw =X c × c +Conv 3×3 (X).

[0105] Specifically, the multi-channel two-dimensional time-frequency representation is used as the input of the time-frequency joint attention network to obtain a feature tensor weighted by the time-frequency joint attention weight, including the following steps:

[0106] (4.1) The multi-channel two-dimensional time-frequency representation obtained in step (3.3) Perform multi-scale two-dimensional convolution to extract features and get

[0107] X m =ReLu(Conv 1×1 (Xcw ))+ReLu(Conv 3×3 (X cw ))+ReLu(Conv 5×5 (X cw ))

[0108] (4.2) respectively using N F ×1 and 1×N T The convolution kernel X m Do a two-dimensional convolution to get the frequency attention weight F f and the temporal attention weight F t :

[0109]

[0110] (4.3) Using the frequency attention weight F f and the temporal attention weight F t , combined with the automatic broadcast mechanism and dynamic gating mechanism, the time-frequency joint attention weight coefficient is calculated

[0111] w tf =Sigmoid(F f ×F t ×AvgPool(FC(F f ×F t )))

[0112] (4.4) Figure 4 As shown in the network structure, the time-frequency joint attention weight coefficient w is used tf With multi-channel two-dimensional time-frequency representation X cw Multiply them together to get the feature tensor X weighted by the time-frequency joint attention weight ctf :

[0113] X ctf =X cw × tf .

[0114] Specifically, the feature tensor is used as the input of the classification network to obtain the predicted probability vector of each category to which the RF signal belongs, including the following steps:

[0115] (5.1) The feature tensor weighted by the time-frequency joint attention weight Feature extraction and classification are performed through a two-dimensional convolutional network. The two-dimensional convolutional network selected in this scheme is AlexNet;

[0116] (5.2) X ctf After input into AlexNet, the predicted probability vector of the original RF signal belonging to each category is output M is the number of transmitter categories.

[0117] Specifically, the network parameters are updated according to the predicted probability vector. The cross entropy is selected as the loss function, and the Adam optimizer is used to update the parameters; the calculation formula of the cross entropy is:

[0118]

[0119] in, is a one-hot encoding vector, if the RF signal transmitter category is m, then p[m] is 1, otherwise it is 0.

[0120] The relevant parameters of the network are determined in the training phase. The testing process consists of eight steps. The first seven steps are the same as the training process (because the network parameters have been determined during the training process, there is no need to update the parameters), and step eight "update the network parameters according to the predicted probability vector" is modified as follows:

[0121] Step eight, output the predicted category of the signal.

[0122] The category with the highest predicted probability is taken as the category predicted by the network

[0123]

[0124] The present invention provides a radio frequency fingerprint recognition system based on a multi-channel time-frequency joint attention mechanism, and the system includes the following modules:

[0125] A receiving module, used for receiving radio frequency signals;

[0126] The training module is used to train the received radio frequency signal, including:

[0127] A normalization processing module, used for performing maximum-minimum normalization processing on the received radio frequency signal;

[0128] A short-time Fourier transform module is used to transform a one-dimensional radio frequency signal into a two-dimensional time-frequency representation;

[0129] Dynamic channel attention module, which is used to extract dynamic channel attention weight coefficients based on two-dimensional time-frequency representation to obtain multi-channel two-dimensional time-frequency representation weighted with channel attention weights;

[0130] The time-frequency joint attention module is used to extract the time-frequency joint attention weight coefficient based on the multi-channel two-dimensional time-frequency representation and then obtain the feature tensor weighted by the time-frequency joint attention weight;

[0131] A probability prediction module is used to use the feature tensor weighted by the time-frequency joint attention weight as the input of the classification network to obtain the prediction probability vector of each category to which the RF signal belongs;

[0132] An updating module, used for updating network parameters according to the predicted probability vectors of each category to which the radio frequency signal belongs;

[0133] An iterative training module, used to repeatedly train the received RF signal until the network converges;

[0134] The output module is used to output the optimal network parameters for the testing process.

[0135] The embodiments of the present application can be provided as methods or computer program products. Therefore, the present application can adopt the form of complete hardware embodiment, complete software embodiment, or the embodiment in combination with software and hardware. Moreover, the present application can adopt the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiments of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and literal scripting language JavaScript, etc.

[0136] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0137] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0138] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1The steps for the functions specified in one or more boxes.

[0139] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

Claims

1. A radio frequency fingerprint recognition method based on a multi-channel time-frequency joint attention mechanism, characterized in that: The method comprises the following steps: receiving radio frequency signals; Training the received RF signal includes: Performing normalization processing on the radio frequency signal; Perform short-time Fourier transform on the normalized RF signal to obtain a two-dimensional time-frequency representation; The two-dimensional time-frequency representation is used as the input of the dynamic channel attention network with a residual structure to obtain a multi-channel two-dimensional time-frequency representation weighted by channel attention weights; The multi-channel two-dimensional time-frequency representation is used as the input of the time-frequency joint attention network to obtain a feature tensor weighted by the time-frequency joint attention weights. The feature tensor is used as the input of the classification network to obtain the predicted probability vector of each category to which the RF signal belongs; Update network parameters according to the predicted probability vector; Repeat the training of the received RF signal until the network converges; Output the optimal network parameters for the testing process.

2. The radio frequency fingerprint recognition method based on multi-channel time-frequency joint attention mechanism according to claim 1 is characterized in that: The normalization processing of the radio frequency signal specifically includes: Assume that the received RF complex signal discrete sequence is The maximum value r of the real and imaginary parts of the complex signal sequence is calculated according to the following formula max and the minimum value r min : r max =max{Re{r[n]},Im{r[n]}} r min =min{Re{r[n]},Im{r[n]}} Perform maximum-minimum normalization on the received RF complex signal to obtain a normalized RF complex signal 3. The radio frequency fingerprint recognition method based on multi-channel time-frequency joint attention mechanism according to claim 1 is characterized in that: The step of performing short-time Fourier transform on the normalized radio frequency signal to obtain a two-dimensional time-frequency representation specifically includes: The one-dimensional N-point RF complex signal sequence r bn [n] Perform short-time Fourier transform to obtain a two-dimensional complex matrix Among them, S(n,m) is the element corresponding to the nth row and mth column of the two-dimensional complex matrix S, N F is the number of Fourier transform points, N T is the number of time windows for short-time Fourier transform, w[k] is the window length N F Hanning window function, h is the sliding step of the window function, 0≤n≤N F -1, 0≤m≤N T -1; Take the two-dimensional complex matrix The real and imaginary parts are spliced ​​into a two-dimensional time-frequency representation of the dual channel As input to the subsequent network.

4. The radio frequency fingerprint recognition method based on multi-channel time-frequency joint attention mechanism according to claim 1 is characterized in that: The two-dimensional time-frequency representation is used as the input of the dynamic channel attention network with a residual structure to obtain a multi-channel two-dimensional time-frequency representation weighted with channel attention weights, specifically including: The dual-channel two-dimensional time-frequency representation is transformed by two-dimensional convolution For channel expansion, N F is the number of Fourier transform points, N T is the number of time windows of the short-time Fourier transform, and a multi-channel two-dimensional time-frequency representation is obtained X c =Conv 3×3 (X) Utilizing X c Do average pooling and standard deviation pooling to get and Using F avg and F std Calculate the dynamic channel attention weight coefficient w c =Sigmoid(Conv 1×1 (ReLu(Conv 1×1 (F avg +F std )))) Using w c The multi-channel two-dimensional time-frequency representation X c Perform channel weighting, and then use the residual structure to add it to the initial two-channel two-dimensional video representation X to obtain a multi-channel two-dimensional time-frequency representation weighted with channel attention weights X cw =X c ×w c +Conv 3×3 (X)。 5. The radio frequency fingerprint recognition method based on multi-channel time-frequency joint attention mechanism according to claim 1 is characterized in that: The multi-channel two-dimensional time-frequency representation is used as the input of the time-frequency joint attention network to obtain a feature tensor weighted by the time-frequency joint attention weight, specifically including: Multi-channel two-dimensional time-frequency representation Perform multi-scale two-dimensional convolution to extract features, N F is the number of Fourier transform points, N T is the number of time windows of the short-time Fourier transform, and we get X m =ReLu(Conv 1×1 (X cw ))+ReLu(Conv 3×3 (X cw ))+ReLu(Conv 5×5 (X cw )) Use N F ×1 and 1×N T The convolution kernel X m Do a two-dimensional convolution to get the frequency attention weight F f and the temporal attention weight F t : Using frequency attention weight F f and the temporal attention weight F t , combined with the automatic broadcast mechanism and dynamic gating mechanism, the time-frequency joint attention weight coefficient is calculated w tf =Sigmoid(F f ×F t ×AvgPool(FC(F f ×F t ))) Using the time-frequency joint attention weight coefficient w tf With multi-channel two-dimensional time-frequency representation X cw Multiply them together to get the feature tensor X weighted by the time-frequency joint attention weight ctf : X ctf =X cw ×w tf 。 6. The radio frequency fingerprint recognition method based on multi-channel time-frequency joint attention mechanism according to claim 1 is characterized in that: The feature tensor is used as the input of the classification network to obtain the prediction probability vector of each category to which the radio frequency signal belongs, specifically including: The feature tensor weighted with time-frequency joint attention weights Feature extraction and classification are performed through a two-dimensional convolutional network, the two-dimensional convolutional network is AlexNet; X ctf After input into AlexNet, the predicted probability vector of the original RF signal belonging to each category is output M is the number of transmitter categories.

7. The radio frequency fingerprint recognition method based on multi-channel time-frequency joint attention mechanism according to claim 1 is characterized in that: The updating of network parameters according to the predicted probability vector specifically includes: Select cross entropy as the loss function and use Adam optimizer to update parameters; the calculation formula of cross entropy is: in, is a one-hot encoding vector, if the RF signal transmitter category is m, then p[m] is 1, otherwise it is 0.

8. The radio frequency fingerprint recognition system based on multi-channel time-frequency joint attention mechanism is characterized by: The system includes the following modules: A receiving module, used for receiving radio frequency signals; The training module is used to train the received radio frequency signal, including: A normalization processing module, used for performing maximum-minimum normalization processing on the received radio frequency signal; A short-time Fourier transform module is used to transform a one-dimensional radio frequency signal into a two-dimensional time-frequency representation; Dynamic channel attention module, which is used to extract dynamic channel attention weight coefficients based on two-dimensional time-frequency representation to obtain multi-channel two-dimensional time-frequency representation weighted with channel attention weights; The time-frequency joint attention module is used to extract the time-frequency joint attention weight coefficient based on the multi-channel two-dimensional time-frequency representation and then obtain the feature tensor weighted by the time-frequency joint attention weight; A probability prediction module is used to use the feature tensor weighted by the time-frequency joint attention weight as the input of the classification network to obtain the prediction probability vector of each category to which the RF signal belongs; An updating module, used for updating network parameters according to the predicted probability vectors of each category to which the radio frequency signal belongs; An iterative training module, used to repeatedly train the received RF signal until the network converges; The output module is used to output the optimal network parameters for the testing process.

9. A computer-readable storage medium storing one or more programs, characterized in that: The one or more programs include instructions which, when executed by a computing device, cause the computing device to perform any one of the methods according to claims 1-7.

10. A computing device, characterized in that: include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs include instructions for executing any of the methods according to claims 1-7.

Images