Radio frequency fingerprint enhancement and reception method, system and device for digital communication system

Abstract

The present invention discloses a radio frequency fingerprint enhancement and receiving method, system and device for a digital communication system. A radio frequency fingerprint neural enhancement unit is added to the transmitting end, and the weight of the radio frequency fingerprint neural enhancement unit is trained by a deep learning method to achieve the effect of enhancing the recognizability of the intrinsic characteristics of the radio frequency signal under a noisy channel. At the same time, a receiving end demodulation network that can automatically match the radio frequency fingerprint neural enhancement unit is constructed to achieve correct demodulation of the received signal.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence application technology, and specifically relates to a radio frequency fingerprint enhancement and receiving method, system and equipment for a digital communication system. Background Art

[0002] With the advent of the Internet of Things era, radio frequency fingerprint recognition has been widely used as a reliable and energy-saving solution to solve problems such as security authentication of wireless devices. Radio frequency fingerprint technology uses the unique hardware defects of the transmitter as the identity of the communication source at the physical layer. The receiver can identify different communication source devices by extracting the characteristic differences of radio frequency fingerprints of different devices. Due to the strict correction mechanism for some disturbances in the transmitter hardware standard design practice, the radio frequency fingerprint itself will be very weak. These weak radio frequency fingerprint features are easily submerged in the channel noise after passing through the channel. At present, deep learning algorithms have a good match with the task of extracting weak radio frequency fingerprint features for identification by mapping the input to a very large feature space, and have shown great potential in the access management and control of wireless devices. However, due to the existence of disturbance factors such as environmental changes and time-varying characteristics of equipment in radio frequency fingerprint recognition, deep learning models are prone to overfitting, learning residuals related to the environment such as channels instead of fingerprint features, and there is a problem of insufficient generalization. Therefore, it is necessary to design a mechanism that can enhance the radio frequency fingerprint characteristics of existing digital communication systems to ensure that the receiver can improve the recognition accuracy of the transmitting device under the premise of normal demodulation of the received signal.

[0003] In the literature “F.Restuccia et al., "DeepRadioID: Real-Time Channel-Re-silientOptimization of Deep Learning-Based Radio Fingerprinting Algorithms", Proc. 20th ACM Int'l. Symp. Mobile Ad Hoc Networking and Computing, pp. 51-60, 2019.”, in order to enhance the RF fingerprint, it is proposed to add a FIR filter at the transmitting end, and use the nonlinear conjugate gradient algorithm to calculate the tap coefficients of the FIR filter to enhance the RF fingerprint. This literature realizes RF fingerprint enhancement based on the FIR filter model. During demodulation, this literature assumes that the receiving end knows the tap coefficients of the transmitting end FIR, and uses DFT to infer the transmitted signal. However, this method is not suitable for actual engineering applications.

[0004] Based on the above technical problems existing in the prior art, the present invention proposes a radio frequency fingerprint enhancement and reception method, system and device for a digital communication system. Summary of the invention

[0005] The purpose of the present invention is to address the deficiencies in the prior art and to provide a method, system and device for radio frequency fingerprint enhancement and reception of a digital communication system. By adding a radio frequency fingerprint neural enhancement unit at the transmitting end and using a deep learning method to train the weight of the radio frequency fingerprint neural enhancement unit, the recognizability of the intrinsic characteristics of the radio frequency signal in a noisy channel can be enhanced. At the same time, a receiving-end demodulation network that can automatically match the radio frequency fingerprint neural enhancement unit is constructed to achieve correct demodulation of the received signal.

[0006] The present invention adopts the following technical solution:

[0007] On the one hand, a radio frequency fingerprint enhancement and reception method for a digital communication system is provided, comprising:

[0008] Step 1, constructing a radio frequency enhancement module including a radio frequency fingerprint neural enhancement unit at the transmitting end, and constructing a demodulation module including a receiving end demodulation network at the receiving end;

[0009] Step 2, constructing a recognition data set for training the RF fingerprint neural enhancement unit and building a receiving end recognition network;

[0010] Step 3, training the radio frequency fingerprint neural enhancement unit;

[0011] Step 4, constructing a receiving data set, a matched filter, a downsampling layer, and demodulation and decision of a signal for training a receiving-end demodulation network;

[0012] Step 5: Train the demodulation network at the receiving end.

[0013] Furthermore, in step 2, constructing a recognition data set for training the radio frequency fingerprint neural enhancement unit includes:

[0014] The transmission signal of the radiation source is collected after modulation, transmission filtering and shaping, and the transmission signal is passed through a Gaussian channel. The received data is intercepted through a sliding window so that the length of each data is N, the data is normalized, and the data is labeled to mark the transmitting device to which the data belongs, and the data is divided into a verification set and a test set.

[0015] Furthermore, in step 2, constructing a receiving end recognition network includes:

[0016] A neural network for signal recognition is trained at the receiving end according to the constructed recognition data set, wherein the recognition data set contains the transmission data of multiple signal sources to be distinguished. During training, the recognition data set is shuffled and sent to the recognition network. After the recognition network is trained, the network parameters are fixed.

[0017] Furthermore, in step 3, the RF fingerprint neural enhancement unit is constructed as a one-dimensional convolutional layer with a length of M. When the signal x with a length of N undergoes a one-dimensional convolution, the internal operation is expressed as:

[0018]

[0019] Among them, the length of the convolution layer M is greater than the number of samples per bit of the signal D.

[0020] Furthermore, in step 4, the received data set is the data that reaches the receiving end after the waveform of the transmitted signal is adjusted by the RF fingerprint neural enhancement unit and channel noise is added to the channel; the collected received data is intercepted using a sliding window to make its length N; and the training set, validation set and test set are divided.

[0021] Furthermore, in step 4, a matched filter is constructed by using a time-delay neural network TDNN, and a one-dimensional convolution layer is constructed as the matched filter part in the neural network structure. The step size is set to 1, and the convolution kernel size, that is, the matched filter order, is set to L. L is greater than the number of samples per bit of the signal D. After passing through the convolution layer, the signal length changes from N to NL. In order to align the processed data with the label, L-1 zeros are added to the end of each data before entering the convolution layer to make its length N+L. After passing through the convolution layer, data with a length of N is obtained.

[0022] Furthermore, in step 4, a downsampling layer is constructed using a time-delay neural network TDNN, which is constructed as a one-dimensional convolutional layer. The step size of the convolution kernel is set to the number of samples D per bit to ensure the alignment condition. The convolution kernel length L is greater than the number of samples D. Before entering the convolution layer, LD zeros are added to the end of each data, thereby ensuring that the data length after the convolution layer is N / D, which is consistent with the corresponding label number.

[0023] Furthermore, in step 4, the demodulation and decision of the signal are constructed using the nonlinear fully connected layer and the softmax activation function.

[0024] Furthermore, in step 5, the loss function is set to a binary cross entropy loss function, and the network parameters are trained by stochastic gradient descent.

[0025] On the other hand, a radio frequency fingerprint enhancement and reception system of a digital communication system is provided, which is used to perform a radio frequency fingerprint enhancement and reception method of a digital communication system, including:

[0026] A transmitter, used to transmit data;

[0027] The RF fingerprint neural enhancement unit is used to train the weights of the RF fingerprint neural enhancement unit to enhance the recognizability of the intrinsic features of the RF signal in a noisy channel;

[0028] A receiving end, used to receive data;

[0029] The demodulation network is used to match the RF fingerprint neural enhancement unit and demodulate the received signal.

[0030] In yet another aspect, a computer device is provided, the computer device comprising a processor and a memory;

[0031] The memory is used to store computer programs; the processor is used to execute the computer programs stored in the memory to implement the radio frequency fingerprint enhancement and reception method of the digital communication system.

[0032] The beneficial effects of the present invention are:

[0033] 1. The radio frequency fingerprint enhancement and receiving method of the digital communication system of the present invention adds a radio frequency fingerprint neural enhancement unit after the transmission filter of the transmitting end, trains the weight of the radio frequency fingerprint neural enhancement unit in a deep learning manner, and can use the radio frequency fingerprint neural enhancement unit at the transmitting end to adjust the waveform of the transmission signal, thereby enhancing the radio frequency fingerprint;

[0034] 2. The RF fingerprint enhancement and reception method of the digital communication system described in the present invention can realize the matched filtering operation of the transmitter filter and the compensation of the newly added RF fingerprint neural enhancement unit by constructing a time delay neural network TDNN at the receiving end, and realize the correct demodulation of the received signal without the need for prior information such as the transmitter filter parameters, and can offset the influence of the transmitter processing on the demodulation. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a schematic diagram of a radio frequency fingerprint enhancement and receiving method of a digital communication system in an embodiment of the present invention;

[0036] Figure 2 A schematic diagram of constructing a transmitting-end radio frequency fingerprint neural enhancement unit in an embodiment of the present invention;

[0037] Figure 3 A schematic diagram of constructing a receiving-end demodulation network in an embodiment of the present invention;

[0038] Figure 4A and Figure 4B A schematic diagram showing a comparison of the recognition confusion matrix of the recognition network under the Gaussian channel condition with a signal-to-noise ratio of 8 dB in an embodiment of the present invention;

[0039] Figure 5 Schematic diagram of bit error rate comparison between neural network demodulation and traditional demodulation method in an embodiment of the present invention. DETAILED DESCRIPTION

[0040] In order to more clearly understand the above-mentioned purposes, features and advantages of the present invention, the present application is further described in detail below in conjunction with the accompanying drawings and specific implementation methods. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0041] Example

[0042] In an embodiment of the present application, the radio frequency fingerprint enhancement and reception method of the digital communication system is implemented by independent modules in the digital domain of the transmitting end and the receiving end respectively. The transmitting end radio frequency enhancement module is completed at the transmitting end, and a radio frequency fingerprint neural enhancement unit (RFF-NEU, Radiofrequency fingerprinting neural enhancement unit) is added after the transmitting filter. The weight of the radio frequency fingerprint neural enhancement unit is obtained by training in a deep learning manner to achieve the function of enhancing the radio frequency fingerprint; the receiving end completes the receiving end demodulation module, and the matched filtering operation of the transmitting filter and the added radio frequency fingerprint neural enhancement unit is completed by constructing a time delay neural network TDNN, so as to achieve correct demodulation of the received signal;

[0043] Specifically, Figure 1-2 As shown, the method includes constructing a transmitting end radio frequency enhancement module:

[0044] S11, constructing a recognition data set suitable for training the RF fingerprint neural enhancement unit, collecting the transmission signal of the radiation source after modulation, transmission filtering and shaping, passing the transmission signal through a Gaussian channel, and then intercepting the received data through a sliding window, so that the length of each data is N, normalizing the data, and labeling the corresponding data, marking the transmitting device to which it belongs, and dividing the training set, verification set and test set;

[0045] S12, constructing a receiving end recognition network, using the data set constructed in S11 to train a neural network for signal recognition at the receiving end, the data set contains the transmission data of multiple signal sources to be distinguished, and the data set is shuffled and sent to the recognition network during training. The recognition network is a multi-classification convolutional neural network, which extracts fingerprint features through the convolution layer and can distinguish the signal as one of the C-class signals with a high accuracy. After the recognition network is trained, the network parameters are fixed;

[0046] S13, train the RF fingerprint neural enhancement unit for a specific device. To enhance the RF fingerprint of a specific device A and offset the weakening of the RF fingerprint by the channel, add a RF fingerprint neural enhancement unit after the transmit filter. The RF fingerprint neural enhancement unit is constructed as a one-dimensional convolution layer with a length of M. When a signal x with a length of N passes through this one-dimensional convolution, the internal operation is expressed as:

[0047]

[0048] The length M of the convolution layer is greater than the number of samples D per bit of the signal, so that the convolution can extract fingerprint features according to different bits. For example, M=3D+1 is set; the weight parameters to be learned are initialized, any one of the parameters is set to 1, and the other parameters are set to 0. Finally, according to the data set constructed in S11, the sub-data set consisting of the data sent by device A is selected. As a new data set, the data in the new data set is input into the RF fingerprint neural enhancement unit, and then enters the fixed parameter recognition network in S12 after passing through the Gaussian channel corresponding to device A and normalization processing. The one-dimensional convolution parameters are updated by back propagation through random gradient descent to maximize the probability that the recognition network predicts the result of A. When the recognition rate of the network reaches 99%, the one-dimensional convolution weight is fixed and derived, and the RF fingerprint neural enhancement unit of device A is obtained;

[0049] like Figure 3 As shown, the method includes constructing a receiving end demodulation network including:

[0050] S21, construct a receiving data set suitable for training a demodulation network. The data set consists of receiving data, that is, the data that reaches the receiving end after the waveform of the sending signal is adjusted by the RF fingerprint neural enhancement unit and the channel noise is added to the channel. The collected receiving data is intercepted by a sliding window to make its length N. Since the sent digital signal is a 01 bit stream, in order to demodulate correctly, the network needs to perform a binary classification on each bit. Its data label is 0 or 1, which is the information bit that has not been sent through filtering, shaping and modulation. The training set, validation set and test set are divided;

[0051] S22, using TDNN to construct a matched filter. The time-delay neural network TDNN takes a signal on a time window as input. The input of each layer is the output of the previous layer of the network at different times, which is used to approximate the matched filter in the communication system. Exemplarily, in the specific implementation, TDNN is realized by controlling the convolution kernel size and step size of the convolution layer. Therefore, a one-dimensional convolution layer is constructed as the matched filter part in the neural network structure, and its step size is set to 1. The convolution kernel size, that is, the order of the matched filter is set to L, and L must be greater than the number of samples per bit of the signal D. After passing through the convolution layer, the signal length changes from N to NL. In order to align the processed data with the label, L-1 zeros are added to the end of each data before entering the layer to make its length N+L. In this way, data with a length of N can be obtained after passing through the convolution layer;

[0052] S23, using TDNN to construct a downsampling layer. Since the signal will be upsampled at the transmitting end, a downsampling operation is required at the receiving end for correct demodulation. The downsampling layer is also a TDNN, constructed as a one-dimensional convolution layer. The step size of the convolution kernel is set to the number of samples per bit D to ensure the alignment condition. The convolution kernel length L of the sampling layer must also be greater than the number of samples D. In this way, other code element information is used to correct the current information. Before entering the sampling layer, LD zeros are added to the end of each data to ensure that the data length after the sampling layer is N / D, which is consistent with the corresponding label number;

[0053] S24, using nonlinear fully connected layer and softmax activation function to build signal demodulation and judgment, the demodulation and judgment of the signal are realized by fitting the nonlinear fully connected layer. The demodulation part is composed of two fully connected layers with activation function as ReLu, and the judgment part is a fully connected layer with activation function as softmax. The neurons in the fully connected layer express 0 and 1 differently, so as to output the predicted value of the signal bit. The demodulation and judgment network structure settings refer to Table 1;

[0054] Table 1

[0055]

[0056]

[0057] S25, training of the entire demodulation network, connects the network layers constructed in S22-S24 respectively, trains the entire demodulation network, sets the loss function as a binary cross entropy loss function, and obtains the network parameters through stochastic gradient descent training. The accuracy of the network output is the accuracy of the bit demodulation.

[0058] The following example takes the case where the transmitted signal comes from three radiation sources, where the three radiation sources are numbered 0, 1, and 2, the transmitted signal symbol frequency is 1 MHz, the sampling frequency is 20 MHz, and the number of samples per bit is D=20. After I / Q modulation and root raised cosine filtering, the signal is added to the carrier frequency and sent to the channel.

[0059] First, build the transmitter RF fingerprint neural enhancement unit;

[0060] S1A, constructs a recognition data set suitable for training RF fingerprint neural enhancement units, collects the transmission signal of the radiation source, passes it through a Gaussian channel with a signal-to-noise ratio of 20dB, and then intercepts the data through a sliding window with a window length of N=512 and a sliding step size of K=64, so that the length of each sample is 512, and each sample is labeled 0, 1, 2 according to the radiation source from which the sample comes. The final data set contains a total of 187476 data. The amount of data in the three categories is the same, each accounting for 1 / 3 of the total data. After normalization, it is divided into training sets Validation set and test set There are 131,232 training data, 28,122 validation data and 28,122 test data respectively (training set: validation set: test set = 0.7:0.15:0.15);

[0061] S1B, build the receiving end recognition network, shuffle the data set in S1A, and send the shuffled data to the recognition network. The recognition network structure and parameter settings are shown in Table 2. In the model training, cross entropy is selected as the loss function, Adam is used as the optimizer, Batch Size is 256, epoch is 15, the learning rate of the model is 0.0, the recognition model converges, and the recognition accuracy is 97.5%;

[0062] Table 2

[0063]

[0064]

[0065] S1C, train the RF fingerprint neural enhancement unit for a specific device, first train and enhance the RF fingerprint of device 0, add a RF fingerprint neural enhancement unit after its transmission filter, construct the RF fingerprint neural enhancement unit as a one-dimensional convolution layer with a convolution kernel length of M = 3D + 1 = 61, and then initialize its parameters, that is, the weight parameters to be learned, set any one of the parameters to 1, and set the other parameters to 0, and finally select the sub-dataset consisting of the transmission data of device 0 according to the data set constructed in S1A As a new data set, the data in the new data set is input into the RF fingerprint neural enhancement unit, and then enters the fixed parameter recognition network in S1B after passing through a Gaussian channel with a signal-to-noise ratio of 8dB and normalization. Adam is used as the optimizer, and the one-dimensional convolution parameters are updated by back propagation in a random gradient descent manner to maximize the probability that the recognition network predicts that the result is device 0, thereby determining the parameter value of the RF fingerprint neural enhancement unit of the 0th type of device. When the recognition rate of the network reaches 99.2%, the training is stopped, and the one-dimensional convolution weights such as φ0 are derived to obtain the RF fingerprint neural enhancement unit of device 0. The training process of the RF fingerprint neural enhancement unit of the first and second type of radiation sources is the same, where φ0=[1.22 21e-01,4.0036e-02,4.7016e-03,-8.0321e-03,-1.1320e-02,-9.70 62e-03,-4.8260e-03,3.4176e-03,1.8396e-02,5.1845e-02,7.2049e -02,4.2256e-02,2.5489e-02,1.4785e-02,6.1415e-03,-2.0669e-03 ,-1.1447e-02,-2.4424e-02,-4.7324e-02,-1.0601e-01,3.9775e-01 ,-4.6033e-02,-2.4929e-02,-1.4499e-02,-8.7318e-03,-5.6794e- 03,-4.9563e-03,-5.8135e-03,-8.7472e-03,-1.7281e-02,2.8197e- 05,6.2824e-03,1.0606e-02,1.5052e-02,2.0136e-02,2.6316e-02, 3.4830e-02,4.7800e-02,7.1148e-02,1.2550e-01,7.0757e-02,1.87 47e-02,-3.3914e-03,-1.6170e-02,-2.5517e-02,-3.3519e-02,-4. 0964e-02,-4.9469e-02,-6.3069e-02,-8.8618e-02,-4.1982e-02,-1 .3186e-02,-5.3584e-03,-4.8632e-03,-8.0319e-03,-1.3479e-02,- 2.1835e-02,-3.1336e-02,-3.8222e-02,-1.3285e-02,2.8069e-02];

[0066] The 0th type of signal passes through a channel with a signal-to-noise ratio of 8dB, the first type of signal passes through a channel with a signal-to-noise ratio of 10dB, and the second type of signal passes through a channel with a signal-to-noise ratio of 7dB. After the signals pass through their respective RF fingerprint neural enhancement units, the recognition network is tested. The three signals are tested under a signal-to-noise ratio of 8dB, and the following results can be obtained: Figure 4A and 4B The effect shown, Figure 4A Displays the recognition confusion matrix of the recognition network without the RF fingerprint neural enhancement unit. Figure 4B Shows the recognition confusion matrix of the recognition network after adding the RF fingerprint neural enhancement unit, compared Figure 4A and Figure 4B ,The RF fingerprint neural enhancement unit improves the recognition accuracy.,Furthermore, the influence of the RF fingerprint neural enhancement unit on the recognition,network under Gaussian channel conditions with different signal-to-noise ratios is tested,,and the influence results of RFF-NEU on the recognition accuracy under Gaussian channel conditions with,different signal-to-noise ratios are obtained as shown in Table 3.,The results show that the addition of the RF fingerprint neural enhancement unit has an,enhanced effect on the RF fingerprint under various Gaussian channel conditions;

[0067] Table 3

[0068]

[0069] Then construct the demodulation network at the receiving end;

[0070] S2A, construct a receiving data set suitable for training a demodulation network. Here, the demodulation network for the Class 0 signal is taken as an example. The demodulation network construction process for the Class 1 signal and the Class 2 signal is the same. The data set of the demodulation network at the receiving end comes from the data that reaches the receiving end after the waveform is adjusted by the Class 0 RF fingerprint neural enhancement unit and the signal-to-noise ratio is 15dB. A sliding window with a window length of N = 40000 and a sliding step size of K = 200 is used to intercept the data. Therefore, each signal to be processed contains N / D = 200 bits, and each bit corresponds to a label, that is, 0 or 1. When processing the original bit stream as a label, a sliding window with a window length of N / D = 200 and a step size of K / D = 10 is used to ensure the alignment of the data and the label. After the data is intercepted, the training set, validation set, and test set are divided;

[0071] S2B, using TDNN to build a matched filter, construct a one-dimensional convolution layer with a convolution kernel size of L = 81 and a stride of 1 as the matched filter part of the neural network structure, and add 80 zeros at the end of each data before entering this layer;

[0072] S2C, TDNN is used to construct the downsampling layer. The downsampling layer is also constructed as a one-dimensional convolutional layer with a convolution kernel size of L = 80 and a step size of D = 20, and LD = 60 zeros are added to the end of each data before entering the convolutional layer to ensure that the data length after the convolutional layer is N / D = 200, which is consistent with the number of labels in the batch;

[0073] S2D, uses nonlinear fully connected layers and softmax activation functions to construct signal demodulation and judgment. Signal demodulation and judgment can be achieved through nonlinear fully connected layer fitting. The demodulation part consists of two fully connected layers with ReLu activation function, and the judgment part is a fully connected layer with softmax activation function and two neurons. The two neurons express 0 and 1 differently and output the predicted value of the signal bit. The demodulation and judgment network structure and parameter settings refer to Table 4;

[0074] Table 4

[0075] Network layer name Number of neurons Activation Function Decoder_layer_1 128 ReLu Decoder_layer_2 64 ReLu Decider 2 Softmax

[0076] S2E, training of the entire demodulation network, connects the network layers constructed in steps S2B-S2D respectively, trains the entire demodulation network, sets the loss function to a binary cross entropy loss function, and obtains the network parameters through stochastic gradient descent training. The accuracy of the network output is the accuracy of the bit demodulation.

[0077] In order to further verify the technical effect of the method described in the embodiment of the present application, Figure 5 The bit error rates of neural network demodulation and traditional demodulation methods are given when the RFF-NEU parameters are unknown and the blue line represents neural network demodulation, and the yellow line represents traditional demodulation. The neural network can still obtain a high demodulation accuracy under the influence of RFF-NEU parameters, while the traditional demodulation can hardly work properly in this case. For comparison, Figure 5 The bit error rate of traditional demodulation without adding RFF-NEU parameters is also given, which shows that when the RFF-NEU parameters are not added, the traditional demodulation system can achieve a relatively high demodulation level, but after the influence of RFF-NEU parameters is applied, the traditional demodulation method is no longer reliable. Although the existence of RFF-NEU parameters makes the demodulation effect of the neural network also worse than the traditional demodulation without adding RFF-NEU parameters, the application of neural networks enables the entire communication link to operate normally, improving the robustness of the system that can both meet the requirements of RF radiation source identification and normal communication.

[0078] The radio frequency fingerprint enhancement and receiving system of the digital communication system comprises:

[0079] A transmitter, used to transmit data;

[0080] The RF fingerprint neural enhancement unit is used to train the weights of the RF fingerprint neural enhancement unit to enhance the recognizability of the intrinsic features of the RF signal in a noisy channel;

[0081] A receiving end, used to receive data;

[0082] The demodulation network is used to match the RF fingerprint neural enhancement unit and demodulate the received signal.

[0083] The computer device comprises a processor and a memory;

[0084] The memory is used to store computer programs; the processor is used to execute the computer programs stored in the memory to implement the described method.

[0085] The present invention is not limited by the above embodiments. The above embodiments and descriptions are only for explaining the principle of the present invention. Without departing from the spirit and scope of the present invention, the present invention may be subject to various changes and improvements, which are within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the appended claims.

Claims

1. A radio frequency fingerprint enhancement and reception method for a digital communication system, characterized in that: include: Step 1, constructing a radio frequency enhancement module including a radio frequency fingerprint neural enhancement unit at the transmitting end, and constructing a demodulation module including a receiving end demodulation network at the receiving end; Step 2, constructing a recognition data set for training the RF fingerprint neural enhancement unit and building a receiving end recognition network; Step 3, train the RF fingerprint neural enhancement unit: The RF fingerprint neural enhancement unit is constructed as a one-dimensional convolutional layer with a length of M. When the signal x with a length of N undergoes a one-dimensional convolution, the internal operation is expressed as: Among them, the length of the convolution layer M is greater than the number of samples per bit of the signal D; Step 4, construct a receiving data set, a matched filter, a downsampling layer, and signal demodulation and judgment for training the receiving end demodulation network; construct a matched filter by using the time delay neural network TDNN, construct a one-dimensional convolution layer as the matched filter part of the neural network structure, set the step size to 1, and set the convolution kernel size, that is, the matched filter order, to L, where L is greater than the number of samples per bit of the signal D. After passing through the convolution layer, the signal length changes from N to NL. In order to align the processed data with the label, L-1 zeros are added to the end of each data before entering the convolution layer to make its length N+L. After passing through the convolution layer, data with a length of N is obtained; Step 5: Train the demodulation network at the receiving end.

2. The radio frequency fingerprint enhancement and reception method of a digital communication system according to claim 1, characterized in that: In step 2, constructing a recognition data set suitable for training the RF fingerprint neural enhancement unit includes: The transmission signal of the radiation source is collected after modulation, transmission filtering and shaping, and the transmission signal is passed through a Gaussian channel. The received data is intercepted through a sliding window so that the length of each data is N, the data is normalized, and the data is labeled to mark the transmitting device to which the data belongs, and the data is divided into a verification set and a test set.

3. The radio frequency fingerprint enhancement and reception method of a digital communication system according to claim 1, characterized in that: In step 2, building a receiving end recognition network includes: A neural network for signal recognition is trained at the receiving end according to the constructed recognition data set, wherein the recognition data set contains the transmission data of multiple signal sources to be distinguished. During training, the recognition data set is shuffled and sent to the recognition network. After the recognition network is trained, the network parameters are fixed.

4. The radio frequency fingerprint enhancement and reception method of a digital communication system according to claim 1, characterized in that: In step 4, the received data set is the data that reaches the receiving end after the waveform of the transmitted signal is adjusted by the RF fingerprint neural enhancement unit and channel noise is added to the channel; the collected received data is intercepted using a sliding window to make its length N; and the training set, validation set and test set are divided.

5. The radio frequency fingerprint enhancement and reception method of a digital communication system according to claim 1, characterized in that: In step 4, a downsampling layer is constructed using a time-delay neural network TDNN, which is constructed as a one-dimensional convolutional layer. The step size of the convolution kernel is set to the number of samples per bit D to ensure the alignment condition. The convolution kernel length L is greater than the number of samples D. Before entering the convolution layer, LD zeros are added to the end of each data to ensure that the data length after passing through the convolution layer is N / D, which is consistent with the corresponding label number.

6. The radio frequency fingerprint enhancement and reception method of a digital communication system according to claim 1, characterized in that: In step 4, the nonlinear fully connected layer and softmax activation function are used to construct the demodulation and judgment of the signal.

7. The radio frequency fingerprint enhancement and reception method of a digital communication system according to claim 1, characterized in that: In step 5, the loss function is set to a binary cross entropy loss function, and the network parameters are trained by stochastic gradient descent.

8. A radio frequency fingerprint enhancement and receiving system for a digital communication system, used to execute the method according to any one of claims 1 to 7, characterized in that: include: A transmitter, used to transmit data; The RF fingerprint neural enhancement unit is used to train the weights of the RF fingerprint neural enhancement unit to enhance the recognizability of the intrinsic features of the RF signal in a noisy channel; A receiving end, used to receive data; The demodulation network is used to match the RF fingerprint neural enhancement unit and demodulate the received signal.

Images