A small sample radio frequency fingerprint intelligent recognition system and method

Abstract

The present invention discloses a small sample radio frequency fingerprint intelligent recognition system and method, which solves the problems that the existing radio frequency fingerprint recognition method relies on a large number of training samples, the recognition is easily disturbed and has low accuracy. Through the data acquisition module, data preprocessing module, data partitioning module, data training module, small sample learning module and data output module, the classification result is finally output. The present invention learns the semantic relationship between classes and maps it into the semantic space. Through the twin network, the accuracy of recognition when the number of samples is small is greatly improved, the interference in the recognition process is reduced, and a high accuracy rate can still be maintained under low signal-to-noise ratio conditions. The present invention uses transfer learning to significantly reduce the dependence on the amount of training samples; the twin network integrates the feature information between different levels, further improving the accuracy of recognition.

Description

Technical Field

[0001] The present invention belongs to the field of wireless communications, and in particular relates to an intelligent recognition system and method for radio frequency fingerprints in a small sample environment. Background Art

[0002] In recent years, with the maturity of 5G technology, research on 6G wireless communication networks has emerged. As a key component of 6G communication networks, intelligent communication technology urgently needs in-depth research. Due to the openness of wireless networks, the risk of illegal user access and large-scale malicious attacks on various wireless devices has increased, and it has become one of the factors that seriously hinder the development and application of wireless network technology. In the process of identifying RF fingerprints, transient signals have a large amount of subtle feature information, which is very suitable for device identification. At the same time, the main content of steady-state signals is communication information data doped with channel noise. This part of the wireless signal can also be extracted as a series of device-specific information as device features. However, whether it is transient features or steady-state features, the number of features that can be artificially extracted is often limited. It also depends on data denoising, signal transformation, and finally feature extraction in the transform domain. However, this method is difficult to extract the optimal features. Feature extraction is the key to RF fingerprint recognition, and its quality directly affects the accuracy of signal recognition. And this method is generally only for a few limited electromagnetic signals and specific environments. Deep neural networks, as a method of autonomously learning different levels of features from data, have been widely used in different fields in recent years. RF fingerprint intelligent recognition technology is one of the key technologies in 6G wireless communication. Therefore, it is very important to carry out research on radio frequency fingerprint intelligent recognition.

[0003] In the RF fingerprint intelligent recognition technology, people close the loop of feature extraction and classification recognition, directly input I / Q data into the network, and achieve more accurate end-to-end individual recognition. M. Ezuma, F. Erden et al. in their paper "Micro-UAV Detection and Classification from RF Fingerprints Using Machine Learning Techniques" (2019 IEEE Aerospace Conference, pp. 1-13, 2019) can identify a variety of signals by analyzing the transient changes of the signal, but the complex preprocessing process makes it difficult to apply. S. Gopalakrishnan et al. in their paper "Robust Wireless Fingerprinting via Complex-Valued Neural Networks" (2019 IEEE Global Communications Conference (GLOBECOM), pp. 1-6, 2019) enhance the data by adding noise, thereby improving the recognition accuracy, but it relies on a large amount of training data. Ender Ozturk et al. proposed a low-SNR UAV signal recognition algorithm in their paper "RF-Based Low-SNR Classification of UAVs Using Convolutional Neural Networks" (arXiv preprint arXiv:2009.05519, 2020.), which can achieve 92% recognition accuracy at -10dB SNR. However, the improvement of RF fingerprint recognition accuracy is still limited when there are fewer samples.

[0004] The above method not only requires a lot of cost to collect relevant signal samples and mark data, but also when new devices are added, the applied model needs to re-collect data and train. The large amount of cost consumption in practical applications makes this technology still remain at the theoretical level. In addition, in actual situations, data-rich categories usually only account for a small part of all data categories. Most data categories have only a few data samples, and even some rare categories have almost no data samples. Finally, the current RF fingerprint recognition method has low recognition accuracy under low signal-to-noise ratio conditions, making it difficult to apply in actual complex communication networks. Therefore, it is urgent to study new RF fingerprint intelligent recognition methods. Summary of the invention

[0005] The present invention provides a small sample radio frequency fingerprint intelligent recognition system and method, which greatly improves the recognition accuracy when the number of samples is small by learning the semantic relationship between classes and mapping them into the semantic space, reduces the interference in the recognition process, and can still maintain a high accuracy under low signal-to-noise ratio conditions through a twin network.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] The present invention provides a small sample radio frequency fingerprint intelligent recognition system, including a data acquisition module, a data preprocessing module, a data division module, a data training module, a small sample learning module and a data output module, wherein:

[0008] A data acquisition module, used for collecting radio frequency signals;

[0009] The data preprocessing module is used to process the noise of the signal from the data acquisition module and convert it into a time-frequency diagram;

[0010] A data classification module is used to classify the time-frequency diagram output by the data preprocessing module;

[0011] A data training module is used to perform data training of a radio frequency fingerprint classification network, a radio frequency fingerprint feature extraction network, and a radio frequency fingerprint feature mapping network on the classified data;

[0012] The small sample learning module builds a machine learning model classifier based on the above-mentioned trained RF fingerprint classification network, RF fingerprint feature extraction network, and RF fingerprint feature mapping network, performs small sample data learning, and selects the optimal classifier;

[0013] The data output module is used to input test data and output classification results.

[0014] The present invention also provides a small sample radio frequency fingerprint intelligent identification method implemented by the above small sample radio frequency fingerprint intelligent identification system, comprising the following steps:

[0015] Step 1: Collect the radio frequency signal in the data acquisition module, and express the collected signal in the form of a vector, where I is the in-phase component and Q is the orthogonal component;

[0016] Step 2: In the data preprocessing module, a corresponding category label file is created according to the collected radio frequency signal, and noise is added to the radio frequency signal;

[0017] Step 3, preprocessing the I / Q data obtained in step 1 and step 2, setting the segmentation point after normalization, and converting the segmentation result into a time-frequency diagram;

[0018] Step 4: In the data partitioning module, the time-frequency graph obtained in step 3 is classified into categories to obtain a basic data set and a new class data set. The basic data set and the new class data set are further divided into a support set and a test set respectively.

[0019] Step 5: In the data training module, a radio frequency fingerprint classification network is constructed and pre-trained, and the support set data of the basic data set obtained in step 4 is input into the network and trained to obtain a radio frequency fingerprint classification pre-training model;

[0020] Step 6, reconstructing the radio frequency fingerprint classification pre-training model obtained in step 5, and obtaining the radio frequency fingerprint feature extraction network using the output parameters in step 5;

[0021] Step 7: construct a twin network training model, wherein the twin network training model is composed of a radio frequency fingerprint feature mapping network, and the features obtained in step 6 are input into the radio frequency fingerprint feature mapping network for training to obtain semantic space information of the radio frequency signal;

[0022] Step 8, determining whether the RF fingerprint feature mapping network training is completed, if so, executing step 9, if not, increasing the number of training iterations by one and continuing to train the RF fingerprint feature mapping network;

[0023] Step 9: In the small sample learning module, a machine learning model classifier is constructed, and the results obtained after a small number of samples are input into the machine learning model classifier in sequence in step 6 and step 7, so as to select the optimal classifier;

[0024] Step 10, in the data output module, input the test set in the new class data set in step 4 into step 6, step 7 and step 9 in sequence;

[0025] Step 11: Output the classification results.

[0026] Furthermore, the RF fingerprint recognition in step 1 is regarded as a K-type hypothesis testing problem. Assume that the received signal of the k-th RF signal is x k (i) = s k (i)+ω k (i) Where s k (i) represents the i-th sampling point of the k-th RF transmission signal, x k (i) represents the i-th sampling point of the k-th RF received signal, ω k (i) means the mean is 0 and the variance is σ 2 The signal received by the receiving end is expressed as an I / Q signal in vector form, x k =I k +Q k , where x k Represents x k The vector form of (i), Ik and Q k Represent the in-phase and quadrature components of the signal respectively.

[0027] Furthermore, in step 2, noise is added to the interference-free data, and the power of the interference-free signal is expressed as:

[0028]

[0029] Where N represents the number of sampling points of the RF received signal. Assuming the signal-to-noise ratio of the required signal is SNR, the noise power that needs to be generated is:

[0030] P noise [dB]=P signal [dB]-SNR

[0031] The final signal generated is

[0032] s k [i] = x k (i)+n k [i]

[0033] where s k [i] represents the final generated signal, n k [i] indicates power P noise [dB] Generated additive white Gaussian noise;

[0034] Furthermore, in step 3, after obtaining the data processed in step 1 and step 2, the segmentation point is determined according to the number of data sampling points, the data is segmented into multiple samples, the mean and variance of the data are calculated, the data is normalized, and a single sample is converted into a time-frequency diagram;

[0035] Furthermore, in step 4, the data and categories of the basic data set and the new class data set are disjoint, and the basic data set and the new class data set are further divided into a support set and a test set, respectively. The data of the support set and the test set are disjoint, but have common categories.

[0036] Furthermore, in step five, the constructed radio frequency fingerprint classification network is an improvement on the existing residual neural network. By changing the number of output categories of the last fully connected layer of the network, it is suitable for the current sample space. The complete radio frequency fingerprint training network consists of four residual units, and the residual unit consists of two convolutional layers with a kernel size of 3×3. The batch normalization layer standardizes the intermediate data in the middle layer of the network to avoid the gradient vanishing problem caused by the saturation of the partial derivative of the intermediate variable. The residual stacking unit is constructed by sequentially connecting a 1×1 convolutional layer and a batch normalization layer. Finally, a linear correction unit is connected after each batch normalization layer as an activation function to introduce nonlinearity into the network. When pre-training the radio frequency fingerprint classification network, the trainable parameters of the network are randomly initialized, the number of network training epochs is initialized to 1, the maximum number of epochs is 50, the learning rate is 0.001, and the stochastic gradient descent optimization algorithm is used as the network training optimizer. The cross entropy loss function is selected to calculate the gap between the network output and the category, and it is provided to the network training optimizer for optimization to obtain the radio frequency fingerprint classification pre-training model.

[0037] Furthermore, in step six, the features output by different layers of the RF fingerprint classification pre-training model represent feature information at different levels. The new RF fingerprint classification pre-training model is reloaded and the optimal parameters obtained in step five are loaded. The output parameters of the network are set to the output results of each residual unit. The output dimensions are 64×56×56, 128×28×28, 256×14×14 and 512×7×7 respectively. The number of output channels of the front output layer is small, indicating that the sample has relatively comprehensive features. The output channels of the deep network are large, indicating the detailed features of the RF signal at different levels, and the fingerprint information of the RF signal can be clearly mapped to the feature space.

[0038] Furthermore, in step seven, the RF fingerprint feature mapping network is trained by the method of training twin networks, and the features learned by the RF fingerprint feature extraction network are mapped to a high-dimensional semantic space. The similarity between the two samples is evaluated by calculating the Euclidean distance between the two input samples. The RF fingerprint feature mapping network is composed of six basic residual triangular units, each of which inputs high-level features and low-level features of two different dimensions. For the high-level features, two convolutional layers with a kernel size of 3×3 and strides of 2 and 1 are connected. The convolutional layer with a stride of 2 reduces the size of the data so that it matches the low-level features. Each convolutional layer is connected by a batch normalization layer. Then, to prevent overfitting of the data, the activation function is used to make the network nonlinear. The residual stacking unit is constructed by sequentially connecting a convolutional layer with a size of 1×1 and a stride of 2 and a batch normalization layer. For low-level features, it is composed of two convolutional layers with a kernel size of 3×3 and a stride of 1. The batch normalization layer is connected after the convolution layer. Finally, the activation function is used to make the network nonlinear. The processed high-level features and low-level features are summed and output after being processed by the basic residual unit. The residual unit here is consistent with the residual unit in step 6. The features finally output by the radio frequency fingerprint feature mapping network are mapped to a 512-dimensional semantic space after the average pooling layer and the linear regression layer.

[0039] Furthermore, in step nine, a small number of samples are support set data in the new class data set. According to different tasks, a small number of samples are input into the optimal model obtained in steps six and seven, and finally the semantic information of each dimension of the samples is obtained. A machine learning model classifier is constructed, and the semantic information is input into the classifier for training. After multiple rounds, the optimal classifier is screened out and saved.

[0040] The beneficial effects of the present invention are:

[0041] 1. The present invention learns the semantic relationship between classes and maps it into the semantic space. Through the twin network, it greatly improves the recognition accuracy when the number of samples is small, reduces the interference in the recognition process, and can still maintain a high accuracy under low signal-to-noise ratio conditions.

[0042] 2. The features of the RF signal are learned using the feature extraction network and the feature mapping network. Compared with the traditional deep learning framework, the present invention uses two different networks and integrates different features, solving the problem of low recognition performance caused by the traditional method of using only a single network and a single feature for recognition.

[0043] 3. Taking advantage of the unique advantages of the twin network, compared to the traditional deep learning framework that requires a large amount of data as training samples in a new environment, the present invention only requires about 5% of the training samples of the traditional method, greatly reducing the dependence on the number of training samples.

[0044] 4. The present invention generalizes the semantic relationship between categories to the semantic space, which solves the problem that the traditional method has unsatisfactory recognition effect under the presence of interference and low signal-to-noise ratio, and significantly improves the recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a flow chart of the present invention;

[0046] Figure 2 It is the overall block diagram of the twin network of the present invention;

[0047] Figure 3 It is a schematic diagram of the radio frequency fingerprint feature extraction and radio frequency fingerprint feature mapping network of the present invention;

[0048] Figure 4 It is a schematic diagram of a basic unit of a radio frequency fingerprint feature mapping network using the present invention;

[0049] Figure 5 This is a comparison chart of classification accuracy using the present invention and other existing technologies under different test methods;

[0050] Figure 6 This is a comparison diagram of the convergence speed of training under different interference levels using the present invention;

[0051] Figure 7 This is a comparison chart of classification accuracy under different signal-to-noise ratios using the present invention. DETAILED DESCRIPTION

[0052] The present invention will be further described below in conjunction with the accompanying drawings.

[0053] Combined with Figure 1 The specific steps of the method of the present invention are described as follows.

[0054] Step 1: RF fingerprint data collection.

[0055] RF fingerprint recognition can be regarded as a K-type hypothesis testing problem. Assume that the received signal of the k-th RF signal is x k (i) = s k (i)+ω k (i) Where s k (i) represents the i-th sampling point of the k-th RF transmission signal, x k (i) represents the i-th sampling point of the k-th RF received signal, ω k (i) means the mean is 0 and the variance is σ 2 The signal received by the receiver is represented as an I / Q signal in vector form, x k =I k +Q k , where xk Represents x k The vector form of (i), I k and Q k Represent the in-phase and quadrature components of the signal respectively.

[0056] Step 2: Create a corresponding category label file based on the collected RF signal and add noise to the RF signal.

[0057] In the present invention, seven types of UAV radio frequency signals are used for training and testing. The UAV states are open, hovering and flying. Each state has signals in four conditions: no interference, wireless network interference, Bluetooth interference, and wireless network and Bluetooth interference. Noise is added to the data without interference. The power of the signal without interference is expressed as:

[0058]

[0059] Where N represents the number of sampling points of the RF received signal. Assuming the signal-to-noise ratio of the required signal is SNR, the noise power that needs to be generated is:

[0060] P noise [dB]=P signal [dB]-SNR

[0061] The final signal generated is

[0062] s k [i] = x k (i)+n k [i]

[0063] where s k [i] represents the final generated signal, n k [i] indicates power P noise [dB] Generated additive white Gaussian noise.

[0064] Step 3, preprocess the I / Q data obtained in steps (1) and (2), set the segmentation point after normalization, and convert the segmentation result into a time-frequency diagram.

[0065] After obtaining the data processed by step (1) and step (2), the segmentation point is determined according to the number of data sampling points, and the data is segmented into multiple samples. The mean and variance of the data are calculated, the data is normalized, and a single sample is converted into a time-frequency diagram to provide it to the RF fingerprint feature extraction network and the RF fingerprint feature mapping network for training.

[0066] Step 4: classify the time-frequency diagram obtained in step (3) to obtain a basic data set and a new class data set.

[0067] The training of the RF fingerprint feature extraction network and the RF fingerprint feature mapping network is completed in the basic data set. The data and categories of the basic data set and the new class data set are disjoint. The basic data set and the new class data set are further divided into a support set and a test set. The data of the support set and the test set are disjoint but have common categories.

[0068] Step 5: construct and pre-train a radio frequency fingerprint classification network, input the support set data obtained in step (4) into the network and perform training to obtain a radio frequency fingerprint classification pre-training model.

[0069] The pre-trained radio frequency fingerprint classification network constructed by the present invention is obtained by improving the existing residual neural network. By changing the number of output categories of the last fully connected layer of the residual neural network, it is suitable for the current sample space. The complete radio frequency fingerprint training network is composed of four residual units, and the residual unit is composed of two convolutional layers with a kernel size of 3×3. The batch normalization layer standardizes the intermediate data in the middle layer of the network to avoid the gradient vanishing problem caused by the saturation of the partial derivatives of the intermediate variables. The residual stacking unit is constructed by sequentially connecting a 1×1 convolutional layer and a batch normalization layer. Finally, a linear correction unit is connected after each batch normalization layer as an activation function to introduce nonlinearity into the network.

[0070] When pre-training the RF fingerprint classification network, the trainable parameters of the network are randomly initialized, the number of network training epochs is initialized to 1, the maximum number of epochs is 50, the learning rate is 0.001, and the stochastic gradient descent (SGD) optimization algorithm is used as the network training optimizer. The cross entropy loss function is used to calculate the gap between the network output and the category.

[0071] Step 6, reconstruct the RF fingerprint classification pre-training model obtained in step (5), and use the output parameters of different layers to obtain the RF fingerprint feature extraction network.

[0072] The features output by different layers of the obtained pre-trained model represent feature information at different levels. By reloading the optimal model, the model outputs feature information extracted from four different layers. The higher layers have fewer output channels, which can represent relatively comprehensive features of the samples. The lower layers have many output channels, and each channel can represent detailed features at different levels. In the lower layers, the fingerprint information of the RF signal can be mapped more clearly.

[0073] Step 7, construct a twin network training model composed of a radio frequency fingerprint feature mapping network, input the features obtained in step (6) into the radio frequency fingerprint feature mapping network for training, and obtain the semantic space information of the radio frequency signal.

[0074] The RF fingerprint feature mapping network is trained by the method of training twin networks, mapping the features learned by the RF fingerprint feature extraction network to a high-dimensional semantic space, and finally evaluating the similarity between two samples by calculating the Euclidean distance between the two input samples. The feature mapping network consists of six basic residual triangular units. Figure 4 As shown, each triangular basic unit inputs high-level features and low-level features of two different dimensions respectively.

[0075] For high-level features, two convolutional layers with a kernel size of 3×3 and strides of 2 and 1 are connected. The convolutional layer with a stride of 2 reduces the size of the data so that it matches the low-level features. Each convolutional layer is connected by a batch normalization layer to prevent overfitting of the data. Finally, the activation function is used to make the network nonlinear. The residual stacking unit is constructed by a convolutional layer with a size of 1×1 and a stride of 2 and a batch normalization layer connected in sequence.

[0076] For low-level features, it consists of two convolutional layers with a kernel size of 3×3 and a step size of 1. The convolutional layer is connected to a batch normalization layer, and finally the activation function is used to make the network nonlinear.

[0077] The processed high-level features and low-level features are summed and output after being processed by the basic residual unit. The residual unit here is the same as the residual unit in step (6). The features finally output by the feature mapping network are mapped to a 512-dimensional semantic space after the average pooling layer and the linear regression layer. The feature mapping network summarizes the global feature information, so it is more robust to interference noise.

[0078] When training the feature mapping network, all training samples are divided into sample pairs. When the sample pairs come from the same category, they are marked as "1", and when they come from different categories, they are marked as "0". The probability of marking "1" and "0" is equal, and the mark is represented as Y. The sample first passes through the feature extraction network to obtain four levels of features. The features are mapped to the high-dimensional feature space through the feature mapping network. During training, only the parameters of the feature mapping network are updated because the feature extraction network has been trained. The way to calculate the loss is the contrast loss, which is expressed as:

[0079]

[0080] Among them, D i (x1,x2)=||G i (x1)-G i (x2)||2 represents the Euclidean distance between the semantic features of the i-th sample pair,

[0081] m is the set threshold. i (x i) represents the result of the time-frequency graph after passing through the RF fingerprint feature extraction network and the RF fingerprint feature mapping network. All training in this step is performed on the training set in the basic data set, and the test set in the basic training set is used for testing.

[0082] Step 8, determine whether the network training is completed. If so, execute step (9). If not, increase the number of training iterations by one and continue training the RF fingerprint feature mapping network.

[0083] Determine whether the current training period has reached the maximum number of training periods. If so, proceed to step (8). If not, continue to train the RF fingerprint feature mapping network in step (7). The RF fingerprint feature mapping network finally saved is the model with the smallest loss function value when testing the test set in the basic data set.

[0084] Step 9: Build a machine learning model classifier. Input a small number of samples into steps (6) and (7) in sequence and then input the results into the model to select the optimal classifier.

[0085] At this point, the three benchmark network models have been trained and entered the small sample learning stage. Step 9 uses the support set in the new class data set for training. According to different tasks, a small number of samples are input into the optimal model obtained in step (6) and step (7), and finally the semantic information of each dimension of the sample is obtained. Construct a machine learning model classifier, input the semantic information into the classifier for training, and after multiple rounds, select the optimal classifier and save it. In the present invention, support-oriented online is used as the machine learning classification method.

[0086] Step 10: Input the validation set data in the new class data set into the network obtained in steps (6), (7), and (9) in sequence.

[0087] Step 11: Output the classification results.

[0088] The effect of the present invention is further described below in conjunction with simulation experiments.

[0089] 1. Simulation conditions and parameter settings:

[0090] The simulation experiment of the present invention is carried out on the simulation platform of Python3.6 and Pytorch1.10.1. The computer CPU model is Intel Core i7, and the equipped model is NVIDIA Geforce RTX 2060 independent graphics card. The data set adopts the public data set drone detect dataset: a radio frequency dataset of unmanned aerial system (UAS) signals for machine learning detection & classification. The data set contains the radio frequency signals of seven drones in three states under four conditions: no interference, Bluetooth interference, WIFI interference, and simultaneous WIFI and Bluetooth interference. The three states are drone on, flying, and hovering. Each data file has 1.2×10 8 Sample points, sampling rate is 60 MHz. In actual testing of the present invention, each data file is divided into 100 parts, that is, each sample duration is 20 milliseconds. The basic data set is four drones randomly selected from the seven drones, and the new class data set is the radio frequency data of the remaining three drones.

[0091] 2. Simulation content:

[0092] Attached Figure 5 It is a comparison chart of classification accuracy when the present invention and the prior art are faced with different tasks under an interference-free data set. Figure 5 The horizontal axis represents different tasks, and "C classification K samples" means that the number of categories of the classification task is C and the number of samples in the support set of each category is K. The vertical axis represents the recognition accuracy. The broken line marked with a circle represents the classification accuracy curve of the method of the present invention, the broken line marked with a square represents the classification accuracy curve of the graph neural network under different tasks, the broken line marked with a triangle represents the classification accuracy curve of the relational network under different tasks, the broken line marked with a cross represents the classification accuracy curve of meta-transfer learning under different tasks, and the broken line marked with a five-pointed star represents the classification accuracy curve of meta-learning under different tasks. A total of four tasks were tested, namely 5 categories and 1 sample, 3 categories and 1 sample, 5 categories and 5 samples, and 3 categories and 5 samples. By comparison, it can be seen that the classification accuracy of the present invention is significantly higher than other existing methods. When executing 5 categories and 1 sample, the classification accuracy of the present invention reached about 75%, which exceeded the accuracy of about 68% of the method using the relational network and the accuracy of about 52% of the method using meta-transfer learning. When executing 3 categories and 5 samples, the classification accuracy of the present invention reached about 93%, which exceeded the accuracy of about 77% of the method based on the graph neural network and the accuracy of about 83% based on meta-learning. At this time, the classification accuracy of the method of the present invention gradually reached saturation, about 95%.

[0093] Attached Figure 6 It is a comparison diagram of the convergence speed of training under different interference levels using the present invention. Figure 6 The horizontal axis represents the number of training periods (times), and the vertical axis represents the training accuracy. The broken line marked with a five-pointed star represents the loss function curve of the method of the present invention in the absence of interference, the broken line marked with a cross represents the loss function of the method of the present invention in the presence of WIFI signal interference, the broken line marked with a triangle represents the loss function of the method of the present invention in the presence of Bluetooth signal interference, and the broken line marked with a pentagon represents the loss function of the method of the present invention in the presence of both WIFI signal and Bluetooth signal interference. By comparing the training speed convergence curves obtained when there are different degrees of interference, it can be seen that the speed of the present invention is very little affected by interference during training. When there is only one signal interference, compared with no interference, the number of training periods is almost the same, and convergence can be completed in about 20 training periods. When there are two signal interferences, the number of training periods is extended to 30 training periods, and training can be completed in a relatively short time overall.

[0094] Attached Figure 7 It is a comparison chart of classification accuracy under different signal-to-noise ratios using the present invention. The horizontal axis represents the signal-to-noise ratio of the signal, and the vertical axis represents the classification accuracy. The broken line marked with a five-pointed star represents the accuracy curve of the method of the present invention under a 3-class 5-sample task, the broken line marked with a cross represents the accuracy curve of the method of the present invention under a 3-class 1-sample task, the broken line marked with a square represents the accuracy curve of the method of the present invention under a 5-class 1-sample task, and the broken line marked with a circle represents the accuracy curve of the method of the present invention under a 5-class 5-sample task. It can be seen that when the signal-to-noise ratio is higher than 15dB, the recognition accuracy is almost unaffected, and the present invention still has a very good recognition accuracy. At the same time, the accuracy of the two methods of 3 categories and 1 sample and 5 categories and 5 samples are similar, which provides guiding significance in practical applications.

[0095] Based on the above simulation results and analysis, the intelligent identification system and method of radio frequency fingerprint in a small sample environment proposed in the present invention can achieve higher classification accuracy than the existing methods, have stronger anti-interference ability, and can still maintain a high accuracy rate under low signal-to-noise ratio conditions.

[0096] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. A small sample radio frequency fingerprint intelligent recognition system, characterized in that: It includes data acquisition module, data preprocessing module, data partitioning module, data training module, small sample learning module and data output module. The data acquisition module is used to collect radio frequency signals. The data preprocessing module is used to perform noise processing on the signal of the data acquisition module and convert it into a time-frequency diagram. The data classification module is used to classify the time-frequency diagram output by the data preprocessing module. The data training module is used to perform data training on the radio frequency fingerprint classification network, radio frequency fingerprint feature extraction network, and radio frequency fingerprint feature mapping network on the classified data. The small sample learning module constructs a machine learning model classifier based on the trained RF fingerprint classification network, RF fingerprint feature extraction network, and RF fingerprint feature mapping network, performs small sample data learning, and selects the optimal classifier. The data output module is used to input test data and output classification results.

2. A small sample radio frequency fingerprint intelligent identification method, characterized in that: Using the small sample radio frequency fingerprint intelligent recognition system described in claim 1, the method comprises the following steps: Step 1: Collect the radio frequency signal in the data acquisition module, and express the collected signal in the form of a vector, where I is the in-phase component and Q is the orthogonal component; Step 2: In the data preprocessing module, a corresponding category label file is created according to the collected radio frequency signal, and noise is added to the radio frequency signal; Step 3, preprocessing the radio frequency signal with noise added obtained in step 2, setting the segmentation point after normalization operation, and converting the segmentation result into a time-frequency diagram; Step 4: In the data partitioning module, the time-frequency graph obtained in step 3 is classified into categories to obtain a basic data set and a new class data set. The basic data set and the new class data set are further divided into a support set and a test set respectively. Step 5: In the data training module, a radio frequency fingerprint classification network is constructed and pre-trained, and the support set data of the basic data set obtained in step 4 is input into the network and trained to obtain a radio frequency fingerprint classification pre-training model; Step 6, reconstructing the radio frequency fingerprint classification pre-training model obtained in step 5, and obtaining the radio frequency fingerprint feature extraction network using the output parameters in step 5; Step 7: construct a twin network training model, wherein the twin network training model is composed of a radio frequency fingerprint feature mapping network, and the features learned by the radio frequency fingerprint feature extraction network in step 6 are input into the radio frequency fingerprint feature mapping network for training to obtain the semantic space information of the radio frequency signal; Step 8, determining whether the RF fingerprint feature mapping network training is completed, if so, executing step 9, if not, increasing the number of training iterations by one and continuing to train the RF fingerprint feature mapping network; Step 9: In the small sample learning module, a machine learning model classifier is constructed, and the results obtained after a small number of samples are input into the machine learning model classifier in sequence in step 6 and step 7, so as to screen out the optimal classifier; Step 10, in the data output module, input the test set in the new class data set in step 4 into step 6, step 7 and step 9 in sequence; Step 11: Output the classification results.

3. The small sample radio frequency fingerprint intelligent identification method according to claim 2 is characterized in that: In the step 1, RF fingerprint recognition is regarded as a K-type hypothesis testing problem. Assume that the received signal of the k-th RF signal is x k (i) = s k (i)+ω k (i), where s k (i) represents the i-th sampling point of the k-th RF transmission signal, x k (i) represents the i-th sampling point of the k-th RF received signal, ω k (i) means the mean is 0 and the variance is σ 2 The signal received by the receiving end is expressed as an I / Q signal in vector form, x k =I k +Q k , where x k Represents x k The vector form of (i), I k and Q k Represent the in-phase and quadrature components of the signal respectively.

4. The small sample radio frequency fingerprint intelligent identification method according to claim 2 is characterized in that: In the step 2, Noise is added to the interference-free data, and the power of the interference-free signal is expressed as: Where N represents the number of sampling points of the RF received signal. Assuming the signal-to-noise ratio of the required signal is SNR, the noise power that needs to be generated is: P noise [dB]=P signal [dB]-SNR The final signal generated is s k [i]=x k (i)+n k [i]. where s k [i] represents the final generated signal, n k [i] indicates power P noise [dB] Generated additive white Gaussian noise.

5. The small sample radio frequency fingerprint intelligent identification method according to claim 2 is characterized in that: In the step three, After obtaining the data processed in step 1 and step 2, the segmentation point is determined according to the number of data sampling points, the data is segmented into multiple samples, the mean and variance of the data are calculated, the data is normalized, and a single sample is converted into a time-frequency diagram.

6. The small sample radio frequency fingerprint intelligent identification method according to claim 2 is characterized in that: In the step 4, The data and categories of the basic data set and the new class data set are disjoint. The basic data set and the new class data set are further divided into a support set and a test set respectively. The data of the support set and the test set are disjoint but have common categories.

7. The small sample radio frequency fingerprint intelligent identification method according to claim 2 is characterized in that: In the step 5, The constructed radio frequency fingerprint classification network is an improvement on the existing residual neural network. By changing the number of output categories of the last fully connected layer of the network, it is suitable for the current sample space. The complete radio frequency fingerprint training network consists of four residual units, which are composed of two convolutional layers with a kernel size of 3×3. The batch normalization layer standardizes the intermediate data in the middle layer of the network to avoid the gradient vanishing problem caused by the saturation of the partial derivatives of the intermediate variables. The residual stacking unit is constructed by sequentially connecting a 1×1 convolutional layer and a batch normalization layer. Finally, a linear correction unit is connected after each batch normalization layer as an activation function to introduce nonlinearity into the network. When pre-training the radio frequency fingerprint classification network, the trainable parameters of the network are randomly initialized, the number of network training epochs is initialized to 1, the maximum number of epochs is 50, and the learning rate is 0.

001. The stochastic gradient descent optimization algorithm is used as the network training optimizer, and the cross entropy loss function is selected to calculate the gap between the network output and the category, which is provided to the network training optimizer for optimization to obtain the radio frequency fingerprint classification pre-training model.

8. The small sample radio frequency fingerprint intelligent identification method according to claim 2 is characterized in that: In the step six, The features output by different layers of the RF fingerprint classification pre-training model represent feature information at different levels. The new RF fingerprint classification pre-training model is reloaded and the optimal parameters obtained in step five are loaded. The output parameters of the network are set to the output results of each residual unit. The output dimensions are 64×56×56, 128×28×28, 256×14×14 and 512×7×7 respectively. The front output layer has a small number of output channels, indicating relatively comprehensive features of the samples. The deep network has more output channels, indicating detailed features at different levels of the RF signal, and can clearly map the fingerprint information of the RF signal to the feature space.

9. The small sample radio frequency fingerprint intelligent identification method according to claim 2 is characterized in that: In the step seven, The RF fingerprint feature mapping network is trained by the method of training twin networks. The features learned by the RF fingerprint feature extraction network are mapped to a high-dimensional semantic space. The similarity between two samples is evaluated by calculating the Euclidean distance between the two input samples. The RF fingerprint feature mapping network consists of six basic residual triangular units. Each triangular unit inputs high-level features and low-level features of two different dimensions. For high-level features, two convolutional layers with a kernel size of 3×3 and strides of 2 and 1 are connected. The convolutional layer with a stride of 2 reduces the size of the data so that it matches the low-level features. Each convolutional layer is connected by a batch normalization layer to prevent data from being lost. The overfitting of the data is prevented, and finally the activation function is used to make the network nonlinear. The residual stacking unit is constructed by sequentially connecting a convolutional layer with a size of 1×1 and a step size of 2 and a batch normalization layer. For low-level features, it is composed of two convolutional layers with a kernel size of 3×3 and a step size of 1. The batch normalization layer is connected after the convolutional layer, and finally the activation function is used to make the network nonlinear. The processed high-level features and low-level features are summed and output after being processed by the basic residual unit. The residual unit here is consistent with the residual unit in step six. The features finally output by the RF fingerprint feature mapping network are mapped to a 512-dimensional semantic space after the average pooling layer and the linear regression layer.

10. The small sample radio frequency fingerprint intelligent identification method according to claim 2 is characterized in that: In the step nine, The small number of samples mentioned above are the support set data in the new class data set. According to different tasks, a small number of samples are input into the optimal model obtained in steps six and seven, and finally the semantic space information of each dimension of the samples is obtained. A machine learning model classifier is constructed, and the semantic space information is input into the classifier for training. After multiple rounds, the optimal classifier is screened out and saved.

Images