Source: Journal of Shanghai Jiaotong University
Authors: Tang Zeyu, Zou Xiaohu, Li Pengfei, Zhang Wei, Yu Jiaqi, Zhao Yaodong
Summary
Under the small sample conditions caused by non-cooperative scenarios, it is difficult and hotspot to robustly extract the target features of communication radiation source and accurately identify the target. Aiming at the problem of small sample individual identification of OFDM communication radiation source, the article is in Based on the data enhancement of phase domain and time domain flipping and the migration of source domain instances, a method for individual identification of non-cooperative communication radiation sources is proposed. The data enhancement method of different domain flipping is used to expand the data set, combined with the improved residual network, to achieve The purpose of improving the accuracy of individual identification of OFDM communication radiation sources is to introduce transfer learning to enhance the generalization ability of the identification model. The experimental results show that: the data enhancement strategy improves the OFDM communication under small samples The accuracy of individual radiation source identification, the introduction of the transfer learning method speeds up the convergence speed of the model, slightly improves the accuracy of model identification and improves the robustness.
Key words
Orthogonal frequency division multiplexing; small sample recognition; data enhancement; transfer learning; deep learning; target recognition
Communication radiation sources are affected by manufacturing, technology, processing, etc., and components of the same type of equipment shipped in the same batch will also have individual differences, resulting in subtle features related to hardware in the transmitted signal, and these features are not affected by the signal. The impact of transmission content. Based on this, the individual identification technology of communication radiation sources is used to analyze the hardware characteristics of the radiation source carried by the electromagnetic signal to judge the source of the electromagnetic signal to determine its use, model, threat and other information. This is targeted It is of great significance in public security management and national defense and military applications[1]. At the same time, since Orthogonal Frequency Division Multiplexing (OFDM) technology is widely used in 4G/5G mobile communication networks, wireless local area networks, digital audio/video broadcasting, military and civilian wireless communication links, military and civilian UAV control signal links and other fields, so the research on individual identification of OFDM communication radiation sources has extensive application significance .
Individual identification of communication radiation sources is essentially a machine learning classification and identification problem. Through the training samples of known communication radiation source category information, learn and train the feature extractor and classifier parameters, and then test and judge the radiation sources of unknown categories of samples Attributes [2-6]. The key to individual identification of communication radiation sources is to extract the effective features of the radiation source. Using the residual network can solve the problem of low efficiency of manual feature extraction in traditional identification methods and maintain a high recognition accuracy[ 7-10]. At the same time, combined with the characteristics of the input data, the structure of the residual network can be adaptively adjusted to further improve the recognition effect[11-12].
However, in practical applications, the signal collector cannot actively control the electromagnetic environment of the signal background or the collected radiation source, making the target electromagnetic signal poor in quality and difficult to intercept. Under such non-cooperative conditions, the data that can be labeled and trained is limited, and it is easy to Over-fitting leads to a decrease in classification accuracy, and it is difficult to form an effective identification model. Therefore, it is necessary to combine data enhancement and small sample identification methods to identify individual targets of communication radiation sources [13-17]. Liu et al. Identify problems that may be disturbed by unstable features, use data enhancement methods to expand data when training networks, reduce preprocessing complexity, and suppress interference from power changes, frequency offsets, phase offsets, and channel noise.Zhou [19] introduced the generation confrontation network into the field of electromagnetic signal classification, and used the sample enhancement capability of the generation confrontation network to construct a semi-supervised learning framework, directly processing the IQ data of electromagnetic signals, and to a certain extent solved the classification of electromagnetic signal types and the identification of individual radiation sources The overfitting problem in . If a neural network can robustly classify objects in different states, it is said to have an invariant property. More specifically, a convolutional neural network (CNN) has a , zooming, flipping and other operations are invariant, which is the premise of data enhancement. In order to solve the problem of overfitting, data enhancement can be used to generate more samples from a small number of available signal samples and expand the training samples.
At the same time, in order to solve the problem of difficult model training and poor generalization caused by difficult data acquisition, knowledge learned in a resource-rich environment can be used to assist learning strategies in another field, that is, the idea of transfer learning. Most data or tasks are related , so transfer learning applies the knowledge or patterns learned in a certain field or task to related fields or problems, and shares the trained model parameters (which can also be understood as the knowledge learned by the model) with the new model, thereby avoiding The network learns from zero to speed up and optimize the learning efficiency of the model [20-21]. Feng et al. [22] use the transfer learning method to screen out the different parts of the source domain and the target domain, so that the distribution of the source domain is close to the distribution of the target domain, and the generated The new data set that can support subsequent classification methods solves the problem of insufficient labels in the target domain of radar radiation sources caused by complex electromagnetic environments. Kuzdeba et al. There are differences in the distribution of the signal source domain and the target domain and the problem of missing sample labels. Through the migration of network model parameters, individual identification of different radiation sources can be realized. Liu et al. [24] pointed out that the subtle characteristics of radiation sources will change over time, place and The distribution of training samples and test samples is different, and the knowledge in the marked source domain data is effectively used and transferred through transfer learning, which improves the individual recognition performance of communication radiation sources. It can be seen that transfer learning It can better explore the data structure information, transfer valuable knowledge from the source domain to the target domain, and improve the recognition accuracy under the condition that the target domain samples are not enough to support model training.
Aiming at the target recognition problem of OFDM signal communication radiation source under the condition of small samples, a data enhancement method based on phase and time domain inversion and source domain instance migration are proposed, the data set is expanded by the inversion data enhancement method, and transfer learning is introduced to Strengthen the generalization ability of the recognition model and improve the recognition accuracy of individual targets.
1 Small-sample OFDM target enhancement recognition method
Taking the OFDM signal generated by the communication radiation source as the research object, the target recognition problem of the communication radiation source is studied in the case of a small sample. The experimental data comes from the most easily obtained OFDM signal of the mobile phone. By judging that the mobile phone currently emitting the OFDM signal belongs to Huawei Enjoy Z, which model of Xiaomi Play4 or vivo Y70s, etc., further determines whether the individual mobile phone belongs to which one of the five Huawei Enjoy Z or five Xiaomi Play4 mobile phones. The individual identification of the mobile phone is helpful to identify criminals in the The mobile phone used in the case of frequent replacement of the Subscriber Identification (SIM) card, and then to identify its identity, has a certain application value in public security management. There is a certain commonality in digital broadcasting, communication radio and other fields.
Figure 1 shows the overall flow of the small-sample OFDM target enhancement recognition method based on transfer learning. Among them, (1, 2) represents the step size, that is, after each convolution, the convolution kernel moves one position vertically, and moves horizontally 2 positions; 1×7 and 3×3 are the size of the convolution kernel; 64 and 128 are the number of channels output by this layer, that is, the number of convolution kernels; 5 in the fully connected layer indicates the dimension of the output, that is, the recognition The model finally outputs a 5-dimensional vector. First, through data preprocessing, segmental screening, normalization, and dual-channel data extraction are performed on the pre-collected mobile phone OFDM signal; second, the convolution kernel structure is modified according to the multi-carrier characteristics of the OFDM signal , to make adaptive adjustments to the residual network; again, use the data enhancement method to expand the data samples; finally, on the basis of the above experiments, use the transfer learning method to train the old model using the source domain data, and use the old model The new model is initialized with parameters, so as to construct the target domain recognition model, complete the target recognition task, and finally achieve the purpose of knowledge transfer.
1.1 OFDM signal preprocessing
In the acquisition of OFDM signals, the acquisition time of each data file is 100 ms. As shown in Figure 2, the data of each 1 ms length includes 14 OFDM symbols and 1 time slot number, and the data is transmitted in two time slots. The 14 OFDM symbols are 3 effective signals, 1 reference signal, 6 effective signals, 1 reference signal, and 3 effective signals. Each OFDM symbol contains 108 subcarriers, and each subcarrier is represented by two IQ channels (2 short data type).
Data preprocessing is divided into three parts: data segment screening, normalization, and dual-channel data generation. Firstly, the data of each 100 ms is segmented, and the signals in each data file are divided into effective signals, reference signals, and time slots, and the effective signals are selected. Then, since the individual fingerprint characteristics of the radiation source have nothing to do with the transmission power of the signal, in order to avoid the influence of the difference in signal power, the data samples were normalized; finally, the processed data of each 100 ms were divided into spectrogram, IQ Two-way real part and real part plus imaginary part, IQ two-way modulus single-channel and modulus-length plus phase dual-channel, a total of 5 ways, each generating a signal data sample.
1.2 Residual network classification method based on adaptive adjustment
The residual network (ResNet) is a convolutional neural network. By using the residual module, the performance degradation problem caused by the increase in the depth of the convolutional neural network can be solved. It contains a new network block, the network block input is x, and the output is Q(x), this network block maps the input feature x to Q(x)-x, and Q(x)-x is recorded as F(x), that is, the network block calculates the original mapping Q(x) and the input feature x The difference F(x) is called "making the network block learn the residual of the original network block Q(x) and the input feature x". The network block structure is shown in Figure 3. Among them, ReLU is a modified linear unit.
Figure 3 shows that the network block formed by the superposition of two weight layers completes the mapping from the input feature x to the residual F(x), and performs an element-level addition operation between F(x) and the short-circuit link on the right, and finally The output of this network block is F(x)+x, that is, the output after passing through all weight layers in Figure 3 is still Q(x).
The residual block structure no longer allows the weight layer to only output the final feature map, but allows the weight layer to output the difference F(x) between the final feature map and the input feature, and then perform element-level addition of F(x) and the input feature x operation to obtain the final feature map. The structure in Figure 3 can be expressed as
y=σ(F(x, {Wi})+x)
(1)
In the formula: F=W2σ(W1x), σ is the corrected linear unit (ReLU); Wi is the weight matrix of the i-th weight layer. After the element addition of F(x) and x is completed, ReLU is also performed in Figure 3 operation, that is, the final output of the structure is σ(Q(x)).
ResNet-18 (residual network with 17 convolutional layers and 1 fully connected layer) is used as the basic model, and its structure is adjusted according to the preprocessed data. OFDM signal data is preprocessed into 3 forms: ① three channels In the form of two-dimensional pixels, for this data form, convolutional neural networks are often used to build models; ② In the form of single-channel one-dimensional data, the stored data is the complete IQ signal module length; ③ In the form of dual-channel two-dimensional data, The phase information between the I-channel signal and the Q-channel signal is increased to maximize the information useful for classification. For the latter two data forms, the network input layer needs to be adjusted first, and the residuals that originally supported the color image RGB three-channel The network input layer is modified to support amplitude and phase dual-channel data as input, and at the same time copy the weights of the first two channels to ensure that the initialization weights are consistent with the residual network. Then, for the special nature of OFDM symbols containing 108 subcarriers , adaptively adjust the network structure, optimize the feature extraction method of the residual network, and reduce the convolution kernel of the original first convolutional layer from 7×7 to 1×7, so that it pays more attention to the characteristics of the same subcarrier. Features. Finally determine the residual network structure as shown in Figure 4.
The loss function is used to determine the closeness between the predicted value and the real value, which helps to optimize the parameters of the neural network. Therefore, cross entropy is used as the loss function to measure the difference between the actual output and the expected output. In addition, through the softmax classifier, let the classification The sum of predicted values is 1. The loss function is defined as follows:
Loss(p, q)=-∑p(x)log q(x)
(2)
In the formula: p is the label value; q is the predicted value.
Gradient descent method (SGD) is a commonly used optimization method in machine learning. The small batch SGD method is used to improve the training speed, improve memory utilization and reduce the number of iterations. The SGD optimization method sets the momentum parameter to 0.9, the learning rate to 0.01, and the batch size to 16, the number of training iterations is 100.
1.3 Data enhancement method based on phase domain and time domain inversion
In order to identify individual radiation source signals under small sample conditions, more sample data need to be obtained, so small changes are made to the existing data sets, such as rotation, shift, flip, etc., to enhance the sample data. Aiming at the time-frequency domain characteristics of OFDM signals, combined with the periodic characteristics of the signal, a method of data enhancement by flipping in the phase domain and time domain is proposed.
In the phase domain, the Hilbert transform is a commonly used method in signal processing. By inverting the phase of the original signal, different angles of the signal are displayed. The Hilbert transform method is used to convolve the signal to construct the time-frequency of the target. Parses the signal, enabling enhancements.
Assuming an existing signal s(t), then define the Hilbert transform of the signal s(t) as
It can be known from the above formula that H(s(t)) is the convolution of s(t) and 1/(πt). By looking up the table to solve, we can get:
In the formula: j is the imaginary unit; ω is the angular frequency.
Therefore, |H(s(t))|=1, let:
Introduce Euler's formula ejφ0n=cos ω0n+jsin ω0n, from Euler's formula, when cos ω0n+jsin ω0n=-j, ω0n=-π/2, namely φ(ω)=-π/2(ω>0 ). Similarly φ(ω)=π/2(ω<0), we can get:
It can be seen from the above formula that when the frequency is greater than 0, the phase shifts to the left by 90°, otherwise, it shifts to the right by 90°, thus realizing the signal inversion in the phase domain.
In the time domain, in view of the periodic characteristics of the signal, the 1200 OFDM effective signals contained in each subcarrier in the 100 ms original signal data are horizontally flipped, and the flipped subcarriers are recombined into a 108 subcarrier OFDM signal, form new training samples, expand the training set, improve the overall performance, so as to achieve the purpose of data enhancement.
1.4 Transfer Learning Method Based on Source Domain Instances
Electromagnetic signals are difficult to intercept, which makes it difficult for the collected signals to cover a large number of radiation source targets. Therefore, when new target signals appear, it is necessary to migrate the existing individual recognition models. In order to achieve faster model migration and achieve better The recognition effect usually requires a large amount of labeled data. Under the condition of small samples, it is necessary to make full use of the existing labeled data to improve the learning effect of the transfer model. Therefore, an example-based transfer learning method based on source domain data is proposed. .
Collect sample data of 5 mobile phones of the same brand and model, and complete model training under the proposed classification and recognition method. After new mobile phone samples of different brands or the same brand and model appear, use the trained model to initialize the new mobile phone recognition task model parameters, that is, select data similar to the target domain from the source domain as the training set of the new task pre-training model. Therefore, the data sources of the new and old recognition tasks have certain similarities, which is equivalent to expanding the new task training set, fully The characteristics of electromagnetic signal data are used. This method improves the robustness of the model, speeds up the fitting speed of network parameters, and can achieve better recognition accuracy faster.
As shown in Figure 5, the same source old data is used to train the old model on the shared network, and then the model parameters of the old model are used to initialize the training of the target task, so as to achieve the purpose of knowledge transfer.
2 experiments
The experimental data is collected from the China Unicom 4G OFDM signal of the individual target of mobile phones sharing a SIM card. There are 30 mobile phones of 5 brands in total, and the mobile phone models of each brand are the same. 5 mobile phones are randomly selected as the target domain samples, including 2 Huawei, 2 iqoo and 1 Xiaomi, each mobile phone is a class in the recognition model, the effective signal samples of each mobile phone are 62, 54, 60, 60, 49 respectively, the signal samples of 5 mobile phones are divided into 4 : 1 is divided into training set and verification set, which is used for the adaptively adjusted residual network model in the training experiment. Using the remaining mobile phone target data to construct 2 source domains, each source domain has 5 mobile phones of the same brand and model, and each A mobile phone is a type of recognition model, with 40 to 60 effective signal samples, which are consistent with the target domain, and used as training data for the old model of transfer learning. Each set of experiments is repeated 10 times to calculate the average recognition accuracy and standard deviation . Among them, the recognition accuracy is defined as:
Recognition accuracy =
By comparing the results of mobile phone individual target recognition, the experiment sequentially analyzes the influence of different data input methods, data enhancement methods and transfer learning methods on the recognition model, so as to study the effect of OFDM communication radiation source target recognition under small samples.
2.1 Impact of data entry
In the data preprocessing, the data is saved in the form of single-channel and dual-channel. The dimension of single-channel data is (108, 1200, 1), and the dimension of dual-channel data is (108, 1200, 2). Channel real part, dual-channel real part and imaginary part, single-channel amplitude and dual-channel amplitude and phase data are input in five ways, and the residual network after modifying the number of channels is trained. The experimental results are shown in Figure 6.
As shown in Figure 6 and Table 1, comparing five data input methods: inputting in the form of an image will cause different subcarriers to superimpose and influence each other, the effect is the worst, and the average recognition accuracy is only 32.23%; The internal connection of IQ signals is lost to a certain extent, and the average recognition accuracy rate is about 51%. The average recognition accuracy rate is 66.16% and 67.28% respectively when input in the form of amplitude, amplitude and phase combination. .The combination of amplitude and phase has the highest recognition accuracy, which shows that the dual-channel method of amplitude and phase combination not only contains the amplitude and phase information of the signal, but also retains the internal relationship of the IQ two-way data to the greatest extent.
Based on the combined dual-channel input of amplitude and phase, the structure of the residual network is adjusted to change its feature extraction method, the convolution kernel is reduced to one dimension, and the step size is modified at the same time, compared with the original residual network The experimental results are shown in Figure 7.
As shown in Figure 7 and Table 2, comparing different network structures: the average recognition accuracy of the original 7×7 convolution kernel network structure is low, and the fluctuation is large and difficult to quickly converge, while the adjusted convolution kernel size is 1× 7. The average recognition accuracy rate is 69.83%, which can keep the model recognition accuracy rate relatively stable. This shows that the adjusted residual network can make the model pay more attention to the characteristics of the same sub-carrier, fuzzy the interference information that may exist between sub-carriers, This enables the model to converge quickly and improves the recognition accuracy.
2.2 Comparison of data augmentation methods
The adjusted residual network results in the previous section were used as the baseline, and the effect comparison experiments of four data enhancement methods were carried out: ① Data enhancement of subcarrier sequencing, generating 9 sets of random sequences from 0 to 108, and adjusting the subcarriers in each original sample. In order, expand the training set to 10 times the original; ② add Gaussian noise, set the variance coefficient to 0.01. In each round, randomly add Gaussian noise to 1/4 of the data; The Hilbert transform is performed on the part to construct the analytical signal, and the amplitude and phase of the complex number results are input as dual channels; ④ time-domain flipping, horizontally flipping all subcarriers in each training set sample, synthesizing new samples, and expanding the training set to the original 2 times. The training results are shown in Figure 8.
As shown in Figure 8 and Table 3, the recognition accuracy of the four types of data enhancement methods increases by 8% to 9% compared with that without data enhancement, which is a significant improvement; the proposed data enhancement method based on phase domain and time domain inversion has achieved The average recognition accuracy is 78.48% and 79.35%, respectively. This shows that the analytical signal data constructed by the data enhancement method of phase domain and time domain inversion has better feature representation ability and makes the model more stable. At the same time, it is also proved that the data enhancement strategy can greatly improve the accuracy of OFDM communication radiation source target recognition under small samples, and has a good effect.
2.3 The role of transfer learning
Through comparative experiments, the effect of adding transfer learning to the model is studied. The data of 5 mobile phones (including Huawei, iqoo, and Xiaomi) in the above experiment is used as the target domain, and then 5 mobile phones of the same model as Xiaomi and 5 phones of the same model as vivo are used in the remaining mobile phone targets. Construct two source domains from the data of the mobile phone, use the source domain target sample data for pre-training, generate two source domain recognition models, and use the parameters of the source domain model to train the model of the target domain when migrating to the target domain to initialize.
There are some common features in the target domain and the source domain, but there are also certain differences. The main manifestations are: First, although the two types of domains contain mobile phones of the same brand and model, the individual mobile phones are different. For example, the source domain and the target domain There are slight differences in the hardware characteristics of Xiaomi mobile phones in , which makes the distribution of the two types of domains somewhat different. The training results are shown in Figure 9(a). Second, the mobile phones in the two domains are of different brands, and these mobile phones are designed by different brand manufacturers , production, and assembly. From the perspective of hardware, the distribution of the two types of domains is quite different. The training results are shown in Figure 9(b). If the difference is too large to produce negative transfer, it can be used as an experimental object for transfer learning.
In Figure 9(a), since both the source domain and the target domain contain Xiaomi brand mobile phones, although the individual mobile phones are different, the pre-training model can better distinguish the difference between other brands and Xiaomi brand mobile phones in the target domain. Therefore, the pre-training The initial target recognition rate of the model is high. In Figure 9(b), the source domain and the target domain contain mobile phones of different brands. The initial target recognition rate of the pre-trained model is consistent with that of the control group without transfer learning, but its convergence speed is faster.
As shown in Table 4, although the recognition accuracy of transfer learning in two different source domains is lower than that of the control group without transfer learning, the difference is not large, and the fluctuation of the target recognition rate is relatively stable, which shows that the addition of transfer learning has improved The generalization and robustness of the OFDM communication radiation source identification model. At the same time, as shown in Figure 9, the pre-trained model converges faster than the control group without transfer learning, indicating that transfer learning can speed up the fitting of network parameters. Faster and better recognition accuracy.
3 Conclusion
Aiming at the problem that the communication radiation source signal data is scarce and it is difficult to effectively carry out target recognition under non-cooperative conditions, the complex features of the subcarrier signal are extracted using a dual-channel method, and the data sample set is expanded by using the flipping data enhancement method, and then based on the adaptively adjusted residual The transfer learning method of the network and examples is used to identify and model the target of the communication radiation source, and realize the model transfer and target recognition of the communication radiation source under the condition of small samples. The experimental results show that the input of the amplitude and phase dual channels Better retain signal data features, combined with data enhancement methods such as phase domain and time domain flipping, can solve the problem of sample scarcity to a certain extent, and significantly improve the recognition performance of the adaptively adjusted residual network. In addition, the transfer learning method can accelerate Network parameter fitting speed, using a small amount of data to quickly build a recognition model that can adapt to new targets, effectively improving the robustness and generalization of the model without reducing the classification performance.
In practical applications, there is not only the problem of scarcity of samples, but also the lack of any data. The identification of such unknown radiation source targets will be the research direction of the next stage.
Disclaimer: The articles and pictures reproduced on the official account are for non-commercial educational and scientific research purposes for your reference and discussion, and do not mean to support their views or confirm the authenticity of their content. The copyright belongs to the original author. If the reprinted manuscript involves copyright and other issues, please contact us immediately to delete it.