YANG Haifen,ZHANG Hao,WANG Houjun,and GUO Zhengyang

School of Information and Communication Engineering,University of Electronic Science and Technology of China,Chengdu 611731,China

Abstract: Radiation source identification plays an important role in non-cooperative communication scene and numerous methods have been proposed in this field.Deep learning has gained popularity in a variety of computer vision tasks.Recently,it has also been successfully applied for radiation source identification.However,training deep neural networks for classification requires a large number of labeled samples,and in non-cooperative applications,it is unrealistic.This paper proposes a method for the unlabeled samples of unknown radiation source.It uses semi-supervised learning to detect unlabeled samples and label new samples automatically.It avoids retraining the neural network with parameter-transfer learning.The results show that compared with the traditional algorithms,the proposed algorithm can offer better accuracy.

Keywords:radiation source identification,deep learning,semisupervised learning.

1.Introduction

Radiation source identification is the process of discriminating individual radiation source by comparing the features carried by the signal with the feature set and choosing the class that matches these features best [1,2].It has been widely used in finding illegal users in non-cooperative communication scene.The identification of radiation source is essentially a pattern recognition problem.Deep learning techniques can directly extract features from the data by updating weights according to back propagation,and it is widely used in pattern recognition,image and voice processing,and so on [3?7].There are many effective network models such as Alexnet [8],VGG [9],and Resnet [10].Recently,it has been introduced into radiation source identification,especially in radar radiation source identification.In [11],deep convolution neural network (CNN) was used to learn the gray scale image of raw Doppler imaging to distinguish mobile radar sensors.In [12],verification support network was used to realize multi-category or multi-target automatic recognition in variable-size synthetic aperture radar (SAR) images.In[13],the signal was processed with Hilbert-Huang transform and residual network was used to identify it.In [14],a deep learning method based on CNN was used to identify internet of things (IoT) wireless devices with representation based on the continuous Wavelet transform.A CNNbased approach was presented to classify specific emitter using the compressed bispectrum [15,16].

Radiation source identification is a difficult issue in the field of deep learning due to the weak signal and complex transient characteristics [17,18].What makes things worse is that in non-cooperative communication scene lots of high-quality labeled training data is hard to get.Special algorithms have been applied in machine learning to process the exploit unlabeled data effectively[19?21].Zhong used semi-supervised learning to perform dialect identification with generative adversarial networks [22].Li proposed an effective approach based on kernel semi-supervised discriminator analysis [23].In [24],the acoustic identification algorithm combined semi-supervised learning with safe semi-supervised support vector machines (SVM) was proposed.Identifying radio signal based on semi-supervised learning with nonlinear feature was performed [25].The pseudo-label algorithm is proposed to realize semi-supervised deep learning [26,27].

Semi-supervise learning can learn the information of unlabeled samples.So the network should be re-trained after predicting unlabeled samples.Re-training the network using random initialized parameters with new sample set costs a lot of time.Transfer learning [28?30]uses the parameters learned before to initialize the model,such as fine-tuning [31].In [32],new fully-connected layer and softmax layer is added to identify new classes and decrease the consumption of computation resources.Tu applied transfer learning on GoogleNet to perform classification on the proposed BreedNet [33].In [34],transfer learning was performed on GoogleNet to ensure low computational complexity.The weights for the entire network except the final one were transferred to another model in [35].

This paper proposes an algorithm based on semi-supervised learning and transfer learning to identify unknown radiation source.Our goal is to use available unlabeled examples to improve the classification accuracy and shorten the network training time.We present a recognition framework for unknown source recognition.It extracts the output of the penultimate layer of the network and compares multiple distance metrics.We propose transfer learning to avoid retaining the network when unknown radiation source appears.The proposed algorithm can dynamically identify unknown radiation source and train the network faster,and the performance is comparable to other algorithms.

The rest of this paper is organized as follows:Section 2 introduces system model of radiation source identification and the feature extraction of signals.The proposed algorithms are demonstrated in Section 3.Simulation results are given and analyzed in Section 4.Section 5 concludes the paper.

2.Preliminaries

The system model of radiation source identification is shown in Fig.1.The input signal data is intercepted and preprocessed firstly.After that,the feature should be extracted.If the feature dimensionality is too high,reduction process is needed.At last we use the classifier for recognition.

Fig.1 A typical radiation source identification process

To identify different communication individuals,we need to extract features from signals.Most of the received communication signals are non-stationary signals.Short-time Fourier transform (STFT) can be performed on non-stationary signals to extract features.STFT can divide the signal to multiple equal-length intervals by setting the window size.Each interval is approximately stable so it can be analyzed with the Fourier transform.Suppose the original signalx(t),t∈(?∞,+∞),the window functionw(t),and the definition of STFT is as follows:

The deep learning method can be used in source identification,such as CNN.It employs a convolution kernel to replace the human field of vision,which can reduce the amount of calculation.Fig.2 shows the structure of a general CNN network.It usually contains convolutional layers,pooling layers,and fully-connected layers.The convolutional layer can extract input features and multiple layers can gradually extract more refined features.The pooling layer can reduce the dimensionality of the hidden layer in the network and lower the amount of computation in the subsequent layers.After multiple convolutional and pooling layers,CNN uses a fully connected layer to classify the extracted features.Finally,the output layer uses the softmax function to obtain the probability distribution based on the input category,and it takes the class of the largest probability as the prediction result of the neural network.

Fig.2 Structure of a general CNN

3.Proposed algorithms

This paper proposes a method for the unlabeled samples of unknown radiation source.It uses the semi-supervised learning to detect the unlabeled samples automatically.After that,in order to avoid re-training the network for a unknown individual,we propose a dynamic classification of unknown radiation source.The proposed algorithm combines fine-tuning algorithm with the unknown radiation source identification algorithm.The flow graph is shown in Fig.3.In the pre-processing,STFT is taken from the labeled and unlabeled class data to extract features.

Fig.3 Dynamic classification of unknown radiation source

3.1 Neural network

We build the network structure shown in Fig.4.The network model has eight layers.It includes three 2-dimensional convolutional (Conv2D) layer.The sizes of the kernels are (11,11),(5,5),and (3,3),respectively.The activation function is the Relu function.The second,fourth,and sixth layer are maxpooling layer.The size of the maxpooling kernel is (3,3).The rate of dropout layer is 0.25.The seventh layer is the flatten layer.This layer transform all the output of the previous layer into one-dimensional.The last layer is dense layer.This layer is the output layer and uses softmax function to output the recognition result.

Fig.4 Network structure

3.2 Semi-supervised learning

The pseudo-label algorithm is used to realize semi-supervised deep learning.The algorithm directly uses the network’s prediction of unlabeled data as the label of the unlabeled sample.The loss function of the network is

whereCis the number of labels,MandM′are the number of mini-batch in labeled and unlabeled data,is the output units,is the unlabeled data,is the label,is the pseudo label of the unlabeled sample predicted by network,and α is a coefficient balancing parameters,which decides the role of unlabeled sample in network cost function.

Euclidean distance and cosine function can be used to measure the difference of two vectors.The Euclidean distance and cosine function of two vectorsA=[x1,x2,···,are shown as follows:

where θ is the angle betweenAandB.

The features of the labeled data in the dropout layer are extracted and distributed into groups.Suppose there areGgroups andSfeature vectors in each group,we compute the Euclidean distance and the cosine function of each vector with the mean vector of each label.The maximum and minimum values of Euclidean distance and the cosine for each group are found respectively.And the minimum range of variation of all groups is computed for each label as follows:

The minimum distance range of unlabeled samples is computed in the same way.Then we compare it with that of labeled samples.If it is less than the corresponding value of labeled samples,the group will be detected as samples of a new unknown individual and add pseudo labels automatically.Otherwise this group will be detected as samples of a known individual.

3.3 Transfer learning

Training a network takes a long time in deep learning.It is inevitable to retrain the network multiple times when unlabeled samples of multiple unknown communication individuals appear.This will greatly increase the time complexity of the algorithm.We propose the transfer learning to reduce the training time effectively in this situation.

The core idea of transfer learning is to find and use the similarity between the source and target domain.Given source domainDsand its source taskTs,target domainDtand its target taskTt,transfer learning learns the conditional distribution ofDtthroughDsandTs.Fine tune is a typical practical method in transfer learning as shown in Fig.5.The target model is compared with the original model and adjusted correspondingly.

Fig.5 Fine tune

4.Simulation and results

To examine the performance of the proposed algorithms,we compare the proposed algorithms with traditional algo rithms,such as principal component analysis (PCA) and tdistributed stochastic neighbor embedding (t-SNE).The dataset is composed of a range of communication signals measurements from seven radiation sources.The bandwidth of the transmitted signal is 5 MHz.Each communication individual has eight working modes with the same signal bandwidth,signal waveform,and modulation mode.

To fully verify the recognition of the proposed algorithm,we design the following test scheme.Partial signal data of four radiation sources are chosen as the labeled sample set.Their remaining data and the total data of the fifth radiation source are unlabeled.Then we detect the unlabeled data sample of the five individuals and label the unknown samples of the fifth individual.This situation is named as 4 to 5 and the accuracy performance is shown in Fig.6.It shows that the proposed method can maintain a high accuracy over 0.99 in the experiment.This is much higher than the t-SNE method and PCA method,which is 0.2 and 0.76 respectively.

Fig.6 Results of continuous dynamic identification

After the samples of fifth individual are labeled,a new network model based on the transfer learning is trained.We repeat the above steps.The sixth and seventh radiation sources are regarded as new unknown individual in order.Fig.6 also shows that the accuracy of the proposed method is much higher than the t-SNE method and PCA method in all three consecutive tests.The proposed method can maintain a high accuracy over 98% in the experiment.

Transfer learning can avoid retraining the network and thus reduce the training time in the situation of unlabeled samples.Fig.7 shows the comparison of the epoch spent between transfer learning and without transfer learning.As is shown,when we use the proposed method based on transfer learning,the number of epoch can be reduced greatly in the system updating from 4 to 5 individuals,5 to 6 individuals,and 6 to 7 individuals.In the overall system update,using transfer learning can save 33% of time.The result shows that using transfer learning can effectively reduce the training time of the network.

Fig.7 Epoch performance of continuous dynamic identification

5.Conclusions

To solve the unknown individual identification problem,this paper proposes dynamic identification system architecture based on semi-supervised learning and transfer learning.It uses Euclidean distance and angle cosine to measure the sample distance.We first train the network on labeled samples and then retrain part of the network for unlabeled ones.The model can automatically identify unknown individual samples,automatically add labels,and dynamically update the identification system model in practical application.Experiments and comparisons show that the proposed algorithms have improved performance of recognition accuracy and calculation efficiency.