Small-Sample Radiation Source Identification Method Based on Siamese Network

Abstract

Aiming at the problems that it is difficult for neural networks to obtain sufficient data and it is difficult to train with imbalanced positive and negative samples in a small-sample environment, the present invention provides an improved small-sample radiation source recognition method based on a siamese network. Data combination is used for data expansion, and at the same time, domain transformation of traditional signal processing is used for data enhancement, so that the sample size of the data set has been increased by almost a square level. At the same time, the present invention uses a siamese network as the structure of the neural network, which fits the expanded data structure. The balanced contrast error is used to replace the probability to control the input of the network training samples, and the data set is organized by full traversal of binary groups, avoiding the deviation of network training caused by imbalanced positive and negative samples. After that, the siamese network output feature vector of the sample to be judged is compared with the feature centers of various categories to obtain the probability that the sample belongs to each type of radiation source, that is, the feature center is used to classify the sample, reducing the jitter of the judgment and having a fast judgment speed.

Description

Technical Field

[0001] The present invention relates to electromagnetic wave and signal recognition technologies, and particularly to radiation source recognition technology. Background Art

[0002] With the continuous development of communication technologies and communication algorithms; the types and quantities of access devices in communication networks are constantly increasing; the signal systems and types of communication protocols emitted by communication devices are constantly expanding, and the difficulty of conducting communication countermeasures against specific radiation source communication targets is increasing continuously. As an important part of communication countermeasures, the radiation source recognition technology has become even more crucial. In an actual complex environment, the obtained radiation source information often has characteristics such as a large amount of data, little effective data, complex channel parameters, redundant signals, and difficulty in extracting radiation source fingerprints and internal modulation information. How to make full use of a small amount of effective information to more effectively extract radiation source features for radiation source recognition has become one of the difficulties in radiation source recognition.

[0003] The radiation source recognition method based on deep learning is more likely to obtain a better feature extractor, and the deep learning method can combine the extraction of radiation source fingerprint information with classification and train them unified in the network, which not only omits the manual fingerprint extraction operation but also has a higher classification accuracy. However, in a complex actual scenario, effective data is extremely scarce, most signal data is unlabeled or even invalid, making it difficult to meet the data requirements for intelligent radiation source recognition based on deep learning. The high training cost of the neural network (usually up to several hours) is also one of its drawbacks.

[0004] In an actual scenario, the channel environment is complex, there is little effective data, and the requirement for real-time processing is high. For the radiation source recognition method based on traditional signal analysis: it is difficult to design feature extraction methods for multiple radiation sources and the performance is not good; for the intelligent radiation source recognition method based on deep learning: the training cost is high, and the training time of several hours cannot meet the requirements of real-time processing; the cost of collecting data is high, and it is very resource-consuming to find enough effective data for network learning from a large amount of data. Although there are attempts to enhance data features in the transform domain, the improvement of the network's rapid learning from small samples is not obvious. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method that can extract radiation source features from a small sample radiation source signal dataset, classify them, and at the same time is easy to train a network to complete the radiation source classification task in a small sample environment, aiming at the difficulties in radiation source recognition and network training under the condition of small samples in a complex communication electromagnetic environment.

[0006] The technical solution adopted by the present invention to solve the above technical problem is a small sample radiation source recognition method based on a siamese network, including the following steps:

[0007] Preprocessing step: Perform data preprocessing on the received electromagnetic signal to obtain radiation source samples;

[0008] Sample expansion step: Combine the radiation source samples in pairs, and form an expanded sample dataset from the sample pairs formed by the pairwise combination. The sample dataset is divided into a training set and a test set; when there are K radiation sources and each radiation source contains N samples, the number of sample pairs in the expanded sample dataset is The label label of the sample pair ∈ {0, 1}, where 0 means the two samples of the sample pair come from different radiation sources, and 1 means the two samples of the sample pair come from the same radiation source;

[0009] Constructing and training the Siamese network model step: The Siamese network model is a Siamese convolutional neural network, including a first sub-network and a second sub-network. The two sub-networks respectively receive one sample in the input sample pair and output the feature vector of the sample; when training the Siamese network model, the balanced contrast loss is used for backpropagation and parameter update of the network parameters. The balanced contrast loss L is:

[0010]

[0011] where W is the weight of the network model, Y is the sample pair label, are the extracted features output by the first sub-network and the second sub-network respectively, D w is and in the Euclidean distance in the latent variable space, m is the target distance between different categories in the feature space of the contrast error; α is the sample balance coefficient. For a classification task with K categories,

[0012] Radiation source target recognition step: First, use the Siamese network model to obtain the mapping of each sample in the training set in the sample dataset in the feature space, and obtain the feature center of each class of samples

[0013]

[0014] is the feature vector of the jth sample belonging to the ith category;

[0015] Then, input a single sample of the test set into one sub-network of the trained Siamese network model to obtain the feature vector The feature vector and the feature vectors of K classes Perform distance measurement, and use the class with the closest distance measurement as the identified radiation source. The input is a single sample of the test set for the trained Siamese network model. After obtaining the feature center, one sub-network of the Siamese network does not need to pass during testing, and the output result of the sub-network that does not need to pass has been replaced by the feature center. Therefore, only the single sample needs to pass through the other sub-network. Simplifying the decision by finding the feature center is one of the innovations.

[0016] The beneficial effects of the present invention are as follows:

[0017] (1) Under the condition of small samples, expand the sample quantity through signal combination expansion and signal domain transformation, and cooperate with the data augmentation algorithm to enable the network to obtain more information; (2) In response to the change in the data structure after expansion, use the Siamese network to better learn the radiation source features from the expanded data; (3) Use the balanced contrast error as the network training error function, use the full traversal dataset of pairs, and achieve sample balance by the loss function, so that under the condition of small samples, the expanded data can be better used to update the network parameters; (4) After the Siamese network is trained, find the feature center of each class, and make a decision based on the feature center to improve the performance during decision-making, and complete the classification of the radiation source on the basis of network update. Description of the Drawings

[0018] Figure 1 Schematic diagram of the embodiment process.

[0019] Figure 2 Accuracy curve of the unknown target identification loss.

[0020] Figure 3 Sample recognition confusion matrix. Detailed Embodiment

[0021] The specific process of the embodiment in implementing the solution of the present invention is as Figure 1 shown:

[0022] Step 1: Perform data preprocessing on the received electromagnetic signal to obtain radiation source samples;

[0023] Step 2: Combine the radiation source samples to complete the dataset combination expansion:

[0024] Currently, many dataset expansions for small samples of neural networks have adopted the Generative Adversarial Networks (GAN). Use the trained GAN network to generate data for data expansion. However, the Generative Adversarial Networks (GAN) have disadvantages such as being prone to network overfitting during training and subsequent network performance degradation when generating data on small sample datasets.

[0025] To effectively expand data, the embodiments adopt a method of data combination to expand data: samples in the small sample dataset are combined pairwise into a new sample, that is, a new dataset is obtained by randomly combining all the radiation source samples. Samples in the new dataset are either composed of two samples from the same radiation source or two samples from different radiation sources.

[0026] When there are K radiation sources and each radiation source contains N samples. The number of samples in the small sample radiation source dataset before expansion is: K×M;

[0027] The number of samples in the radiation source dataset after signal combination and expansion according to the embodiment method is:

[0028] At the same time, since the input sample changes from a single one to a combination of a pair of samples, the learning target of the network afterwards also needs to change from extracting radiation source features and classifying to comparing radiation source features and classifying, and the structure of the network will be different from that of a conventional network.

[0029] Furthermore, the embodiments also combine traditional signal analysis methods. On the basis of the combined sample expanded dataset, all the information of the signal quadrature component (Q channel) and the in-phase component (I channel) is retained in complex form, and the short-time Fourier transform STFT (short-time Fourier transform), wavelet transform WT (wavelet transform) or bispectrum transform is used to perform dimension elevation on the sample sequence. The samples after dimension elevation can not only use a two-dimensional convolutional layer with better effect to build a network, but also the internal modulation characteristics of the radiation source signal are more obvious in the transform domain, which can further improve the performance.

[0030] Step 2 enables the purpose of expanding the sample quantity, elevating the sample dimension, expanding data and data enhancement under the condition of small samples through signal combination expansion and signal domain transformation.

[0031] Step 3: Construct a siamese network model based on balanced contrast error:

[0032] The data structure after expansion changes, and a siamese network is used to better learn the characteristics of radiation source samples from the expanded data.

[0033] The model selection and data interface of the network need to be changed accordingly. First is the change of the label corresponding to the sample: for the combined samples and the expectation of learning to compare radiation source characteristics, the label of the new sample pair label changes from the original sample set formed by K radiation source types label∈{emitter 1 ,emitter 2 ,…,emitter K} It becomes a binary label label ∈ {0, 1}, where "0" means the radiation sources from which the two samples come are different, and "1" means the radiation sources from which the two samples come are the same. The network structure selects a Siamese Network, that is, the Siamese Network contains two sub-networks with exactly the same structure: a simple Siamese Network with 2 hidden layers, which is used for binary classification that needs to output the prediction probability P. The structure of the network is replicated at the top and bottom to form a Siamese Network, and each layer shares the weight matrix to ensure that two radiation source signal samples with extremely similar features cannot be mapped to very different positions in the feature space by their respective networks. To adapt to the new sample pair structure, each input branch of the Siamese Network accepts one sample in the sample pair. For the original data set with K types of radiation sources and M samples in each type, after the combined data is expanded, the number N of sample pairs with label = 1 from the same radiation source 1 is:

[0034]

[0035] The number N of sample pairs with label = 0 from different radiation sources 0 is:

[0036]

[0037] Obviously, the unbalanced data volume of different labels is not suitable for network training. Therefore, when the traditional Siamese Network inputs the network, the occurrence probabilities of sample pairs with label = 1 and label = 0 are made equal, that is

[0038]

[0039] to improve the problem of unbalanced sample volume during network training, which is equivalent to repeating the same pair of samples to balance with the number of different pairs of samples.

[0040] The embodiment adopts the full traversal of the binary groups of the sample set to organize the data set, and the sample balance is completed by balancing the contrast error.

[0041] Step 4: Use the balanced contrast loss L to update the network parameters:

[0042] In Step 3, the forward part of the entire network is constructed, and in Step 4, the backward part of the network is constructed; the embodiment adopts the Balanced Contrastive loss to perform backpropagation and parameter update of the network parameters, and its loss function L is:

[0043]

[0044] where W is the network weight, Y is the paired label, The features extracted by the first sub-network and the second sub-network respectively, if this pair of samples belongs to the same class, Y = 1, and if they belong to different classes, Y = 0. α is the sample balance coefficient. For a classification task with K classes, when the number of samples in each class is comparable and much greater than 1:

[0045]

[0046] D w is the Euclidean distance in the latent variable space. When Y = 0, adjust the parameters to minimize the distance between and When Y = 1, if the distance between and is greater than m, no optimization is performed; if the

[0047] distance between and is less than m, then increase the distance between the two to m. Sample pairs that belong to different classes but whose pairwise distances are inherently greater than m are completely ignored by the contrastive loss, greatly reducing the computational amount. On the other hand, since no restrictions are imposed on those samples, when clustering similar samples, the high-dimensional space points to which different classes shrink are not affected by too many artificial restrictions.

[0048] After completing the training of the Siamese network, enter Step 5.

[0049] Step 5: Calculate the feature center

[0050] Based on the network update in Step 4, identify the radiation source target, that is, obtain the feature center of each class and make a decision on the radiation source type based on the feature center.

[0051] Due to the characteristic that the Siamese network needs to input sample combinations into the network for decision-making, the following optimizations are performed before network decision-making:

[0052] After the network training is completed, first use the network model to obtain the mapping of each training set sample in the feature space: where i.j indicates that the feature vector belongs to the j-th sample of the i-th class. Then, obtain the feature center of each class of samples

[0053]

[0054] When determining which class of radiation source a test sample belongs to, use the input sample to obtain its feature vector through the sub-network of the Siamese network Compare the sample feature vector with the feature vectors of K classes Perform distance measurement, and the one with the closest distance in the feature space represents that the signal sample is judged to belong to the l-th type of radiation source. Compared with the previous twin network discrimination method, this discrimination method uses the feature center to replace the extraction of known samples and test samples for combination to reduce the jitter of the decision; the feature center is obtained in advance, and only the test sample needs to pass through the network, and the decision speed is fast.

[0055] From Figure 2 it can be seen from the training set loss curve train loss and the accuracy rate train acc curve, the test set loss val loss curve and the accuracy rate val acc curve that the network training converges normally, and the accuracy rate can reach more than 90% in the case of a small number of samples. The calculation method of the accuracy rate acc for radiation source target recognition is as follows:

[0056] acc = (the number of correctly discriminated known targets + the number of correctly discriminated unknown targets) / the total number of samples.

[0057] Figure 3 is the confusion matrix for small sample recognition. The targets corresponding to the 0-8 labels are the targets to be recognized. It can be seen that the embodiment can accurately recognize the targets to be recognized in the small sample scenario.

Claims

1. A method for identifying small - sample radiation sources based on a Siamese network, characterized in that, it includes the following steps: 1) A pre - processing step: performing data pre - processing on the received electromagnetic signal to obtain radiation source samples; 2) Sample expansion step: Combine the radiation source samples in pairs, and form an expanded sample data set from the sample pairs after pairwise combination. The sample data set is divided into a training set and a test set; when there are K radiation sources and each radiation source contains N samples, the number of sample pairs in the expanded sample data set is The label of the sample pair label ∈ {0, 1}, where 0 means that the two samples of the sample pair come from different radiation sources, and 1 means that the two samples of the sample pair come from the same radiation source; 3) A step of constructing and training a Siamese network model: The Siamese network model is a Siamese convolutional neural network, including a first sub - network and a second sub - network. The two sub - networks respectively receive one sample in the input sample pair and output the feature vector of the sample. When training the Siamese network model, a balanced contrast loss is used for backpropagation of network parameters and parameter update. The balanced contrast loss L is: where W are the weights of the network model, and Y are the labels of the sample pairs, are the extracted features output by the first sub-network and the second sub-network respectively, D w is the Euclidean distance in the latent variable space with respect to, and m is the target distance between different classes in the feature space of the contrast error; α is the sample balance coefficient. For a classification task with K classes, 4) Steps for identifying radiation source targets: First, use the siamese network model to obtain the mapping of each single sample in the training set of the sample dataset in the feature space, and calculate the feature center of each class of samples is the feature vector of the j-th sample belonging to the i-th category; Then, input a single sample of the test set into a sub-network of the trained Siamese network model to obtain a feature vector Take the feature vector and perform a distance metric with the feature vectors of K classes and regard the class with the closest distance metric as the identified radiation source.

2. The method according to claim 1, characterized in that, each sample in the expanded sample dataset is in the form of a complex number retaining the quadrature component and the in - phase component, and then each sample is dimension - expanded to form two - dimensional sample data; The Siamese convolutional neural network is a Siamese convolutional neural network built by two - dimensional convolutional layers.

Images