Electromagnetic signal target recognition system based on transformer feature extraction

Abstract

The application discloses an electromagnetic signal target recognition system based on a Transformer feature extraction, which comprises a data processing module for image-like data conversion, a feature extractor module based on a lightweight Transformer and a feature classifier module; the data processing module generates a fixed-size time-frequency matrix from each pulse signal through a short-time Fourier transform, thereby converting an electromagnetic signal target recognition task into an image-like classification task; the feature extractor module extracts global features through a Mobile-Vit model according to the fixed-size time-frequency matrix, obtains a one-dimensional feature activation vector, and obtains the feature activation vector through a full connection layer; and the feature classifier module calculates the probability of a class to which a sample belongs according to the feature activation vector through the distance between the feature of each known class electromagnetic signal target and the reciprocal point, so that more comprehensive features and more accurate recognition rates are achieved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of electromagnetic signal processing, and is an electromagnetic signal target recognition system based on a Transformer feature extraction and applied to an open set scene. BACKGROUND

[0002] In a traditional closed set scene, the electromagnetic signal target categories used in the test stage are often the same as those used in the training stage, while in an open set scene, the number of electromagnetic signal target categories cannot be predicted, so it is difficult to collect and extract feature information of all electromagnetic signal targets in the training stage. The electromagnetic signal target categories appearing in the training stage are referred to as known class electromagnetic signal targets, and the electromagnetic signal target categories appearing in the test stage and not appearing in the training stage are referred to as unknown class electromagnetic signal targets. Extracting only the features of the known class electromagnetic signal targets cannot accurately identify the unknown class electromagnetic signal targets.

[0003] At present, most solutions to this scene are to project each electromagnetic signal target into a high-dimensional feature space, and separate the known class electromagnetic signal target feature space and the unknown class electromagnetic signal target feature space in the space. The reciprocal point-based method defines a reciprocal point for each electromagnetic signal target as a potential representation of the unknown class electromagnetic signal target and other electromagnetic signal targets, so as to separate the known class electromagnetic signal target feature space and the unknown class electromagnetic signal target feature space. SUMMARY

[0004] The application proposes an electromagnetic signal target recognition system based on a Transformer feature extraction to solve the problems that the prior art is only applicable to a closed set scene and does not make full use of the time sequence relationship in the time-frequency matrix generated by a short-time Fourier transform. The electromagnetic signal target recognition problem is converted into an image-like classification problem by using a short-time Fourier transform, a feature extractor integrates a lightweight Transformer model to extract global features of the image-like time-frequency matrix, and a center loss is fused to compress the unknown class electromagnetic signal target feature space, so that a method with higher recognition accuracy than the prior art is realized.

[0005] The application is implemented by the following technical solutions:

[0006] The application relates to a kind of electromagnetic signal target identification systems based on Transformer feature extraction, comprising: data processing module for class image data conversion, feature extractor module based on lightweight Transformer and feature classifier module, wherein: data processing module generates fixed size time-frequency matrix from each pulse signal by short-time Fourier transform, thereby converting electromagnetic signal target identification task into class image classification task;Feature extractor module extracts global feature by Mobile-Vit model according to fixed size time-frequency matrix, obtains one-dimensional feature activation vector, and obtains feature activation vector by full connection layer;Feature classifier module calculates the probability of the class to which the sample belongs according to the distance between the feature activation vector and the reciprocal point of each known electromagnetic signal target feature, and realizes high-precision identification.

[0007] The data processing module includes an energy detection unit and a short-time Fourier transform unit, wherein: the energy detection unit removes noise signals in the original electromagnetic signal, and the remaining pulse signals are directly spliced; the short-time Fourier transform unit divides the long signal of the entire class into a single small signal of fixed size, and then performs short-time Fourier transform on each small signal to generate a time-frequency matrix of fixed size class image, obtaining frequency, phase and amplitude information in the signal.

[0008] The data processing module includes energy detection denoising and short-time Fourier transform of the original electromagnetic signal. Wherein: energy detection is to set a threshold, and the part of the original signal with a signal value less than the threshold is removed as noise, and the remaining signal is spliced as usable pulse signal. After generating the usable pulse signal, the entire electromagnetic signal is divided into a single signal with a fixed length of 2000, and the single signal is subjected to short-time Fourier transform to generate a time-frequency matrix with a size of 2*64*64. For the convenience of direct model training, the first and second dimensions are averaged and spliced into the third dimension, and the final time-frequency matrix has a size of 3*64*64.

[0009] The feature extractor module includes a plurality of MobileNetv2 structures, a plurality of Transformer structures and a plurality of layers of full connection layers, wherein: the MobileNetv2 structure undertakes feature extraction and convolution in the model, the sample feature is reduced after MobileNetv2, then the global feature is extracted by using the Transformer structure to obtain more complete sample feature;Full connection layer regards the output of each layer as feature activation vector, splices the feature activation vector output in front, and finally obtains more powerful feature, which is convenient for classification of feature classifier.

[0010] The MobileNetv2 structure convolves the sample features, converts large samples into small sample features while extracting features; the Transformer structure extracts global features from samples with small data volume; the Mobile-Vit uses the MobileNetv2 and the Transformer structure to extract global features, and avoids the disadvantage of large calculation caused by the introduction of the Transformer structure. The Mobile-Vit of the feature extractor finally inputs a time-frequency matrix with a size of 3*64*64, outputs a feature activation vector with a size of 1*512, and is spliced with a 1*512 vector through two fully connected layers to form a feature activation vector with a size of 1*768, which is input into the feature classifier.

[0011] The feature classifier module includes a classification loss unit and a center loss unit, wherein: the classification loss unit calculates the distance of the reciprocal point of the known class electromagnetic signal target from the corresponding feature, and then takes the distance as the input parameter of the cross-entropy loss to calculate the cross-entropy loss; the center loss unit maintains a clustering center for each known class electromagnetic signal target in the iteration process, and the loss is calculated by the distance of the sample from the clustering center, and the greater the distance, the greater the loss, and finally the classification loss and the center loss are minimized through model training to solve the problem.

[0012] The present application relates to a kind of electromagnetic signal target identification method in open set scene based on above-mentioned system, through data processing module to the electromagnetic signal data received by energy detection, extract the pulse signal with signal information, obtain the time-frequency matrix with electromagnetic signal frequency, phase, amplitude information after short-time Fourier transform, and it is converted into image classification problem and output to feature extractor module;Feature extractor module extracts the feature activation vector and the feature activation vector of full connection output splicing of sample by convolution and Transformer processing;Feature classifier module separates known class electromagnetic signal feature space and unknown class electromagnetic signal feature space under the constraint of the classifier loss function of center loss fusion, realizes the identification of unknown class electromagnetic signal.

[0013] Technical effects

[0014] The present application converts electromagnetic signal target recognition problem into image classification problem, can directly use powerful neural network in deep learning to classify, and introduces center loss on the basis of reciprocal point to compress known class electromagnetic signal feature space, improves the recognition ability of method in open set scene, compared with baseline algorithm, the recognition rate of unknown class electromagnetic signal is improved by 6.7%, and the recognition rate of known class electromagnetic signal is improved by 5.5%. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a flowchart of the present application;

[0016] Figure 2 is a signal processing schematic diagram of the present application;

[0017] Figure 3 is a feature extractor module structure diagram of the present application;

[0018] Figure 4 is a Transformer structure diagram of the present application;

[0019] Figure 5 is a MobileNetv2 structure diagram of the present application. DETAILED DESCRIPTION

[0020] The data of the embodiment is collected from electromagnetic signal data of ten different models of mobile phones, and a class image time-frequency matrix is generated through a data processing module, which is classified according to the belonging category to generate a mobile phone electromagnetic signal data set. In the embodiment, 10000 time-frequency matrices are generated for each electromagnetic signal target, 6 electromagnetic signal targets are randomly selected in the training process, 5000 are selected from each target as a training set, and the remaining data and the data of another 4 electromagnetic signal targets are used as a test set. The full connection layer in the feature extractor is two layers, and the output of the first layer is a feature with a size of 1*256. In the training process, Adam is selected as the model optimizer, the batch size is set to 128, the training epoch is set to 100, and the model parameters are obtained through back propagation. In the test stage, the test is carried out in units of test categories, and the recognition accuracy of each electromagnetic signal category is counted.

[0021] As shown in Figure 1 , the present application is a method for electromagnetic signal target recognition in an open set scene, comprising:

[0022] Step one: energy detection is performed on the original electromagnetic signal data, and the noise signal smaller than the threshold value in the signal is removed, and the remaining usable pulse signal is spliced;

[0023] Step two: the usable pulse signal is cut into equal length signals, and a short-time Fourier transform is performed on each signal to generate a single time-frequency matrix sample;

[0024] Step three: input the time-frequency matrix into the Mobile-Vit feature extractor, output the feature activation vector of the time-frequency matrix sample, and splice the feature activation vector output by the Mobile-Vit and the feature activation vector output by the full connection layer into a new feature activation vector;

[0025] Step four: the feature activation vector passes through the loss function, so that the model parameters are trained as the state of the separable feature activation vector in the back propagation process, realizing the separation of the known class electromagnetic signal feature space and the unknown class electromagnetic signal feature space;

[0026] Step five: after the time-frequency matrix sample input in the test stage is extracted by the feature extractor, the feature is input into the feature classifier to directly output the result;

[0027] Step six: statistics of the recognition result of each electromagnetic signal target.

[0028] As Figure 3 shown, the open set scene electromagnetic signal target recognition task includes a data processing module, a feature extraction module, and a feature classification module. The data processing module generates a fixed size time-frequency matrix from one-dimensional electromagnetic signal denoising, and converts the electromagnetic signal target recognition problem into an image classification problem. Then, the Transformer series model with strong feature extraction capability is used to realize better feature extraction (see table for specific effects). Finally, the feature classifier integrates the classification loss and the center loss to realize the separation of the features and the compression of the known class feature space. An open set scene electromagnetic signal target recognition method with high recognition capability is realized.

[0029] Since the unknown class and the known class are randomly divided, in order to avoid the contingency caused by random classification, the average value of 5 experimental results is taken in the experiment, and the method is compared with other methods, and the ablation experiment results are added to show the advantages of the method.

[0030] Table 1

[0031]

[0032] The method is superior to common open set recognition algorithms, because it is improved on the basis of the SOTA method ARPL. As can be seen from the above experiments, after replacing the feature extractor, the recognition rate of the known class electromagnetic signal target of the model is obviously improved, and after integrating the center loss of the feature classifier, the recognition rate of the unknown class electromagnetic signal target is obviously improved. Compared with the prior art, each index of the present application reaches the highest level.

[0033] Compared with the prior art, the present method converts one-dimensional electromagnetic signal data into a time-frequency matrix, converts the problem into an image problem, and facilitates the use of neural networks with strong feature extraction capability. Moreover, on the basis of separating the known class electromagnetic signal feature space and the unknown class electromagnetic signal feature space at the reciprocal point, the center loss is integrated to realize the compression of the known class electromagnetic signal feature space, and improve the recognition performance of the method.

[0034] The above specific embodiments can be partially adjusted in different ways by those skilled in the art without departing from the principles and purposes of the present application, the protection scope of the present application is subject to the claims and is not limited by the above specific embodiments, each implementation within the scope is subject to the present application.

Claims

1. A system for electromagnetic signal target recognition based on Transformer feature extraction, characterized in that, The application relates to a data processing module for image-like data conversion, a feature extractor module based on a lightweight Transformer and a feature classifier module. The data processing module comprises an energy detection unit and a short-time Fourier transform unit, wherein the energy detection unit removes noise signals in original electromagnetic signals, and remaining pulse signals are directly spliced; the short-time Fourier transform unit divides long signals of the whole category into single small signals of a fixed size, then performs short-time Fourier transform on each small signal to generate a time-frequency matrix of a fixed-size image-like signal, and frequency, phase and amplitude information in the signal is obtained. The feature extractor module comprises a plurality of MobileNetv2 structures, a plurality of Transformer structures and a plurality of full connection layers, wherein the MobileNetv2 structure undertakes feature extraction and convolution in the model, sample features are reduced after the MobileNetv2 structure, then global features are extracted by the Transformer structure to obtain more complete sample features; each layer of the full connection layer regards the output as a feature activation vector, the feature activation vector is spliced with a feature activation vector output in the front, and finally more powerful features are obtained, which are convenient for classification of the feature classifier.

2. The electromagnetic signal target recognition system based on the Transformer feature extraction according to claim 1, characterized in that, The feature classifier module comprises a classification loss unit and a center loss unit, wherein the classification loss unit calculates the distance of each feature from the reciprocal point of the known category electromagnetic signal target, then the distance is taken as an input parameter of cross-entropy loss to calculate the cross-entropy loss; the center loss unit maintains a clustering center for each known category electromagnetic signal target in the iteration process, and the loss is calculated by the distance of the sample from the clustering center; the greater the distance, the greater the loss; finally, the problem is solved by minimizing the classification loss and the center loss through model training. 3.The electromagnetic signal target recognition system based on the Transformer feature extraction according to claim 1, characterized in that, When the pulse signals after invalid noise signals are removed by the threshold value are spliced again, the relative position relationship of the original front and back positions is maintained; when the short-time Fourier transform is performed, the length of a single small signal sample is 2000, a time-frequency matrix of 2*64*64 is generated, the first and second dimensions are spliced to obtain the third dimension, and finally the size is 3*64*64. 4.The system of claim 1, wherein, The weight ratio of the classification loss and the center loss is 5:

4.

5. The electromagnetic signal target recognition system based on the Transformer feature extraction according to claim 3, characterized in that, ​ 6. A method for target recognition of electromagnetic signals in open set scenarios based on the system of any one of claims 1-5. The electromagnetic signal data received is subjected to energy detection by the data processing module, and the pulse signal with signal information is extracted, and after a time-frequency matrix with electromagnetic signal frequency, phase and amplitude information is obtained through short-time Fourier transform, it is converted into an image-like classification problem and output to the feature extractor module; The feature extractor module extracts the feature activation vector of the sample and the feature activation vector spliced by full connection output through convolution and Transformer processing; the feature classifier module separates the known class electromagnetic signal feature space and the unknown class electromagnetic signal feature space under the constraint of the classifier loss function with the center loss, and realizes the identification of the unknown class electromagnetic signal.

Images