A method for identifying individual radiation sources

Abstract

The application discloses a radiation source individual identification method, and relates to the field of target identification, which comprises the following steps: firstly, detecting air electromagnetic signals; then, extracting time-frequency feature vectors in the electromagnetic signals through fractional Fourier transform; finally, sending the time-frequency feature vectors into a trained complex residual radiation source complex individual identification network with a complex lightweight convolutional neural network and an adaptive noise reduction module for identification; the application can extract more subtle features of complex signals through fractional Fourier transform, reduce the algorithm calculation amount and storage space through the lightweight complex residual network, and hardly reduce the individual identification rate, then, the adaptive noise reduction module is added to effectively eliminate signal noise, improve the radiation source individual identification rate, and eliminate the complexity of manually setting a complex denoising algorithm.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of target recognition, and in particular to a radiation source individual identification method based on adaptive noise reduction and light-weight complex residual network. BACKGROUND

[0002] The statements in this section merely provide background information related to the present disclosure and can not constitute the prior art.

[0003] In the radio frequency signal recognition task of a complex electromagnetic environment, most deep neural networks are based on real number data for operation and description, which only utilizes real part or imaginary part information, ignoring the phase information between the real part and the imaginary part. Meanwhile, complex number data has advantages such as easier optimization, better generalization features and better representation ability. The existing radiation source individual identification algorithm research is mostly carried out on real number data, which not only discards the information amount of imaginary part data, but also discards the phase information of complex number signals, resulting in damage to the integrity and effectiveness of data features, and leading to unsatisfactory classification performance.

[0004] For the complex neural network radiation source individual identification technology, the current main methods are the complex network based on residual and the radiation source individual identification technology based on complex residual network and attention mechanism. The complex network based on residual has the advantage of directly processing complex signals, but does not make the network light-weight, and the whole network parameters and calculation amount are large. The radiation source individual identification technology based on complex residual network and attention mechanism can make the residual network pay more attention to useful subtle features, but does not provide a unified signal denoising module for the signal, and the denoising module needs to be manually set. Under the same signal-to-noise ratio or in the case of reduction, compared with the residual network using the complex denoising module, the classification probability of the residual network without using the complex denoising module will be reduced. SUMMARY

[0005] The present application aims to solve the problems that the existing feature extraction method cannot well extract the subtle features of radio frequency signals (I / Q two-way, complex signals), the light-weight technology of complex neural network is not much researched, there is no simple and practical light-weight technology, and there is no automatic denoising module for complex convolution network, which needs to manually select the appropriate denoising algorithm for different data, the setting is complex, and the denoising algorithm may be invalid after the change of the intercepted data, and the generalization is poor. The present application provides a radiation source individual identification method based on adaptive noise reduction and light-weight complex residual network, which uses the light-weight technology and the adaptive denoising technology of the complex residual network to reduce the parameter amount and the calculation amount, and improve the radiation individual identification rate, thereby solving the above problems.

[0006] The technical scheme of the present application is as follows:

[0007] A radiation source individual identification method, comprising:

[0008] Step S1: intercepting an electromagnetic signal in the air;

[0009] Step S2: extracting a time-frequency feature vector in the electromagnetic signal through fractional Fourier transform;

[0010] Step S3: sending the time-frequency feature vector into a trained complex light-weight convolutional neural network and adaptive noise reduction module complex residual radiation source complex individual identification network for identification.

[0011] Further, the order p of the fractional Fourier transform ranges from -2 to 2, and the corresponding rotation angle a ranges from -pi to pi. By using multiple different orders p, multiple time-frequency feature vectors can be generated.

[0012] Further, the complex light-weight convolutional neural network and adaptive noise reduction module complex residual radiation source complex individual identification network comprises:

[0013] Six adaptive noise reduction and light-weight complex residual modules, splicing, global average pooling, multi-dimensional to one-dimensional, and full connection layer.

[0014] Further, the adaptive noise reduction and light-weight complex residual module is composed of two complex light-weight convolution modules and one complex adaptive noise reduction module, and the complex adaptive noise reduction module performs noise reduction processing on the output of the complex light-weight convolution module.

[0015] Further, the complex light-weight convolution module comprises:

[0016] Complex batch normalization, complex activation function, and complex depth separable convolution;

[0017] The complex depth separable convolution is composed of a complex convolution with a convolution kernel of 3*3 and a group number equal to the number of input channels, and a complex convolution with a convolution kernel of 1*1.

[0018] Further, the complex adaptive noise reduction module is composed based on soft thresholding and channel attention mechanism, and specifically comprises:

[0019] Complex global average pooling, two complex full connection networks, complex CRelu, and complex sigmoid.

[0020] Further, the training in step S3 comprises:

[0021] Step A: intercepting an electromagnetic signal in the air;

[0022] Step B: extracting a time-frequency feature vector in the electromagnetic signal through fractional Fourier transform, and dividing it into a training set, a validation set, and a test set in proportion.

[0023] Step C: The complex individual identification network with the complex light-weight convolutional neural network and the adaptive noise reduction module is trained by using the training set and the validation set, and a plurality of identification networks are generated by using the hyperparameters;

[0024] Step D: The optimal identification network is selected by the classification accuracy index after the test set is input into the identification network, and is determined as the final complex individual identification network with the complex light-weight convolutional neural network and the adaptive noise reduction module.

[0025] Compared with the existing technology, the beneficial effects of the present application are:

[0026] 1. A radiation source individual identification method, which can extract more subtle features of complex signals through fractional Fourier transform, reduce algorithm calculation amount and storage space through a light-weight complex residual network, and hardly reduce individual identification rate, then add an adaptive noise reduction module to effectively eliminate signal noise and improve radiation source individual identification rate, and eliminate the complexity of manually setting a complex denoising algorithm, and the complex residual radiation source individual identification algorithm realized through the above method can improve classification probability, and in the case of 18db, for 100 types of ADS-B intercepted signals, the individual identification rate can be improved by more than 2% compared with the complex parameter network algorithm, and in the case of reducing to 15db, the individual identification rate has almost no loss compared with 18db. Therefore, the complex residual algorithm has better classification accuracy in low signal-to-noise ratio compared with the traditional complex residual network.

[0027] 2. A radiation source individual identification method, which can better extract signal subtle features when intercepting radio frequency signals (having I / Q two-way) as complex signals, a light-weight complex residual network can reduce storage space and calculation amount while hardly losing accuracy, an adaptive noise reduction module can automatically remove signal noise and retain effective features, and therefore the algorithm has good generalization ability, high classification accuracy and timeliness. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 It is a flowchart of a radiation source individual identification method;

[0029] Figure 2 It is a schematic diagram of an adaptive noise reduction and light-weight complex residual module;

[0030] Figure 3 It is a schematic diagram of a complex individual identification network with a complex light-weight convolutional neural network and an adaptive noise reduction module;

[0031] Figure 4 It is a waveform diagram of fractional Fourier transform under different orders;

[0032] Figure 5a Complex depth convolution schematic diagram;

[0033] Figure 5b Complex 1x1 convolution schematic diagram. DETAILED DESCRIPTION

[0034] It should be noted that the relational terms herein, such as first and second, and the like, are used solely to distinguish one from another entity or action without necessarily requiring or implying any actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by "comprises... a" does not, without more constraints, exclude the existence of additional identical elements in the process, method, article, or apparatus that comprises the element.

[0035] The features and advantages of the present application will be further described in the following embodiments.

[0036] Embodiment One

[0037] Please refer to Figure 1 A radiation source individual identification method, specifically comprising the following steps:

[0038] Step S1: Intercepting an air electromagnetic signal; preferably, the present embodiment takes a civil ADS-B signal (I / Q two-way) as an example for illustration, i.e., step S1 intercepts an ADS-B signal;

[0039] Step S2: extracting time-frequency feature vectors in the electromagnetic signal by fractional Fourier transform; that is, extracting the first four synchronization headers of the ADS-B signal, and converting the ADS-B time-domain signal into a time-frequency feature vector by a feature extraction method; preferably, the plurality of time-frequency feature vectors formed are divided into a training set, a validation set and a test set according to a ratio of 7:1.5:1.5; wherein the training set and the validation set are used to train the complex lightweight convolutional neural network and the adaptive noise reduction module complex residual radar source complex individual identification network, the test set is used to verify the identification accuracy of the complex lightweight convolutional neural network and the adaptive noise reduction module complex residual radar source complex individual identification network, so as to select the best complex lightweight convolutional neural network and adaptive noise reduction module complex residual radar source complex individual identification network; the fractional Fourier transform (FrFT) as a new type of generalized time-frequency analysis method can be understood as a fractional order domain representation obtained by counterclockwise rotating the original point on the time-frequency domain coordinate axis by an arbitrary angle, which not only inherits the excellent characteristics of the Fourier transform, but also has its own unique advantages, so better analysis results can be obtained in some applications;

[0040] Step S3: sending the time-frequency feature vector into the trained complex lightweight convolutional neural network and adaptive noise reduction module complex residual radar source complex individual identification network for identification; that is, sending the to-be-detected ADS-B sample into the selected best complex lightweight convolutional neural network and adaptive noise reduction module complex residual radar source complex individual identification network to determine the category of the to-be-detected target; it should be noted that the to-be-detected ADS-B sample refers to 6 time-frequency feature vectors generated by fractional Fourier transform on the ADS-B signal received by air reconnaissance.

[0041] In this embodiment, specifically, the order p of the fractional Fourier transform is in the range of -2 to 2, and the corresponding rotation angle a is in the range of -p to p, and a plurality of time-frequency feature vectors can be generated by a plurality of different orders p; when p = 1, the FrFT is equivalent to the Fourier transform; when p = 0, the result of the FrFT is the same as the time-domain result;

[0042] It should be noted that when the fractional Fourier transform order p of the four synchronization header original signals of the ADS-B signal (complex signal of I / Q two-way) is 0, 0.25, 0.75, 1, 1.25 and 1.75, 6 different time-frequency feature vectors are obtained, and each time-frequency feature vector has a length of 1024 dimensions (complex vector); that is, the training set, the validation set and the test set mainly refer to the time-frequency feature sample set obtained by fractional Fourier transform, the total number of sample sets is 600, the number of training sets is 420, the number of validation sets is 90, and the number of test sets is 90, each sample contains 6 time-frequency feature vectors, such as Figure 4, as shown.

[0043] In this embodiment, specifically, as shown in the formula (1), the complex light weight convolutional neural network and the adaptive noise reduction module complex residual source individual identification network comprises: Figure 3

[0044] Six adaptive noise reduction and light weight complex residual modules, splicing, global average pooling, multi-dimensional to one-dimensional, and full connection layer.

[0045] In this embodiment, specifically, the adaptive noise reduction and light weight complex residual module is composed of two complex light weight convolutional modules and one complex adaptive noise reduction module, and the complex adaptive noise reduction module performs noise reduction processing on the output of the complex light weight convolutional module.

[0046] In this embodiment, specifically, the complex light weight convolutional module comprises:

[0047] Complex batch normalization, complex activation function, and complex depth separable convolution;

[0048] The complex depth separable convolution is composed of a complex convolution with a convolution kernel of 3*3 and a group number of input channel numbers and a complex convolution with a convolution kernel of 1*1.

[0049] In this embodiment, specifically, the complex adaptive noise reduction module is composed based on soft thresholding and channel attention mechanism, and specifically comprises:

[0050] Complex value global average pooling, two complex value full connection networks, complex value CRelu, and complex value sigmoid; preferably, all convolution channel numbers are set to 64.

[0051] It should be noted that the light weight complex residual module is the complex light weight convolutional neural network, and the complex light weight convolutional neural network also refers to the complex depth separable convolution network, and the adaptive noise reduction module refers to a technology of applying soft thresholding technology to the complex residual network for automatic noise reduction.

[0052] The idea of the complex depth separable convolution network is to divide a complex convolution operation into two steps of complex depth convolution and complex 1x1 convolution. The complex depth convolution refers to convolution of each channel separately, and the number of input feature maps is the same as the number of output feature maps; the complex 1x1 convolution is mainly used for channel dimension increasing and dimension decreasing and channel fusion of the complex depth convolution feature; and a complex depth separable convolution diagram is as shown in Figure 5a and Figure 5b .

[0053] It should be noted that the two-dimensional complex convolution is taken as a column, and the complex input feature map of the complex convolution layer is represented as V=V R +iV​I V R and V I These represent the real and imaginary parts of the complex input feature map, respectively; the complex convolution kernel is represented as W = W R +iW I W R and W I These represent the real part and the imaginary part, respectively. The output complex feature map is represented as U = U R +U I The complex convolution is represented as follows:

[0054]

[0055] Where * denotes a real-valued convolution filter, and the convolution assumes the input is W. in ×H in ×C in The output is W out ×H out ×C out The kernel size is k×k×2, and the number of parameters in the complex convolution is 2×C. in ×k×k×C out The computational cost of complex convolution is equivalent to that of four real convolutions; therefore, the computational cost of complex depthwise separable convolution is 4 × C. in ×k×k×C out ×W out ×H out .

[0056] Depend on Figure 5a and Figure 5b It can be seen that the computational cost of parameters for complex depthwise separable convolution is 2 × (C in ×k×k×1+C out ×C in The computational cost of complex depthwise separable convolution is 4×(C × 1× 1). in ×k×k×1+C out ×C in ×1×1)×W out ×H out Therefore, the number of parameters in a complex depthwise separable convolution is equal to that of a complex convolution. The computational cost is Because C in the network structure out >>k 2 The size k of the convolution kernel space is generally taken as 3, that is, the computational complexity of complex depthwise separable convolution can be significantly reduced by 1 / 8 to 1 / 9 of the computational complexity of complex convolution; similarly, the number of parameters of complex depthwise separable convolution modules can be significantly reduced by 1 / 8 to 1 / 9 of the number of parameters of complex convolution layers.

[0057] It should be noted that the complex adaptive noise reduction module is mainly composed of soft thresholding and channel attention mechanism and the like. The soft thresholding is the most commonly used technology in signal noise reduction algorithm, which converts the original signal to the feature domain, wherein the noise and redundant part of the signal becomes near zero value, and finally a threshold is set to convert the near zero feature to zero. The conversion formula is:

[0058]

[0059] Wherein x is the input feature, y is the output feature, and τ is the threshold value which is positive;

[0060] The traditional soft thresholding technology needs to manually design the threshold size, which is troublesome and has large workload, and in the embodiment, a complex neural network for automatically designing threshold is designed in combination with the complex residual network and the channel attention mechanism; the complex adaptive noise reduction module is as shown in Figure 2 .

[0061] In the module, first, the real part and the imaginary part of the feature map output by the last layer network of the residual unit are respectively taken absolute value and global average pooling (GAP), to obtain a one-dimensional vector with the same number of convolution kernels as the last layer. The specific operation is as follows:

[0062] CGAP = GAP (|U R |) + iGAP (|U I |)

[0063] The one-dimensional vector is input into two layers of complex value full connection network, and the complex value full connection calculation reference formula (1) is complex value convolution operation, wherein CReLU (U) = CReLU (U R ) + iCReLU (U I ), and a complex value Sigmoid function is applied at the end of the two layers of complex value FC network, and the scaling parameter is normalized according to the following formula:

[0064]

[0065] Wherein, d i is the output value of the i-th neuron of the last layer full connection network of the shrinkage module, real (d i ) and imag (d i ) represent the real part and the imaginary part of d i respectively, α i is the result of normalizing d i , and the threshold value is calculated as follows:

[0066] τ i = real (α i ) · real (Average {g m,n,i}m,n )+iimag(α i )·imag(Average{g m,n,i} m,n ) (4)

[0067] where τ i represents the threshold value of the i-th channel of the feature map, m and n represent the width and height of the feature map, and the average value of {g} is multiplied by the scaling parameter α to obtain the complex threshold value, which is to limit the threshold value to remain within a reasonable range. If the threshold value is greater than the maximum absolute value of the feature map, the output of the soft threshold will be 0.

[0068] As Figure 2 shown, the complex threshold value is equal to the noise reduction on the real part feature vector and the imaginary part feature vector respectively, and then spliced into a complex form.

[0069] In this embodiment, specifically, the training in step S3 comprises:

[0070] Step A: intercepting the electromagnetic signal in the air;

[0071] Step B: extracting the time-frequency feature vector in the electromagnetic signal by fractional Fourier transform, and dividing it into training set, validation set and test set in proportion;

[0072] Step C: training the complex individual identification network of the complex residual radiation source with the complex light-weight convolutional neural network and the adaptive denoising module by using the training set and the validation set, and generating multiple identification networks by using the hyperparameters;

[0073] Step D: selecting the optimal identification network as the final complex individual identification network of the complex residual radiation source with the complex light-weight convolutional neural network and the adaptive denoising module by using the classification accuracy index of the test set input identification network.

[0074] The above-described embodiments only express the specific implementation of the present application, and the description is more specific and detailed, but it cannot be understood as a limitation on the protection scope of the present application. It should be noted that for ordinary skilled persons in the art, without departing from the technical concept of the present application, a number of modifications and improvements can be made, which are within the protection scope of the present application.

[0075] This background section is provided to generally present the context of the application, the work of the current named inventors, the work described in this background section to the extent that it is described, and the work described in this section at the time of filing, neither expressly nor implicitly, is recognized as prior art of the present application.

Claims

1. A method of identifying individual radiation sources, characterized by, The method comprises the steps of: Step S1: intercepting electromagnetic signals in the air; Step S2: extracting time-frequency feature vectors in the electromagnetic signals through fractional Fourier transform; Step S3: sending the time-frequency feature vectors into a trained complex residual emitter complex individual identification network with a complex lightweight convolutional neural network and an adaptive noise reduction module for identification; The order p of the fractional Fourier transform ranges from -2 to 2, and the corresponding rotation angle a ranges from -p to p. By using multiple different orders p, multiple time-frequency feature vectors can be generated. The complex residual emitter complex individual identification network with the complex lightweight convolutional neural network and the adaptive noise reduction module comprises: Six adaptive noise reduction and lightweight complex residual modules, splicing, global average pooling, multi-dimensional to one-dimensional, and a fully connected layer. The adaptive noise reduction and lightweight complex residual module is composed of two complex lightweight convolution modules and a complex adaptive noise reduction module. The complex adaptive noise reduction module performs noise reduction processing on the output of the complex lightweight convolution module. The complex lightweight convolution module comprises: A complex batch normalization, a complex activation function, and a complex depth separable convolution. The complex depth separable convolution is composed of a complex convolution with a 3*3 convolution kernel and a group number equal to the number of input channels, and a complex convolution with a 1*1 convolution kernel.

2. The method of claim 1, wherein, The complex adaptive noise reduction module is based on soft thresholding and channel attention mechanism, and specifically comprises: A complex global average pooling, two complex fully connected networks, a complex CRelu, and a complex sigmoid.

3. The method of claim 1, wherein, The training in step S3 comprises the steps of: Step A: intercepting electromagnetic signals in the air; Step B: extracting time-frequency feature vectors in the electromagnetic signals through fractional Fourier transform, and dividing them into a training set and a validation set according to a proportion; Step C: training the complex residual emitter complex individual identification network with the complex lightweight convolutional neural network and the adaptive noise reduction module using the training set, and generating multiple identification networks using hyperparameters; Step D: selecting the optimal identification network as the final complex residual emitter complex individual identification network with the complex lightweight convolutional neural network and the adaptive noise reduction module according to the classification accuracy of the identification network after inputting the validation set.

Images