Frequency domain robust direction finding method based on coherent subspace deep learning

Abstract

The application discloses a frequency domain robust direction finding method based on coherent subspace deep learning, and belongs to the technical field of radio direction finding, and the specific steps comprise: step one, constructing a direction finding model based on coherent subspace deep learning; step two, carrying out digital sampling to obtain a digital sampling signal; step three, carrying out frequency domain transformation and subband decomposition; step four, focusing processing, signal subspace of different subbands is transformed to reference signal subspace; step five, convolution processing, a convolution layer adopts a local connection mode, and nonlinear feature mapping of data is realized through convolution operation; step six, pooling processing, based on the shift-invariant characteristic pooling processing, the input feature mapping is divided into multiple non-overlapping regions; step seven, coherent subspace wideband direction finding, on the basis of learning and training, frequency domain robust direction finding based on coherent subspace deep learning is realized; the coherent subspace focusing preprocessing is increased, and intelligent direction finding of a wideband signal is realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of radio direction finding, and specifically relates to a frequency domain robust direction finding method based on coherent subspace deep learning. BACKGROUND

[0002] Direction finding of a radiation source target is that electronic reconnaissance equipment estimates the direction of arrival of a target signal through signal processing by using the received target signal. The direction finding method for a radiation source target mainly includes amplitude direction finding, phase direction finding, spatial spectrum estimation direction finding and the like, and these methods are all traditional non-intelligent direction finding methods.

[0003] At present, intelligent direction finding methods have become a research hotspot. The patent "arbitrary array direction of arrival estimation method based on deep learning" proposes an array signal direction finding method based on a convolutional neural network (CNN). The method is aimed at a given multi-element array. By extracting the phase difference and other characteristic information of the sampling data of each array element, the method realizes fast and high-precision direction finding of a radiation source target on the basis of deep learning.

[0004] However, the above-mentioned CNN direction finding method is designed for a fixed working frequency. However, a wireless direction finding system usually needs to adapt to a relatively wide frequency range (such as several MHz to several tens of GHz). In the face of wide-band direction finding requirements, when the frequency range is relatively narrow, the above-mentioned CNN direction finding method needs to establish a separate function model for each carrier frequency in the frequency range for training, which greatly increases the training cost and storage requirement. When the frequency range is very wide, the number of function models and training samples are significantly increased, which leads to difficulty in neural network training and even difficulty in convergence.

[0005] In order to solve the intelligent direction finding problem of frequency domain wideband signals, the application proposes a frequency domain robust direction finding method based on coherent subspace deep learning. SUMMARY

[0006] In order to solve the problems existing in the above-mentioned scheme, the application provides a frequency domain robust direction finding method based on coherent subspace deep learning. The idea of the coherent subspace method is applied to wideband CNN direction finding. First, the coherent subspace focusing preprocessing is used to focus the wideband signal subspace to the reference signal subspace, and then the convolutional neural network narrowband signal direction finding model is used to realize frequency-independent wideband signal robust direction finding.

[0007] The purpose of the application can be achieved by the following technical solutions:

[0008] The frequency domain robust direction finding method based on coherent subspace deep learning includes the following specific steps:

[0009] Step 1: Construct a direction finding model based on coherent subspace deep learning;

[0010] Step two: digital sampling is performed to obtain a digital sampling signal;

[0011] Step three: frequency domain transformation and sub-band decomposition are performed;

[0012] Step four: focus processing, signal subspace transformation of different sub-bands to reference signal subspace;

[0013] Step five: convolution processing;

[0014] Step six: pooling processing, based on the shift-invariant characteristic, the input feature map is divided into multiple non-overlapping regions;

[0015] Step seven: coherent subspace wideband direction finding, based on learning and training, frequency domain robust direction finding based on coherent subspace deep learning is realized.

[0016] Further, the coherent subspace deep learning direction finding model is divided into coherent subspace focus preprocessing and CNN direction finding processing.

[0017] Further, the digital sampling is: the radiation source signals received by each array element are digitally sampled and processed to obtain a digital sampling signal .

[0018] Further, the method for performing frequency domain transformation and sub-band decomposition is:

[0019] Based on the digital sampling signal , Fourier transformation is performed to obtain:

[0020] ;

[0021] The wideband signal is decomposed into multiple narrow sub-bands, and for a narrow sub-band with a center frequency in the wideband, there is:

[0022] .

[0023] Further, the focus processing method includes:

[0024] Let , obtain:

[0025] ;

[0026] The focus matrix is designed according to the above formula, the signal subspace of different sub-bands is transformed to the reference signal subspace , and the data vector after focus processing is:

[0027] .

[0028] Further, the convolution layer in the convolution processing adopts a local connection mode, and realizes nonlinear feature mapping of data through a convolution operation.

[0029] Compared with the prior art, the beneficial effects of the application are that the coherent subspace focusing preprocessing is added, and the intelligent direction finding of the wideband signal is realized; the direction finding frequency domain robustness is good, and when the frequency range reaches 8 times of the frequency range, the direction finding precision deterioration degree is less than 2%. BRIEF DESCRIPTION OF DRAWINGS

[0030] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0031] Fig. 1 The flow chart of the direction finding based on the coherent subspace deep learning of the present application is shown in

[0032] Fig. 2 The model schematic diagram of the direction finding based on the coherent subspace deep learning of the present application is shown in DETAILED DESCRIPTION

[0033] The technical solutions of the present application will be described clearly and completely in combination with the embodiments. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0034] As shown in Figs. 1-2 , the frequency domain robustness direction finding method based on the coherent subspace deep learning is to solve the intelligent direction finding problem of the frequency domain wideband signal. The method applies the idea of the coherent subspace method to the wideband CNN direction finding. First, the coherent subspace focusing preprocessing is used to focus the wideband signal subspace to the reference signal subspace, and then the convolution neural network narrowband signal direction finding model is used to realize the frequency-independent wideband signal robustness direction finding. The specific steps include:

[0035] Step 1: Constructing the direction finding model based on the coherent subspace deep learning;

[0036] The direction finding model based on the coherent subspace deep learning is divided into two parts. The first part is the coherent subspace focusing preprocessing, and the second part is the CNN direction finding processing. The design of the direction finding model based on the coherent subspace deep learning is shown in Fig. 2

[0037] ​The CNN uses a sigmoid activation function, and an average pool layer is followed after each convolution layer, and a cross-entropy loss function is used in the classification layer.

[0038] Step two: digital sampling is performed to obtain a digital sampling signal;

[0039] The radiation source signal received by each array element is subjected to digital sampling processing to obtain a digital sampling signal .

[0040] Step three: frequency domain transformation and sub-band decomposition are performed;

[0041] Based on Fourier transform, specifically:

[0042] Based on the digital sampling signal , Fourier transform is performed to obtain:

[0043] ;

[0044] The wideband signal is decomposed into multiple narrow sub-bands, and for a narrow sub-band with a center frequency in the wideband, there is:

[0045] .

[0046] Step four: focusing processing, signal subspace transformation of different sub-bands to reference signal subspace;

[0047] To realize frequency domain focusing, it is necessary to make , and the following can be obtained:

[0048] ;

[0049] The focusing matrix is designed according to the above formula , the signal subspace of different sub-bands is transformed to the reference signal subspace , and the data vector after focusing processing is:

[0050] .

[0051] Step five: convolution processing, the convolution layer uses a local connection method to realize nonlinear feature mapping of data through convolution operation;

[0052] The convolution layer mainly analyzes each small block of data more deeply to abstract higher-level features. The convolution layer uses a local connection method to realize nonlinear feature mapping of data through convolution operation of different convolution kernels and input data, thereby realizing feature extraction.

[0053] Step six: pooling processing, based on the shift-invariant characteristics, the input feature map is divided into multiple non-overlapping regions;

[0054] The pooling layer realizes data compression through downsampling operation, which is used to reduce the dimension of the feature after convolution and reduce the number of neurons required by the network. The pooling processing is based on the shift-invariant characteristics, which divides the input feature map into multiple non-overlapping regions. After the pooling processing, the automatic extraction of the target angle feature vector is realized.

[0055] Step seven: coherent subspace broadband direction finding, based on learning and training, frequency domain robust direction finding based on coherent subspace deep learning is realized.

[0056] After focusing processing, convolution processing and pooling processing, the full connection layer couples the generated features to different spatial angles, and realizes frequency domain robust direction finding based on coherent subspace deep learning on the basis of learning and training.

[0057] Coherent subspace method:

[0058] Let the wideband signal vector of the radiation source be The wideband signal received by the direction finding system is The noise is Then we have: ; Where, is the direction vector.

[0059] Let the number of sampling shots be N, The Fourier transform of

[0060] ;

[0061] For the narrow sub-band with the center frequency in the wideband, we have:

[0062] ;

[0063] For different sub-bands, there is a matrix such that:

[0064] .

[0065] The matrix is called the focusing matrix, where is a selected reference frequency. Obviously, through focusing, the signal subspace of different sub-bands can be transformed into the reference signal subspace .

[0066] The converted data vector obtained by wideband focusing is:

[0067] .

[0068] The essence of the above formula for CSM focusing is to construct a consistent direction vector for each wideband signal source without changing the signal information, in other words, to focus the independent direction vector at each frequency point into the direction vector at a single frequency point.

[0069] In order to verify the performance of the method, by sampling 11 elements of a uniform linear array, for different frequency range of the radiation source signal, the patent "arbitrary array direction of arrival estimation method based on deep learning" (hereinafter referred to as method 1) and the method of the application are applied respectively, statistical test of direction finding is carried out, and the statistical results of the direction finding accuracy deterioration degree of different frequency range signals are shown in Table 1.

[0070] Table 1 Direction finding accuracy deterioration degree of different frequency range signals

[0071]

[0072] The experimental results show that when the frequency range reaches 8 octaves, the direction finding accuracy deterioration degree of method 1 is greater than 10%, while the direction finding accuracy deterioration degree of the method is less than 2%, it can be seen that compared with method 1, the direction finding robustness of the method of the application is obviously improved.

[0073] The above formulas are all dimensionless values, the formulas are obtained by software simulation of a large amount of data to obtain a formula closest to the actual situation, and the preset parameters and the preset threshold in the formula are set by the person skilled in the art according to the actual situation or obtained by a large amount of data simulation.

[0074] The working principle of the application is as follows:

[0075] Through the coherent subspace focusing preprocessing, the wideband signal subspace is focused to the reference signal subspace, and then the convolutional neural network narrowband signal direction finding model is used to realize the frequency-independent wideband signal robust direction finding.

[0076] The above examples are only used to illustrate the technical method of the present application but not limit the present application. Although the present application is described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present application can be modified or equivalently replaced without departing from the spirit and scope of the technical method of the present application.

Claims

1. A frequency domain robust direction finding method based on coherent subspace deep learning, characterized in that, The specific steps include: Step one: constructing a coherent subspace deep learning direction finding model; Step two: digital sampling to obtain a digital sampling signal; Step three: frequency domain transformation and subband decomposition; Step four: focusing processing, transforming the signal subspace of different subbands to the reference signal subspace; Step five: convolution processing; Step six: pooling processing, based on the shift-invariant characteristic, the input feature mapping is divided into multiple non-overlapping regions; Step seven: coherent subspace wideband direction finding, based on the frequency domain robustness direction finding of the coherent subspace deep learning on the basis of learning and training; The digital sampling is performed as follows: the radiation source signals received by each array element are subjected to digital sampling processing to obtain digital sampling signals ; The method for frequency domain transformation and subband decomposition is: Based on digitized sampling signals a Fourier transform is performed to obtain ； The wideband signal is decomposed into a plurality of narrow subbands, for a narrow subband with a center frequency of within the wideband, there is: ； The focusing processing method includes: Let , we obtain: ； The focusing matrix is designed according to the above formula The signal subspace of different subbands is transformed to the reference signal subspace The signal subspace of different subbands is transformed to the reference signal subspace The data vector after the focusing processing is: ； In the convolution processing, the convolution layer adopts a local connection mode, and the nonlinear feature mapping of data is realized through convolution operation.

2. The frequency domain robust direction finding method based on coherent subspace deep learning of claim 1, wherein, The coherent subspace deep learning direction finding model is divided into coherent subspace focusing preprocessing and CNN direction finding processing.

Images