A method for identifying a model of a UAV based on a distortionless spectrum and a light network

Abstract

The application discloses a kind of based on distortionless spectrogram and lightweight network's unmanned aerial vehicle model identification method, belong to signal processing field.First of all, the minimum distortionless response spectrogram of unmanned aerial vehicle measurement and control signal or downlink signal is characterized to construct feature database, then feature database is proportionally training set and test set division, finally through specially constructed based on grouping convolution and channel rearrangement lightweight neural network model carries out feature map identification.Compared with the existing recognition method based on wavelet scattering transform, the minimum distortionless response spectrogram based on the present application can effectively improve the recognition rate of unmanned aerial vehicle equipment, and the specially constructed lightweight neural network can effectively reduce the computational complexity.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of signal processing, in particular to a method for radio frequency fingerprint identification of unmanned aerial vehicles. TECHNICAL BACKGROUND

[0002] “Low, slow and small” unmanned aerial vehicles have the advantages of miniaturization, low cost, multi-function, flexibility and convenience, low requirement for use environment, strong field survival ability, etc., and are widely used in remote sensing surveying and mapping, weather monitoring, precision agriculture, remote medical treatment, emergency rescue, Internet of Things and next-generation broadband mobile communication, etc. However, while the technology of unmanned aerial vehicles is rapidly evolving and innovative applications are developing explosively, various civilian unmanned aerial vehicles continue to fly “black” and cases of interference with civil aviation and public safety are frequently reported. The existing unmanned aerial vehicle monitoring technology needs to be strengthened in terms of detection capability and identification accuracy.

[0003] At present, the basic idea of detecting and identifying “low, slow and small” unmanned aerial vehicles is to first select some features of the unmanned aerial vehicles to form a feature set, and then use a suitable classification algorithm to classify the feature set. Different features and their selection methods, different classification methods can be combined to form different solutions. In terms of feature selection, the acoustic features, optical images, radar reflections and radio frequency fingerprints of unmanned aerial vehicles are commonly used, or the fusion of these features. Using the acoustic features of unmanned aerial vehicles is easily affected by background noise and distance, and is not suitable for long-distance identification. Using the optical image and radar reflection features of unmanned aerial vehicles are severely limited by hardware performance and each has its own limitations. Using the radio frequency signals of unmanned aerial vehicles, although different unmanned aerial vehicle devices have differences, the radio frequency fingerprint features of different types of devices have certain identity, and the identification effect of simple time domain or frequency domain features is not ideal. Moreover, the transient characteristics of the radio frequency signals radiated by unmanned aerial vehicles are significantly affected by the space propagation path, and the detection and identification effect of the existing methods under low signal-to-noise ratio conditions is very poor, so it is urgent to construct a set of unmanned aerial vehicle radio frequency extraction and identification system to improve the identification accuracy of unmanned aerial vehicle devices. SUMMARY

[0004] The present application aims to overcome the deficiencies of the prior art. The present application selects the minimum variance distortionless response (MVDR) spectrum of the unmanned aerial vehicle measurement and control signal as the feature, maps the signal feature to the image feature through the MVDR spectrum diagram; at the same time, a lightweight network model UAV-G-ConvNet based on group convolution is constructed, and the network is used to train and identify the MVDR spectrum diagram. The method proposed in the present application is divided into the following two steps, that is, first, the time domain unmanned aerial vehicle radio frequency signal is converted into an MVDR spectrum diagram, and then the lightweight network model UAV-G-ConvNet is used to identify the MVDR spectrum diagram.

[0005] The technical scheme of the present application is a kind of unmanned aerial vehicle model identification method based on distortionless spectrum diagram and lightweight network, which comprises:

[0006] Step 1: MVDR spectrum diagram transformation is performed on the radio frequency time domain signal of the unmanned aerial vehicle;

[0007] Step 1.1: the M+1 dimensional autocorrelation matrix R of the radio frequency time domain signal x(n) of the unmanned aerial vehicle is solved M+1 :

[0008]

[0009] Wherein, r(M) represents the first row and the Mth column autocorrelation coefficient;

[0010] Step 1.2: according to the autocorrelation matrix R obtained in step 1.1 M+1 The output average power P of the radio frequency signal is derived;Suppose that the original input signal passes through the M-order finite impulse response FIR filter, and the output y(n) is:

[0011] y(n)=h H x(n)=x T (n)h *

[0012] Wherein, h=[h0,h1,...,h M ] is the weight vector;The output average power P of the radio frequency signal is:

[0013] P=E{h H x(n)x H (n)h}=h H R M+1 h

[0014] Step 1.3: according to the MVDR idea, the average power P in step 1.2 is rewritten;Thus, the MVDR spectrum of the radio frequency time domain signal of the unmanned aerial vehicle is solved;

[0015] Step 1.3.1: find the equivalent condition of the undistorted condition of the original input signal through the FIR filter; the frequency response of the M-order FIR filter at the observation frequency ω i has a unit gain, and the undistorted condition is expressed as:

[0016]

[0017] where ω i represents the i-th frequency component of the M-order FIR filter;

[0018] The equivalent condition is expressed as: α H ω i h i = 1; where

[0019] Step 1.3.2: from the basic idea of MVDR, we get:

[0020]

[0021] The optimal solution is obtained, that is, the following equation is solved:

[0022]

[0023] Step 1.3.3: rewrite the average power P equation in step 1.2 from the optimal solution equation in step 1.3.2, to get the general expression of the MVDR spectrum:

[0024]

[0025] Then the MVDR spectrum of any M-order is:

[0026]

[0027] Step 1.3.4: find the MVDR spectrum of the UAV radio frequency time domain signal according to the expression of the MVDR spectrum of any M-order in step 1.3.3;

[0028] Step 2: construct a lightweight network model UAV-G-ConvNet based on grouped convolution to identify the MVDR spectrum of the UAV radio frequency time domain signal obtained in step 1;

[0029] Step 2.1: construct a feature extraction module (e-unit) to reduce the size of the input MVDR spectrum and extract the shallow information of the picture;

[0030] Step 2.2: construct four grouped convolution modules (g-unit) to sequentially extract the shallow information of the picture extracted in step 2.1 in detail, and further reduce the size of the picture features;

[0031] Step 2.3: Constructing the classification module (C-unit) according to the fine features of the pictures extracted in step 2.2 to give the recognition result of the radio frequency signal of the unmanned aerial vehicle.

[0032] Further, the feature extraction module (e-unit) in step 2.1 includes 3 layers:

[0033] The first layer is a two-dimensional convolution layer (conv1) with a size of 3*3pixel and a step length s of 2;

[0034] The second layer is a BatchNorm (BN) layer, which divides the output of the convolution layer into multiple batches for normalization, and in the BN layer, the output of each batch is

[0035]

[0036] where y n is the output of the conv1 layer, n=1,2,…,N, N is the sample size in each batch; ε is a constant to ensure the stability of the mean square error, usually ε=10 -5 ; the parameters μ b and represent the mean and mean square error of the output respectively, and the batch output is translated and scaled to get

[0037]

[0038] where k represents the scale factor and ξ represents the translation factor, both k and ξ are network training parameters;

[0039] The third layer is a nonlinear activation function f(υ n ), which is:

[0040]

[0041] where α is a constant, and υ n represents the result of translation and scaling of the batch output .

[0042] Further, the classification module (C-unit) is a fully connected layer.

[0043] The present application improves the recognition rate of different unmanned aerial vehicle equipment types, and at the same time, the present application also uses a lightweight grouping network model to reduce the computational resource overhead in the process of identifying the MVDR spectrum. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 ​Time-domain waveforms (left) and their corresponding MVDR spectrograms (right) of the control signals of DJI_M600 and BEEBEE RUN.

[0045] Figure 2 The overall structure diagram of UAV-G-ConvNet.

[0046] Figure 3 The detection (left) and identification (right) of the control signals of five networks under different SNRs. DETAILED DESCRIPTION

[0047] The first problem to be solved in the UAV model recognition method based on the minimum variance distortionless spectrum and the lightweight network is the conversion of the UAV radio frequency time domain signal to the minimum variance distortionless spectrum. The MVDR spectrum is used as the candidate feature of the UAV model recognition, because it is essentially a narrow-band filter in the frequency domain or the spatial domain. The application uses the suppression effect of MVDR on signals outside the observation frequency to filter out interference, and uses the frequency resolution of the MVDR algorithm which is much higher than that of the conventional Fourier transform to finely depict the UAV radio frequency signals including the control signals and the downlink signals; the frequency domain features are mapped to images through the MVDR spectrogram, so as to achieve the purpose of training and prediction using the convolutional neural network. The basic idea of the MVDR spectrum estimation is to design a filter group which has a distortionless frequency response at the observed frequency, and at the same time minimizes the energy of the output signal to ensure maximum suppression of other frequencies.

[0048] Let the input signal of the M-order finite impulse response (FIR) filter be x(n), and the output y(n) be

[0049]

[0050] where h i (i=0, 1,..., M) is the weight coefficient of the FIR filter.

[0051] The input vector of the FIR is defined as x(n)=[x(n), x(n-1),..., x(n-M)] T , and the weight vector is h=[h0, h1,..., h M ]. The output vector y(n) and the output average power P can be respectively represented as:

[0052] y(n)=h H x(n)=x T (n)h *

[0053] P=E{|y n | 2}=E{hH x(n)x H (n)h}=h H R M+1 h

[0054] wherein the matrix is the M+1 dimensional autocorrelation matrix of the vector x(n):

[0055]

[0056] The frequency response of the Mth order FIR filter at the observation frequency ω i has unit gain, i.e.

[0057]

[0058] The above formula is equivalent to α H ω i h i =1, where αω=[1, e jω ,e j2ω ,…,e jMω ] T .

[0059] From the basic idea of MVDR, there is

[0060]

[0061] To obtain the optimal solution, the following formula is solved:

[0062]

[0063] Rewriting the average power P formula, the general expression of the MVDR spectrum is obtained:

[0064]

[0065] This formula is the MVDR spectrum, and the MVDR spectrum of any Mth order can be written as:

[0066]

[0067] The present application sets the order M to 256. Taking DJI_M600 and BEEBEERUN as examples, the time domain data of the unmanned aerial vehicle control signal of the two models is given, and the MVDR spectrum obtained by referring to step 1 is shown as follows. Figure 1

[0068] To solve the problems of excessive parameter quantity and excessive FLOPS of the traditional deep network, the present application uses the grouping convolution idea to construct a lightweight network model "UAV-G-ConvNet" for unmanned aerial vehicle radio frequency signal recognition, and the overall structure is shown as follows. Figure 2 ​​

[0069] By Figure 2 It can be seen that the UAV-G-ConvNet first sets a feature extraction module (e-unit) after the input layer, then four consecutive grouped convolution modules (g-unit), and finally a classification module (C-unit) to realize classification.

[0070] The first layer of the feature extraction module e-unit is a two-dimensional convolution layer (conv1) with a size of 3x3 pixels and a step s of 2. The role of this convolution layer is to reduce the size of the input picture while extracting the shallow information of the picture.

[0071] In order to make the output of the intermediate layer more stable and have fast learning performance, the e-unit sets a BatchNorm (BN) layer after the convolution layer. The BN layer divides the output of the convolution layer into multiple batches for normalization, thereby reducing the sensitivity of the network to initial values and can inhibit overfitting to a certain extent;

[0072] In the BN layer, the output of each batch is For

[0073]

[0074] Where y n is the output of the conv1 layer (n = 1, 2,..., N, N is the sample size in each batch); ε is a constant to ensure the stability of the mean square error, usually take ε = 10 -5 ; Parameters μ b and respectively represent the mean and mean square error of the output, defined as follows:

[0075]

[0076] Where i (i = 1, 2,..., M) is the batch number, and M is the batch size, i.e. the number of batches.

[0077] The batch output is translated and scaled to get

[0078]

[0079] The BN layer is followed by a nonlinear activation function eLU, defined as:

[0080]

[0081] where a is a constant, which can be taken as 1 for simplicity. Compared with linear activation function, eLU can improve classification accuracy and convergence speed. After E-Unit, the input picture size is changed from 3x224x224 of the original input layer to 64x112x112.

[0082] The idea of grouped convolution is to divide the input feature map into multiple groups, perform normal convolution within each group, and then splice the feature maps obtained by each group convolution to obtain the same output feature dimension as normal convolution. The G-Unit constructed in the application is used to extract fine features of the unmanned aerial vehicle radio frequency signal, and four G-Units are designed. The first layer of each G-Unit is a grouped convolution layer (g-conv), and g is the number of groups. By selecting a suitable g value, the calculation complexity and parameter amount of convolution can be reduced; the larger the g value, the more obvious the reduction of the parameter amount; the layered grouped convolution can further reduce the parameters, thereby achieving the purpose of network lightweight. In the application, the value of g is 8.

[0083] The data information of grouped convolution is only within the group, and there is no interaction between channels, which may cause a phenomenon similar to inbreeding and reduce the accuracy of the classifier. Therefore, a channel rearrangement (shuffle) layer is added after the grouped convolution layer, and the shuffle operation is used to interact the information of channels in different groups. The BN layer and the eLU layer are arranged after the shuffle layer, which are the same as the E-unit.

[0084] In the G-unit, the deepening of the network structure makes the size of the convolution filter smaller and smaller, and the arrangement and combination possibility of the features is more and more, so as to cover as many key features as possible. At this time, the channel dimension of the output feature layer should be increased to make the network have stronger expression. The output channel numbers of the four G-Units are set to 96, 128, 192 and 256 respectively in the application, the convolution kernel size is 3x3pixel, and the step s is 2.

[0085] After the four G-Units, the feature size becomes 256x7x7. The feature is reduced to 256x4x4 through a max pooling layer; then the multi-dimensional feature output by the pooling layer is compressed into a one-dimensional feature through a Flatten layer, to prepare for input to the fully connected layer (FC).

[0086] After the fully connected layer, a softmax function is used as the last output layer of the UAV-G-ConvNet network, which is defined as follows

[0087]

[0088] where, F fc is the output of the fully connected layer.

[0089] The cross-entropy loss function is selected in the present application. It is defined as

[0090]

[0091] where T l (l = 1, 2, …, L, L is the total number of classifications) is the expected class probability, v l is the class probability output by the network.

[0092] Training and verification

[0093] The data set used in the present application includes: 6 types of remote controller signals of unmanned aerial vehicles as the measurement and control signal data set, and the unmanned aerial vehicle models are DJI Phantom4, DJI InSpire, DJI Matrice600, DJI MavicPro1, BeeBeerunFPV RC drone mini quadcopter, and 3DR Iris FS-TH9x. The radio frequency signals of 2 Bluetooth devices (including 1 smart phone Apple Iphone 6S and 1 smart watch FitBit Charge3 smartwatch) and 2 WiFi routers (models are Cisco Linksys E3200 and TP-link TL-WR940N, respectively) are selected as interference signals.

[0094] The frequencies of all the above radio frequency signals are 2.4 GHz, the sensor sampling frequency is 5 Gsps, and the sampling time is 0.001 s; the number of all radio frequency signal data is 300, and all are divided into training set and test set in the ratio of 2:1.

[0095] The above radio frequency signal time domain data set is used to generate MVDR spectrograms of the same size according to step 1, and then the network model is used to train and test the pictures. All steps are developed and implemented based on Windows 11 environment, Python 3.10 platform. The CPU is Intel Xeon GOLD 5222, the GPU is NVIDIA GeForce RTX 2080TI, the neural network training framework is Pytorch 1.21.1, the network optimizer is Adam, the learning rate is fixed at 0.0001, the batch-size is 16, and the number of training rounds is 100.

[0096] The input and output sizes of each layer of the UAV-G-ConvNet are shown in Table 1.

[0097] Four benchmark network models are selected for comparison and verification: ResNet50, ShuffleNetV2, MobileNetV2, and SqueezeNet. The recognition of the six UAV control signals and four interference signals by the UAV-G-ConvNet proposed in the application is compared with the benchmark network models, as shown in Table 2.

[0098] As shown in Table 2, the parameter amount of the UAV-G-ConvNet is only 20.47% of the SqueezeNet, which has the least parameter amount; and the FLOPS of the UAV-G-ConvNet is only 47.31% of the ShuffleNetV2, which has the least FLOPS.

[0099] Without additional channel noise injection, the MDVR spectrum-based MVDR+UAV-G-ConvNet scheme is compared with four benchmark network models that also use the MVDR spectrum as a feature, and a scheme in which the existing method uses a wavelet scattering transform graph as a feature and uses two network models. The results are shown in Table 3. As shown in Table 3, the UAV-G-ConvNet of the application improves the recognition rate by 2.8% compared with the best scheme in the existing method. The accuracy of the solution of the application is only 0.5% lower than that of the MVDR+ResNet50 scheme, which performs best in model identification. According to Table 2, the parameter amount of the UAV-G-ConvNet is more than one order of magnitude less than that of ResNet50; and the FLOPS is more than two orders of magnitude less than that of ResNet50.

[0100] After adding Gaussian white noise with different signal-to-noise ratios, the performance of the UAV-G-ConvNet in the application and the four benchmark network models for the two classification problems is shown in Table 4. Figure 3 As can be seen, the existence detection accuracy of the UAV-G-ConvNet is 99.4% and the model identification accuracy is 87.5% at a signal-to-noise ratio of 0 dB, under the premise of small parameter amount and computational amount.

[0101] Table 1: Input and output size of each layer of UAV-G-ConvNet

[0102]

[0103] Table 2: Comparison of parameter amount and FLOPS of each network model (total number of classifications L = 10)

[0104]

[0105] Table 3: Comparison of detection and identification accuracy of each network model for control signals

[0106]

Claims

1. A UAV model recognition method based on distortionless spectrogram and lightweight network, the method comprising: Step 1: MVDR spectrogram transformation is performed on the UAV radio frequency time domain signal; Step 1.1: Obtain the M+1 dimensional autocorrelation matrix R of the drone radio frequency time domain signal x(n) M+1 : Wherein, r(M) represents the first row and the Mth column autocorrelation coefficient; Step 1.2: Autocorrelation matrix R obtained from step 1.1 M+1 The output average power P of the radio frequency signal is derived; assuming that the original input signal is filtered by an M-order finite impulse response (FIR) filter, the output y(n) is: y(n) = h H x(n) = x T (n)h * Where, h = [h0, h1, ..., h M Let ] be the weight vector; then the average output power P of the RF signal is: P = E{h H x(n)x H (n)h} = h H R M+1 h Step 1.3: According to the MVDR idea, the average power P in step 1.2 is rewritten; thus the MVDR spectrum of the UAV radio frequency time domain signal is obtained; Step 1.3.1: Find the equivalent condition of the undistorted condition for the original input signal passing through the FIR filter; the frequency response of the Mth order FIR filter at the observation frequency ω i has unity gain, and the undistorted condition is expressed as: where ω i represents the i-th frequency component of the M-th order FIR filter; The equivalent condition is expressed as: α H ω i h i = 1; wherein Step 1.3.2: According to the MVDR basic idea, we get: The optimal solution is obtained, that is, the following equation is solved: Step 1.3.3: According to the optimal solution equation in step 1.3.2, the average power P equation in step 1.2 is rewritten, and the general expression of the MVDR spectrum is obtained: Then the MVDR spectrum of any M order is: Step 1.3.4: According to the MVDR spectrum expression of any M order obtained in step 1.3.3, the MVDR spectrogram of the UAV radio frequency time domain signal is obtained; Step 2: Construct a lightweight network model UAV-G-ConvNet based on grouped convolution to identify the MVDR spectrogram of the UAV radio frequency time domain signal obtained in step 1; Step 2.1: Construct a feature extraction module to reduce the size of the input MVDR spectrogram while extracting the shallow information of the picture; Step 2.2: Construct four grouped convolution modules to extract the shallow information of the picture extracted in step 2.1 in turn, and further reduce the size of the picture features; Step 2.3: Construct a classification module to give the UAV radio frequency signal recognition result according to the fine features of the picture extracted in step 2.

2.

2. The unmanned aerial vehicle model recognition method based on a distortionless spectrum and a lightweight network according to claim 1, wherein, The feature extraction module in step 2.1 includes 3 layers: The first layer is a two-dimensional convolution layer with a size of 3x3pixel and a step s of 2; The second layer is a BatchNorm layer. The BN layer divides the output of the convolutional layer into multiple batches for normalization. In the BN layer, the output of each batch is is: where y n is the output of the conv1 layer, n = 1, 2, …, N, N is the sample size in each batch; ε is a constant to ensure the stability of the mean square error, usually taken ε = 10 -5 ; Parameters μ b and respectively represent the mean and mean square error of the output, and the batch output is translated and scaled to obtain Wherein, k represents the scale factor, and ξ represents the translation factor, both k and ξ are network training parameters; The third layer is a non-linear activation function f(u n ), which is: where a is a constant, u n represents a batch output The result of the translation and scaling. 3.The UAV model recognition method based on distortionless spectrogram and lightweight network according to claim 1, wherein, The classification module is a fully connected layer.

Images