Unmanned aerial vehicle detection and identification method based on multi-dimensional feature fusion

Abstract

The invention discloses an unmanned aerial vehicle detection and identification method based on multi-dimensional feature fusion, and belongs to the technical field of unmanned aerial vehicle detection, the unmanned aerial vehicle detection and identification method comprises a signal receiving module and a signal processing module, is suitable for the field of unmanned aerial vehicle detection and classification identification, and is mainly characterized in that a software wireless electronic board is connected with a 2.4-2.485 GHz omnidirectional antenna to collect wireless signals; the received signals are preprocessed, the unmanned aerial vehicle is detected by comparing whether the variance of the received signals is larger than a denoising threshold value solved by a self-adaptive triangular threshold value method or not, and an unmanned aerial vehicle communication signal segment is confirmed; and extracting Mel-frequency cepstral coefficient and axial integral bispectrum features of the signal segments, taking fusion features obtained through canonical correlation analysis and processing as fingerprint features, and finally performing classification and identification on the unmanned aerial vehicle by using a classification and identification algorithm.

Description

Technical Field

[0001] The present invention belongs to the field of drone detection and identification, and specifically proposes a drone detection and identification method. Background Art

[0002] With the development of economy and the progress of science and technology, drones have been widely used in various fields. With the increase in the number of drones, the incidents of "illegal flying" of drones have occurred frequently. Therefore, it is necessary to develop a system that can quickly and accurately detect and identify "illegal flying" drones to improve the efficiency of management and control.

[0003] At present, the main drone detection technologies include: vision-based detection, voiceprint-based detection, radar-based detection, etc. Vision-based detection uses high-definition cameras to obtain image information and combines it with drone detection algorithms to detect drones. Voiceprint-based detection uses the audio information generated by the rotor and motor of the drone during flight to extract corresponding features to detect and identify drones. However, vision-based detection and voiceprint-based detection are greatly affected by environmental factors. Bad weather and noisy environments will greatly reduce the recognition accuracy. Radar-based detection uses radar to obtain drone echo information and detects and identifies drones by extracting echo information features. However, this technology has high requirements for equipment, and due to the small scattering area of ​​drones, it is not easy to detect and identify.

[0004] The drone identification technology based on multidimensional feature fusion proposed in the present invention has low equipment requirements, simple calculation process and high recognition accuracy. The present invention extracts the Mel-frequency cepstral coefficients and axially integrated bispectral features of the drone radio frequency signal, and obtains new features by performing feature fusion processing through canonical correlation analysis. The subtle differences between different drone communication signals are described by multidimensional fusion features, thereby realizing the classification of drones and improving the management and control efficiency of drones. Summary of the invention

[0005] The purpose of the present invention is to provide a method for detecting and identifying unmanned aerial vehicles based on multi-dimensional feature fusion.

[0006] The present invention provides the following technical solutions:

[0007] A method for detecting and identifying unmanned aerial vehicles based on multi-dimensional feature fusion, the technical solution includes:

[0008] Firstly, the 2.4-2.485 GHz omnidirectional antenna is connected to the software radio electronic board to collect and save the wireless signal. The adaptive triangular threshold method is used to calculate the denoising threshold of the captured signal. Then the data is segmented, and the variance of the segmented data is calculated and compared with the denoising threshold. The data greater than the threshold is the UAV communication signal data. The Mel cepstral coefficients and axial integral bispectral features of the UAV signal data are calculated. The fused features obtained by feature fusion processing through canonical correlation analysis are used as the signal fingerprint features of the UAV. Finally, the UAV is classified and identified.

[0009] A method for identifying drones based on multi-dimensional feature fusion includes the following steps:

[0010] Step 1: Use the software-defined radio electronic board with a 2.4-2.485 GHz omnidirectional antenna to collect and store wireless signals.

[0011] Step 2: Analyze and process the signal obtained in step 1, and use the adaptive triangular threshold method to calculate the denoising threshold.

[0012] Step 3: Based on step 2, the variance of the data segment is calculated and compared with the denoising threshold. The data segment greater than the threshold is the drone communication signal data, and the drone detection is completed at the same time.

[0013] Step 4: Extract features based on the effective drone signal obtained in step 3 and calculate the Mel-frequency cepstral coefficients of the data.

[0014] Step 5: Perform feature extraction based on the effective UAV signal obtained in step 3 and calculate the axial integrated bispectrum of the data.

[0015] Step 6: Based on the UAV signal features obtained in steps 4 and 5, the fusion features are obtained through canonical correlation analysis.

[0016] Step 7: Based on the fusion features obtained in step 6, the SVM algorithm is used to classify and identify the drone.

[0017] Beneficial Effects

[0018] The beneficial effects of the present invention are as follows: the present invention first uses a software wireless electronic board to connect a 2.4-2.485GHz omnidirectional antenna to collect wireless signals, pre-processes the received signals, and detects the drone and confirms the drone communication signal segment by comparing whether the variance of the received signal is greater than the denoising threshold; extracts the Mel-frequency cepstrum coefficients and axially integrated bispectral of the signal segment, and uses typical correlation analysis to fuse the features as fingerprint features to achieve classification and identification of drones. Drone communication signals have strong anti-interference capabilities. This method extracts fingerprint features from this signal to achieve drone identification, which can effectively solve the problems of complex systems, time-consuming feature extraction, and low recognition accuracy in current drone detection and identification equipment. And compared with recognition methods based on voiceprints, optoelectronics, radar, etc., this method is less affected by the environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a schematic diagram of the flow chart of the method for detecting and identifying unmanned aerial vehicles in the present invention;

[0020] Figure 2 It is a schematic diagram of the process of the adaptive triangular threshold method in the present invention;

[0021] Figure 3 It is a schematic diagram of the Mel-cepstral coefficient and axially integrated bispectrum extraction process in the present invention;

[0022] Figure 4 It is a schematic diagram of the dual spectrum axial integration path in the present invention; DETAILED DESCRIPTION

[0023] The present invention will be further described below in conjunction with the accompanying drawings:

[0024] The present invention provides a method for detecting and identifying unmanned aerial vehicles based on Mel-frequency cepstral coefficients, which can be applied to unmanned aerial vehicle detection and identification systems, such as Figure 1 As shown, a method for detecting and identifying drones based on Mel-frequency cepstral coefficients comprises the following steps:

[0025] Step 1: Use the software radio electronic board with a 2.4-2485 GHz omnidirectional antenna to collect and store wireless signals.

[0026] Step 2: Analyze and process the signal obtained in step 1, and use the adaptive triangular threshold method to calculate the denoising threshold. The threshold calculation process is as follows: Figure 2 As shown:

[0027] 2.1 Input data D = {d 1 ,d 2 ,...,d n}, where d nis the amplitude of the nth data point, take the absolute value of each data, count the amplitude of each element, and generate the corresponding histogram. The horizontal axis of the histogram represents the amplitude of the input signal. The range of the amplitude is [d' min ,d' max ], where d' min is the minimum amplitude, d' max is the maximum amplitude; the ordinate represents the number of times this amplitude occurs, and H(d') represents the number of times data with an amplitude of d' occurs, where d'∈[d' min ,d' max ].

[0028] 2.2 Find the point with the highest frequency in the histogram and record it as point A, (d' A ,H(d' A ))

[0029] 2.3 Find the first point on the opposite side of point A with a non-zero frequency and record it as point B, (d' B ,H(d' B )).For example When From the amplitude d' max Past Direction to find the first point whose frequency is not zero.

[0030] 2.4 Draw a straight line L through points A and B, traverse all points between A and B, calculate their distance to the straight line L, and the amplitude corresponding to the point with the largest distance is the denoising threshold.

[0031] Step 3: Based on step 2, calculate the variance of the received drone data. If the value is greater than the required denoising threshold, then this signal is a drone communication signal, and the drone detection is completed.

[0032] Step 4: Select the drone communication signal segment data based on the threshold value obtained in step 3, calculate the Mel cepstral coefficient of this data and use it for subsequent recognition, such as Figure 3 As shown, the steps for extracting Mel cepstral coefficients are as follows:

[0033] 4.1 Let the signal segment data be s(n), and perform windowing on this data to obtain data s', where N is the length of the input data and μ is the Hamming window coefficient:

[0034]

[0035] 4.2 Perform discrete Fourier transform on the windowed data to transform the time domain data into frequency domain data:

[0036]

[0037] 4.3.1 Determine the minimum frequency f L , maximum frequency f H , the number of Mel filters M.

[0038] 4.3.2 The center frequency of each filter is calculated as follows, where 1≤m≤M, N is the length of the discrete Fourier transform, and f s is the sampling frequency:

[0039]

[0040] In the formula

[0041] 4.3.3 The function expression of the Mel filter bank is as follows:

[0042]

[0043] 4.4 The spectrum data obtained after discrete Fourier transform is filtered through the Mel filter bank, and the logarithm of each Mel filter output is calculated, that is, the logarithmic power spectrum of the corresponding frequency band. The calculation formula is as follows:

[0044]

[0045] Where, m = 1, 2, ..., M, M is the number of filters; H m (k) is the mth filter transfer function.

[0046] 4.5 Perform discrete cosine transform on the desired logarithmic power spectrum, and the result is the Mel frequency cepstrum coefficient, which is calculated as follows:

[0047]

[0048] Step 5: Select the drone communication signal segment data based on the threshold value obtained in step 3, calculate the axial integral bispectrum of this data and use it for subsequent identification, such as Figure 3 As shown, the steps for extracting the axially integrated bispectrum are as follows:

[0049] 5.1 Divide the signal segment data s(1), s(2), ..., s(N) into P segments, each of which has M data points, that is, N = M × P, where N is the total length of this segment of signal data. The Pth segment of data is represented by s p (0),s p (1),...,s p (M-1), where p=1,2,...,P.

[0050] 5.2 Calculate the discrete Fourier transform coefficients for each segment of data:

[0051]

[0052] Where k = 0, 1, ..., M / 2, p = 1, 2, ..., P.

[0053] 5.3 Calculate the third-order correlation of discrete Fourier transform coefficients:

[0054]

[0055] where p=1,2,...,P,0≤k 2 ≤k 1 , k 1 +k 2 ≤f s / 2,Δ 0 =f s / N,f s is the sampling frequency, Δ 0 represents the frequency sample space, and N, M and L 1 The relationship between them satisfies M=(2L 1 +1)N.

[0056] 5.4 Take the average of the bispectral estimates of the P segment data as the bispectral estimates of the signals s(1), s(2), ..., s(N):

[0057]

[0058] Among them, ω 1 =2πf s k 1 / N,ω 2 =2πf s k 2 / N.

[0059] 5.5 Calculate the axial integral of the bispectrum. The integral path is as follows: Figure 4 As shown:

[0060]

[0061] Among them, B(ω 1 ,ω 2 ) is the bispectral estimation value of the UAV signal.

[0062] Step 6: Based on the drone features obtained in steps 4 and 5, feature fusion processing is performed through canonical correlation analysis, and the obtained fused features are used as classification training data.

[0063] 6.1 Let the Mel cepstral coefficient and the axially integrated bispectral eigenvector be Each eigenvector has n groups, so the feature matrix is ​​recorded as Let the projections of the two eigenvectors be w x and w y, then the goal of canonical correlation analysis is to maximize the following formula:

[0064]

[0065] where Σ xx =XX T ,Σ yy =YY T ,Σ xy =XY T , are the covariance matrix and cross-covariance matrix of X and Y respectively.

[0066] 6.2 Convert the formula in 6.1 into an optimization problem with the following constraints:

[0067]

[0068] Using the Lagrangian function, we get the following formula:

[0069]

[0070] For w x and w y Derivative:

[0071]

[0072] Let λ = λ x =λ y Then simplify it to get:

[0073]

[0074] Substituting into the partial derivative equation, we can get

[0075]

[0076] At this point, the problem is transformed into solving the characteristic equation of the transformation matrix. The transformation matrix to be solved is recorded as The same can be said Where l≤min{p,q}.

[0077] 6.3 Note is the transformed feature, and the fusion feature is

[0078] Step 7: Based on the drone fusion features obtained in step 6 as training data, use the SVM algorithm for machine learning to establish a radio frequency fingerprint library. Then, this method is used to extract drone fingerprint features and compare them with the radio frequency fingerprint library to achieve classification and identification of drones.

Claims

1. A drone identification method based on multi-feature fusion, characterized in that: include: Step 1: Use the software radio electronic board with a 2.4-2485 GHz omnidirectional antenna to collect and store wireless signals. Step 2: Analyze and process the signal obtained in step 1, and use the adaptive triangular threshold method to calculate the denoising threshold. The threshold calculation process is as follows: 2.1 Input data D = {d1, d2, ..., d n }, where d n is the amplitude of the nth data point, take the absolute value of each data, count the amplitude of each element, and generate the corresponding histogram. The horizontal axis of the histogram represents the amplitude of the input signal. The range of the amplitude is [d' min ,d' max ], where d' min is the minimum amplitude, d' max is the maximum amplitude; the ordinate represents the number of times this amplitude occurs, and H(d') represents the number of times data with an amplitude of d' occurs, where d'∈[d' min ,d' max ]. 2.2 Find the point with the highest frequency in the histogram and record it as point A, (d' A ,H(d' A )) 2.3 Find the first point on the opposite side of point A with a non-zero frequency and record it as point B, (d' B ,H(d' B )).For example When From the amplitude d' max Past Direction to find the first point whose frequency is not zero. 2.4 Draw a straight line L through points A and B, traverse all points between A and B, calculate their distance to the straight line L, and the amplitude corresponding to the point with the largest distance is the denoising threshold. Step 3: Based on step 2, calculate the variance of the received drone data. If the value is greater than the required denoising threshold, then this signal is a drone communication signal, and the drone detection is completed. Step 4: Select the drone communication signal segment data based on the threshold value obtained in step 3, calculate the Mel-frequency cepstral coefficient of this data and use it for subsequent recognition. The steps for extracting the Mel-frequency cepstral coefficient are as follows: 4.1 Let the signal segment data be s(n), and perform windowing on this data to obtain data s', where N is the length of the input data and μ is the Hamming window coefficient: 4.2 Perform discrete Fourier transform on the windowed data to transform the time domain data into frequency domain data: 4.3.1 Determine the minimum frequency f L , maximum frequency f H , the number of Mel filters M. 4.3.2 The center frequency of each filter is calculated as follows, where 1≤m≤M, N is the length of the discrete Fourier transform, and f s is the sampling frequency: In the formula 4.3.3 The function expression of the Mel filter bank is as follows: 4.4 The spectrum data obtained after discrete Fourier transform is filtered through the Mel filter bank, and the logarithm of each Mel filter output is calculated, that is, the logarithmic power spectrum of the corresponding frequency band. The calculation formula is as follows: Where, m = 1, 2, ..., M, M is the number of filters; H m (k) is the mth filter transfer function. 4.5 Perform discrete cosine transform on the desired logarithmic power spectrum, and the result is the Mel frequency cepstrum coefficient, which is calculated as follows: Step 5: Select the drone communication signal segment data based on the threshold value obtained in step 3, calculate the axial integral bispectrum of this data and use it for subsequent identification. The steps for extracting the axial integral bispectrum are as follows: 5.1 Divide the signal segment data s(1), s(2), ..., s(N) into P segments, each of which has M data points, that is, N = M × P, where N is the total length of this segment of signal data. The Pth segment of data is represented by s p (0),s p (1),...,s p (M-1), where p=1,2,...,P. 5.2 Calculate the discrete Fourier transform coefficients for each segment of data: Where k = 0, 1, ..., M / 2, p = 1, 2, ..., P. 5.3 Calculate the third-order correlation of discrete Fourier transform coefficients: where p=1,2,...,P,0≤k2≤k1,k1+k2≤f s / 2, Δ0=f s / N,f s is the sampling frequency, Δ0 represents the frequency sample space, and N, M and L1 satisfy the relationship M=(2L1+1)N. 5.4 Take the average of the bispectral estimates of the P segment data as the bispectral estimates of the signals s(1), s(2), ..., s(N): in, ω1=2πf s k1 / N,ω2=2πf s k2 / N. 5.5 Calculate the axial integral of the bispectrum: Among them, B(ω1,ω2) is the bispectral estimation value of the UAV signal. Step 6: Based on the drone features obtained in steps 4 and 5, feature fusion processing is performed through canonical correlation analysis, and the obtained fused features are used as classification training data. 6.1 Let the Mel cepstral coefficient and the axially integrated bispectral eigenvector be Each eigenvector has n groups, so the feature matrix is ​​recorded as Let the projections of the two eigenvectors be w x and w y , then the goal of canonical correlation analysis is to maximize the following formula: where Σ xx =XX T ,Σ yy =YY T ,Σ xy =XY T , are the covariance matrix and cross-covariance matrix of X and Y respectively. 6.2 Convert the formula in 6.1 into an optimization problem with the following constraints: Using the Lagrangian function, we get the following formula: For w x and w y Derivative: Let λ = λ x =λ y Then simplify it to get: Substituting into the partial derivative equation, we can get At this point, the problem is transformed into solving the characteristic equation of the transformation matrix. The transformation matrix to be solved is recorded as The same can be said Where l≤min{p,q}. 6.3 Note is the transformed feature, and the fusion feature is Step 7: Based on the drone fusion features obtained in step 6 as training data, use the SVM algorithm for machine learning to establish a radio frequency fingerprint library. Then, this method is used to extract drone fingerprint features and compare them with the radio frequency fingerprint library to achieve classification and identification of drones.

Images