Unmanned aerial vehicle detection method based on super-directivity beam forming

Abstract

The invention provides an unmanned aerial vehicle detection method based on super-directivity beam forming, which comprises the following steps: firstly, scanning the amplitudes of plane waves in different directions, finding out a target sound source direction and collecting a sound signal, and then adjusting the phase and gain of the sound signal based on a super-directivity beam forming algorithm to realize signal enhancement. And finally, sound signal identification is carried out according to the convolutional neural network. Compared with the prior art, the method is based on the super-directivity beam forming algorithm, sound signal enhancement of the target sound source is achieved, and the recognition accuracy of the unmanned aerial vehicle is improved; meanwhile, a hybrid beam former is provided, a frequency-independent beam former and a traditional delay summation beam former are combined, the respective defects of the two beam formers are overcome, and the beam formers have high signal-to-noise ratios no matter low-frequency signals or high-frequency signals.

Description

Technical Field

[0001] The present invention relates to the field of unmanned aerial vehicle detection, and in particular to a method for detecting unmanned aerial vehicles based on super-directional beamforming. Background Art

[0002] With the rapid development of drone technology, drones are increasingly used in civil and military fields. However, the disorderly flight of drones and potential safety hazards have gradually become prominent, posing a considerable challenge to public safety and aviation safety. Therefore, it is particularly important to research and develop efficient and accurate drone detection technology.

[0003] There are some drone detection technologies, but they all have their own limitations. Radio spectrum detection cannot detect radio-silent drones, and detection performance will be greatly affected in areas with severe radio signal interference. Photoelectric signal detection cannot detect obscured targets. In loaded urban environments, due to the high background interference, the error rate of photoelectric signal detection is high. Radar detection has low detection reliability in urban environments because obstacles such as buildings and trees will affect the detection range and accuracy of the radar, and the cost of radar detection equipment is very high. The drone detection method based on microphone arrays has received widespread attention in recent years due to its low cost, easy deployment, and strong anti-interference ability.

[0004] The existing drone detection method based on microphone array mainly includes two parts. First, the target sound source is located based on the microphone array, the sound signal of the target sound source is collected, and then the target sound source is recognized for speech to determine whether the target object is a drone. The clarity of the sound signal of the target sound source affects the accuracy of the target sound source identification. When the sound source is far away, the signal reaching the array will be very weak, or when the environmental noise interference is relatively large, it will cover the sound signal of the target sound source. These factors will reduce the accuracy of sound source identification. There are also some methods that integrate microphone arrays with image recognition, RF antenna arrays and other technologies to achieve drone detection, but as the complexity of the technology increases, the difficulty of implementation will also increase. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention proposes a UAV detection method based on super-directional beamforming, the steps of which include: S1: designing a microphone array and establishing a spatial coordinate system, wherein the microphone array includes a plurality of microphones; S2: Since the distance from the target sound source to the microphone array is much greater than the distance between each microphone, the sound wave emitted by the target sound source is equivalent to a plane wave. The sound field at the location of any microphone in the microphone array can be equivalent to the superposition of plane waves in all directions. The direction of the target object is determined according to the amplitude of the plane wave. S3: According to the angle and sound signal of the target sound source, the phase and gain of the sound signal are adjusted through the super-directional beamforming algorithm to enhance the sound signal of the target object and suppress the noise in other directions; S4: Extract features of the enhanced sound signal, and determine whether the target object is a drone based on a convolutional neural network.

[0006] Preferably, the microphone array is composed of a plurality of microphones arranged linearly for collecting sound signals, and the spatial coordinate system takes the left end of the microphone array as the coordinate center, and extends toward the right end as the positive direction of the z-axis.

[0007] Preferably, in step S2, the plane wave amplitude in a certain direction is calculated at any microphone, and then traverses from 0 degrees to 180 degrees, each angle corresponds to a plane wave amplitude, and the angle corresponding to the maximum value is the incident angle of the sound signal of the target object.

[0008] Preferably, in the step S3, the super-directional beamforming algorithm mainly calculates high-order derivatives in the wave number domain to obtain weight values, brings the weight values ​​into the beamformer, and outputs the processed sound signal. The specific steps include: S31: approximating the sound field to a finite superposition of plane waves, associating the sound signals received by the multiple microphones through a wave equation, reconstructing the sound field, and differentiating the reconstructed sound field to obtain a high-order derivative; S32: Calculate the high-order derivative by multiplication to obtain a weight value of the sound signal, bring it into the beamformer, and output an enhanced signal.

[0009] Preferably, in step S4, MFCC is used to extract first-order dynamic response features of the processed sound signal, and then a convolutional neural network modified based on AlexNet is used to identify the processed sound signal, and a result of whether the target object is a drone is output.

[0010] Preferably, the beamformer is obtained by mixing a frequency-independent beamformer and a traditional delay-sum beamformer, which still has a high signal-to-noise ratio for high-frequency signals.

[0011] The present invention proposes a method for detecting drones based on super-directional beamforming, which first scans the plane wave amplitude in different directions to find the direction of the target sound source and collect sound signals, then adjusts the phase and gain of the sound signal based on the super-directional beamforming algorithm to enhance the signal, and finally recognizes the sound signal based on the convolutional neural network. Compared with the prior art, the present invention has the following beneficial effects: improving the accuracy of drone recognition, enhancing the sound signal of the target sound source based on the super-directional beamforming algorithm, and better identifying the drone sound signal; combining the frequency-independent beamformer with the traditional delay-sum beamformer to overcome the shortcomings of the two beamformers, and the beamformer has a higher signal-to-noise ratio for both low-frequency and high-frequency signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 The following is the overall flow chart; Figure 2 It is the flow chart of the hyper-directional algorithm; DETAILED DESCRIPTION

[0013] According to the following Figure 1 and Figure 2 , giving the preferred embodiments of the present invention and describing them in detail, so as to better understand the functions and features of the present invention.

[0014] according to Figure 1 , a method for detecting a drone based on super-directional beamforming according to an embodiment of the present invention comprises the following steps: S1: Design a microphone array and establish a spatial coordinate system. The microphone array consists of M microphones arranged linearly. The distance between adjacent microphones is d. It is used to collect sound signals. The spatial coordinate system takes the left end of the microphone array as the coordinate center, and extends to the right end as the positive direction of the z axis. The angle between the target sound source direction and the z axis is The sound signal received by a microphone can be expressed as: (1) in is the wave number, is the frequency, is the speed of sound,

[0015] S2: Since the distance from the target sound source to the microphone array is much greater than the distance between each microphone, the sound wave emitted by the target sound source is equivalent to a plane wave. The sound field at the location of any microphone in the microphone array can be equivalent to the superposition of plane waves in all directions. In the Cartesian coordinate system, any point The sound field is expressed as the superposition of multiple plane waves. The formula is as follows: (2)

[0016] In the z-axis direction, formula (2) can be further expressed as: (3) Among them, the angle The corresponding plane wave amplitude is , traverses from 0 to 180 degrees, when hour, There is a maximum value, that is, the plane wave amplitude in the direction of the target sound source is the largest. The direction of the target sound source.

[0017] S3: According to the angle and sound signal of the target sound source, the phase and gain of the sound signal are adjusted through the super-directional beamforming algorithm to enhance the sound signal of the target object and suppress the noise in other directions. The mathematical model of the beamformer is constructed as follows: In practice, in addition to receiving the signal of the target object, the microphone also receives a noise signal. By introducing the noise signal vector (4) in, T stands for transpose, represents the position of the microphone, By introducing the weight vector To improve the signal-to-noise ratio, the output of the beamforming algorithm is: (5) in ,The relevant beamformer can be defined as: (6) in stands for complex conjugate transpose; The beamforming algorithm is as follows: Figure 2 As shown, the steps are as follows: S31: The sound field is approximated as a finite superposition of plane waves, the sound signals received by multiple microphones are associated together through wave equations, the sound field is reconstructed, and the reconstructed sound field is differentiated to obtain high-order derivatives; according to formula (3), the sound field of any microphone position superposed by finite plane waves can be obtained: (7) in L is the number of angular discrete points, For the M microphones on the z-axis, the sound field can be expressed by a vector: (8) in and A is a full rank matrix, so formula (8) can be equivalent to: (9) The matrix A can be decomposed into: (10) Among them and is the identity matrix, It can be expressed as: (11) Among them are singular values, is a matrix with all elements equal to zero. By using formulas (10) and (11), we can obtain: (12) According to formulas (10) and (12), we can obtain: (13) In order to increase the robustness of x, in the matrix Introducing filter factor get: (14) Among them is the cutoff parameter, which can be obtained through experiments. Substituting formula (14) into (13), we can get: (15) Among them By differentiating formula (3), we can obtain the nth-order derivative of the sound field at any microphone: (16) S33: Calculate the high-order derivative by multiplication to obtain a weight value of the sound signal, substitute it into the beamformer, and output an enhanced signal.

[0018] By calculating the derivative in the spatial domain through multiplication in the wavenumber domain, similar to the discretization of formula (3), formula (16) can be expressed as: (17) in and ,for , substituting formula (15) into (17), we can obtain: (18) In formula (18), the left side is the real high-order derivative of the sound field, and the right side is the high-order derivative of the reconstructed sound field. For the angle of the target sound source , whose derivative is: (19) The present invention proposes a hybrid beamformer, which is obtained by mixing a frequency-dependent beamformer and a delay-sum beamformer. The frequency-dependent beamformer is: (20) The delay-sum beamformer is: (21) Formulas (20) and (21) can be unified into a beamformer: (twenty two) in Steering vector, represents the Hadamard product, Column vector, when the vector Only the When elements are 1 and the rest are 0, Represents the integer part, when the vector When all elements in are 1,

[0019] According to Euler's formula and binomial theorem, It can be expressed as: (twenty three) Substituting formula (23) into (22) we can obtain: (24) According to formula (24), the beamformer can finally be expressed as: (25) Among them ,for ,

[0020] Substituting formula (19) into (25), we can obtain: (26) Substituting formula (26) into (5), we obtain Directional enhancement signal.

[0021] S4: Extract features of the enhanced sound signal, and determine whether the target object is a drone based on a convolutional neural network.

[0022] After obtaining the enhanced signal of the target sound source, first perform MFCC feature extraction on it, perform windowing and framing on the signal, and then use Fast Fourier Transform (FFT) to convert the time domain sequence of the signal to frequency domain representation, and then use Mel filtering to map the signal to Mel domain representation, so that it is more in line with the auditory characteristics of the human ear. The Mel filter bank is a set of triangular filter banks, in which the center frequency c(l) of each filter is nonlinearly distributed, and the relationship between the Mel frequency m and the actual frequency f is: (27) Discrete cosine transform (DCT) is used to obtain the cepstrum coefficients of the signal. DCT can separate the envelope and details of the spectrum and eliminate the correlation between the output parameters of each filter. The calculation formula is: (28) in represents the MFCC coefficients, The Mel spectrum of the amplitude spectrum is obtained by multiplying the amplitude by a triangular Mel-weighted filter, where m is the coefficient of the mth filter, M is the number of filters, C is the length of the DCT, and C = 13 is usually selected. The dynamic characteristics of the signal are then extracted by first-order difference.

[0023] The extracted MFCC features are used as the input of the convolutional neural network. The convolutional neural network is modified based on AlexNet. The network mainly consists of 5 convolutional layers and 3 fully connected layers. The convolutional layer is activated by the relu activation function, which reduces the complexity of network calculation. Pooling layers are added after the first, second and fifth convolution layers for downsampling to reduce the dimension of the feature matrix. A dropout layer is added after the first two fully connected layers to improve the generalization ability of the network by inactivating some neurons. The last fully connected layer is activated by the softmax function, so that the output is a binary probability distribution, and finally determines whether the target object is a drone.

Claims

1. A method for detecting a drone based on super-directional beamforming, the steps comprising: S1: designing a microphone array and establishing a spatial coordinate system, wherein the microphone array includes a plurality of microphones; S2: Since the distance from the target sound source to the microphone array is much greater than the distance between each microphone, the sound wave emitted by the target sound source is equivalent to a plane wave. The sound field at the location of any microphone in the microphone array can be equivalent to the superposition of plane waves in all directions. The direction of the target object is determined according to the amplitude of the plane wave. S3: According to the angle and sound signal of the target sound source, the phase and gain of the sound signal are adjusted through the super-directional beamforming algorithm to enhance the sound signal of the target object and suppress the noise in other directions; S4: Extract features from the enhanced sound signal, and determine whether the target object is a drone based on a convolutional neural network.

2. A method for detecting unmanned aerial vehicles based on super-directional beamforming according to claim 1, characterized in that: The microphone array is composed of a plurality of microphones arranged linearly and is used to collect sound signals. The spatial coordinate system takes the left end of the microphone array as the coordinate center and extends to the right end as the positive direction of the z-axis.

3. A method for detecting unmanned aerial vehicles based on super-directional beamforming according to claim 1, characterized in that: In the step S2, the plane wave amplitude in a certain direction is calculated at any microphone, and then traverses from 0 degrees to 180 degrees, each angle corresponds to a plane wave amplitude, and the angle corresponding to the maximum value is the incident angle of the sound signal of the target object.

4. A method for detecting unmanned aerial vehicles based on super-directional beamforming according to claim 1, characterized in that: In the step S3, the super-directional beamforming algorithm mainly calculates high-order derivatives in the wave number domain to obtain weight values, brings the weight values ​​into the beamformer, and outputs the processed sound signal. The specific steps include: S31: approximating the sound field to a finite superposition of plane waves, associating the sound signals received by the multiple microphones through a wave equation, reconstructing the sound field, and differentiating the reconstructed sound field to obtain a high-order derivative; S32: Calculate the high-order derivative by multiplication to obtain a weight value of the sound signal, bring it into the beamformer, and output an enhanced signal.

5. A method for detecting unmanned aerial vehicles based on super-directional beamforming according to claim 1, characterized in that: In the step S4, MFCC is used to extract the first-order dynamic response features of the processed sound signal, and then a convolutional neural network modified based on AlexNet is used to identify the processed sound signal, and the result of whether the target object is a drone is output.

6. A method for detecting unmanned aerial vehicles based on super-directional beamforming according to claim 4, characterized in that: The beamformer is obtained by mixing a frequency-independent beamformer with a traditional delay-sum beamformer, and still has a high signal-to-noise ratio for high-frequency signals.

Images