Unmanned aerial vehicle acoustic identification method and system based on frequency domain analysis and machine learning

Abstract

The invention provides an unmanned aerial vehicle acoustic recognition method and system based on frequency domain analysis and machine learning. The unmanned aerial vehicle acoustic recognition method comprises the steps of obtaining a to-be-recognized sound signal; performing windowing processing, Fourier transform processing and parabola interpolation processing on the sound signal to be identified, and determining a plurality of inter-harmonic frequency characteristics in the sound signal to be identified; and inputting the plurality of inter-harmonic frequency characteristics into a pre-trained SVM classification model, calculating Euclidean distances between the plurality of inter-harmonic frequency characteristics and two sides of an optimal hyperplane, and determining whether the unmanned aerial vehicle exists in the sound signal to be identified. The unmanned aerial vehicle is identified by using the sound signal, so that the identification accuracy of the unmanned aerial vehicle can be improved.

Description

Technical Field

[0001] The present application relates to the technical field of drone identification, and in particular to a drone acoustic identification method and system based on frequency domain analysis and machine learning. Background Art

[0002] The core goal of drone detection and identification is to detect, analyze and identify targets using visual, audio, radar, and radio frequency sensors. Among them, visual sensors are used to capture and process visible light and other band images; audio sensors sense environmental sounds and convert them into electrical signals; radar sensors emit pulses and receive echoes; and radio frequency sensors use radio signals to sense characteristics such as distance, position, movement, and temperature. Among the above detection and identification methods, the accuracy of the vision-based detection and identification method will change due to factors such as lighting conditions, weather, angle, and distance, especially in the dark, the accuracy will drop sharply. Radar-based detection and identification technology relies on measuring the scattering cross-sectional area to identify targets through electromagnetic backscattering. Due to the small size of drones and the use of low-conductivity materials, the target echo is hidden in strong clutter or noise background, and the signal-to-clutter ratio is low; the target maneuvering flight causes Doppler diffusion, and the target echo is difficult to accumulate; the radar echo is weak, and it is difficult to extract and estimate target features; the radar refinement process faces challenges, and target classification and identification is difficult. Detection and identification technology based on RF detection signals uses RF signals emitted by drones during telemetry, communication or navigation to determine the presence, type, location, behavior, and possible threats of drones. Drones usually contain an onboard transmitter that exchanges data via RF signals to control and operate the drone. Since drones use specific frequencies in the RF spectrum, drone communications can be distinguished from all radio frequencies emitted by other devices (such as smartphones and laptops). In the process of drone RF detection and identification, the core is to capture its RF signal. However, when the intruding drone flies along a pre-set route (that is, the onboard transmitter does not emit RF signals), this method will lose its original effect. Therefore, how to improve the accuracy of drone identification has become a technical issue that cannot be underestimated. Summary of the invention

[0003] In view of this, the purpose of this application is to provide a drone acoustic identification method and system based on frequency domain analysis and machine learning, which can improve the accuracy of drone identification by using sound signals.

[0004] The embodiment of the present application provides a drone acoustic recognition method based on frequency domain analysis and machine learning, and the drone acoustic recognition method includes:

[0005] Acquiring a sound signal to be recognized;

[0006] Performing windowing processing, Fourier transform processing, and parabola interpolation processing on the sound signal to be identified, so as to determine a plurality of interharmonic frequency features in the sound signal to be identified;

[0007] The plurality of interharmonic frequency features are input into a pre-trained SVM classification model, and the Euclidean distances between the plurality of interharmonic frequency features and both sides of the optimal hyperplane are calculated to determine whether a drone exists in the sound signal to be identified.

[0008] In a possible implementation manner, performing windowing processing, Fourier transform processing, and parabola interpolation processing on the sound signal to be identified to determine a plurality of interharmonic frequency features in the sound signal to be identified includes:

[0009] Performing windowing processing on the sound signal to be identified to determine the sound signal to be identified after windowing;

[0010] Performing Fourier transform processing on the windowed sound signal to be identified to determine the windowed power spectrum density of the windowed sound signal to be identified;

[0011] Screening out multiple frequency points of maximum windowed power spectrum density values ​​from the windowed power spectrum density, performing parabolic interpolation processing on each frequency point of the maximum windowed power spectrum density value, and determining the interpolated peak position corresponding to the frequency point of the maximum windowed power spectrum density value;

[0012] For each frequency point of the maximum windowed power spectrum density value, an estimation process is performed based on the interpolated peak position, sampling frequency and frequency point of the maximum windowed power spectrum density value to determine the interharmonic frequency characteristics.

[0013] In a possible implementation manner, the interpolated peak position corresponding to the frequency point of the maximum windowed power spectrum density value is determined by the following formula:

[0014]

[0015] Wherein, m is the interpolated peak position corresponding to the frequency point of the maximum windowed power spectrum density value, and PSD(K) is the frequency point of the maximum windowed power spectrum density value.

[0016] In a possible implementation manner, the interharmonic frequency characteristic is determined by the following formula:

[0017]

[0018] Among them, f estis the interharmonic frequency feature, k is the frequency point of the maximum windowed power spectrum density value, m is the interpolated peak position corresponding to the frequency point of the maximum windowed power spectrum density value, and f s is the sampling frequency, and N is the number of sampling points.

[0019] In a possible implementation, the SVM classification model is trained by the following steps:

[0020] Inputting the inter-sample harmonic frequency features of the plurality of sample sound signals under two categories into the initial SVM classification model, translating the initial classification hyperplane to the inter-sample harmonic frequency features under the two categories, and determining the inter-sample harmonic frequency features under each category that are closest to the initial classification hyperplane;

[0021] Based on the inter-sample harmonic frequency features closest to the initial classification hyperplane under each category, a positive classification hyperplane and a negative classification hyperplane are determined; wherein the categories include the presence of a drone and the absence of a drone;

[0022] On the basis of maximizing the geometric interval between the positive classification hyperplane and the negative classification hyperplane, solving the initial classification hyperplane based on the Lagrange multiplier method and the dual problem solution to determine the optimal classification hyperplane;

[0023] Classify the inter-sample harmonic frequency features under each category based on the optimal classification hyperplane to determine a predicted classification result;

[0024] The initial SVM classification model is iteratively adjusted based on the predicted classification result and the actual classification result corresponding to the inter-sample harmonic frequency feature to determine the SVM classification model.

[0025] In a possible implementation manner, the iterative adjustment of the initial SVM classification model based on the classification result and the true classification result corresponding to the inter-sample harmonic frequency feature to determine the SVM classification model includes:

[0026] Determining a loss value between a classification result of the inter-sample harmonic frequency feature and a true classification result of the inter-sample harmonic frequency feature;

[0027] If the loss value is greater than or equal to a preset threshold, the network parameters of the initial SVM classification model are changed; if the loss value is less than the preset threshold, the change of the network parameters of the initial SVM classification model is stopped to determine the SVM classification model.

[0028] The embodiment of the present application further provides a drone identification device based on frequency domain analysis and machine learning, the drone identification device comprising:

[0029] An acquisition module, used for acquiring a sound signal to be recognized;

[0030] A frequency processing module, used for performing windowing processing, Fourier transform processing and parabola interpolation processing on the sound signal to be identified, so as to determine a plurality of interharmonic frequency features in the sound signal to be identified;

[0031] The determination module is used to input the multiple interharmonic frequency features into a pre-trained SVM classification model, calculate the Euclidean distance between the multiple interharmonic frequency features and the two sides of the optimal hyperplane, and determine whether there is a drone in the sound signal to be identified.

[0032] In a possible implementation manner, the frequency processing module performs windowing processing, Fourier transform processing, and parabola interpolation processing on the sound signal to be identified to determine multiple interharmonic frequency features in the sound signal to be identified, including:

[0033] Performing windowing processing on the sound signal to be identified to determine the sound signal to be identified after windowing;

[0034] Performing Fourier transform processing on the windowed sound signal to be identified to determine the windowed power spectrum density of the windowed sound signal to be identified;

[0035] Screening out multiple frequency points of maximum windowed power spectrum density values ​​from the windowed power spectrum density, performing parabolic interpolation processing on each frequency point of the maximum windowed power spectrum density value, and determining the interpolated peak position corresponding to the frequency point of the maximum windowed power spectrum density value;

[0036] For each frequency point of the maximum windowed power spectrum density value, an estimation process is performed based on the interpolated peak position, sampling frequency and frequency point of the maximum windowed power spectrum density value to determine the interharmonic frequency characteristics.

[0037] An embodiment of the present application also provides an electronic device, including: a processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor and the memory communicate through the bus, and when the machine-readable instructions are executed by the processor, the steps of the above-mentioned drone acoustic identification method based on frequency domain analysis and machine learning are performed.

[0038] An embodiment of the present application also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the drone acoustic recognition method based on frequency domain analysis and machine learning are executed.

[0039] The embodiment of the present application provides a method and system for drone acoustic identification based on frequency domain analysis and machine learning, the method comprising: obtaining a sound signal to be identified; performing windowing, Fourier transform and parabola interpolation processing on the sound signal to be identified, and determining multiple interharmonic frequency features in the sound signal to be identified; inputting the multiple interharmonic frequency features into a pre-trained SVM classification model, calculating the Euclidean distance between the multiple interharmonic frequency features and both sides of the optimal hyperplane, and determining whether there is a drone in the sound signal to be identified. Using sound signals to identify drones can improve the accuracy of drone identification.

[0040] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, preferred embodiments are specifically cited below and described in detail with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0042] Figure 1 A flowchart of a drone acoustic recognition method based on frequency domain analysis and machine learning provided in an embodiment of the present application;

[0043] Figure 2 One of the structural schematic diagrams of a drone acoustic recognition system based on frequency domain analysis and machine learning provided in an embodiment of the present application;

[0044] Figure 3 The second structural diagram of a drone acoustic recognition system based on frequency domain analysis and machine learning provided in an embodiment of the present application;

[0045] Figure 4 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0046] To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. The components of the embodiments of the present application usually described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, each other embodiment obtained by those skilled in the art without making creative work belongs to the scope of protection of the present application.

[0047] First, the application scenarios to which this application is applicable are introduced. This application can be applied to the field of drone identification technology.

[0048] Research has found that the core goal of drone detection and identification is to use visual, audio, radar, radio frequency and other sensors to detect, analyze and identify targets. Among them, visual sensors are used to capture and process visible light and other band images; audio sensors sense environmental sounds and convert them into electrical signals; radar sensors emit pulses and receive echoes; radio frequency sensors use radio signals to sense characteristics such as distance, position, movement and temperature. Among the above detection and identification methods, the accuracy of vision-based detection and identification methods will change due to factors such as lighting conditions, weather, angles, and distances, especially in the dark, the accuracy will drop sharply. Radar-based detection and identification technology relies on measuring the scattering cross-sectional area to identify targets through electromagnetic backscattering. Due to the small size of drones and the use of low-conductivity materials, the target echo is hidden in strong clutter or noise background, and the signal-to-clutter ratio is low; the target maneuvering flight causes Doppler diffusion, and the target echo is difficult to accumulate; the radar echo is weak, and the target feature extraction and estimation are difficult; the radar refinement process faces challenges, and target classification and identification is difficult. Detection and identification technology based on RF detection signals uses RF signals emitted by drones during telemetry, communication or navigation to determine the presence, type, location, behavior, and possible threats of drones. Drones usually contain an onboard transmitter that exchanges data via RF signals to control and operate the drone. Since drones use specific frequencies in the RF spectrum, drone communications can be distinguished from all radio frequencies emitted by other devices (such as smartphones and laptops). In the process of drone RF detection and identification, the core is to capture its RF signal. However, when the intruding drone flies along a pre-set route (that is, the onboard transmitter does not emit RF signals), this method will lose its original effect. Therefore, how to improve the accuracy of drone identification has become a technical issue that cannot be underestimated.

[0049] Based on this, an embodiment of the present application provides a drone acoustic identification method based on frequency domain analysis and machine learning, which uses sound signals to identify drones and can improve the accuracy of drone identification.

[0050] See also Figure 1 , Figure 1 The following is a flow chart of a method for drone acoustic recognition based on frequency domain analysis and machine learning provided in an embodiment of the present application. Figure 1 As shown in , the drone acoustic recognition method provided by the embodiment of the present application includes:

[0051] S101: Acquire a sound signal to be recognized.

[0052] In this step, a sound signal to be recognized is acquired using a sound collection device.

[0053] Here, the drone sound signal can be regarded as an interharmonic signal. The commonly used interharmonic signal model can be described by the following mathematical formula:

[0054]

[0055] In formula (1), p is the number of simple harmonic components, A k represents the amplitude of the kth interharmonic component, f k represents the frequency of the kth interharmonic component, represents the phase of the kth interharmonic component, {n(t), t=1,2,…,T} represents the noise, {A k ,k=1,2,…,p} takes a positive value.

[0056] S102: performing windowing processing, Fourier transform processing and parabola interpolation processing on the sound signal to be identified, and determining a plurality of interharmonic frequency features in the sound signal to be identified.

[0057] In this step, the sound signal to be identified is subjected to windowing processing, Fourier transform processing and parabola interpolation processing to determine a plurality of interharmonic frequency features in the sound signal to be identified.

[0058] In a possible implementation manner, performing windowing processing, Fourier transform processing, and parabola interpolation processing on the sound signal to be identified to determine a plurality of interharmonic frequency features in the sound signal to be identified includes:

[0059] A: Perform windowing processing on the sound signal to be recognized to determine the sound signal to be recognized after windowing.

[0060] Here, the sound signal to be recognized is subjected to windowing processing to determine the windowed sound signal to be recognized.

[0061] The windowing function may be a Hamming window function.

[0062] Here, assuming that the window function is ω(n), the length of the window function is N, the sound signal to be recognized after windowing is:

[0063] x ω (n)=x(n)·ω(n),n=0,1,…,N-1 (2)

[0064] Where x(n) represents the time domain signal.

[0065] B: Performing Fourier transform processing on the windowed sound signal to be recognized to determine the windowed power spectrum density of the windowed sound signal to be recognized.

[0066] Here, Fourier transform processing is performed on the windowed sound signal to be identified to determine the windowed power spectrum density of the windowed sound signal to be identified.

[0067] Among them, the discrete Fourier transform is performed on the windowed sound signal to be recognized:

[0068]

[0069] Among them, X(k) is its corresponding frequency domain representation.

[0070] Among them, the amplitude X of the frequency after windowing is determined by the following formula: ω (k):

[0071]

[0072] Among them, the power spectrum PW(k) after windowing is determined by the following formula:

[0073]

[0074] Among them, the power spectrum density after windowing is determined by the following formula:

[0075]

[0076] Here, PSD(K) is the windowed power spectrum density, and the frequency points of the maximum windowed power spectrum density value corresponding to a plurality of maximum power spectrum densities are screened out from the windowed power spectrum density.

[0077] C: Filter out multiple frequency points of the maximum windowed power spectrum density value from the windowed power spectrum density, perform parabolic interpolation processing on each frequency point of the maximum windowed power spectrum density value, and determine the interpolated peak position corresponding to the frequency point of the maximum windowed power spectrum density value.

[0078] Here, multiple frequency points of the maximum windowed power spectrum density value are screened out from the windowed power spectrum density, and parabolic interpolation processing is performed on each frequency point of the maximum windowed power spectrum density value to determine the interpolated peak position corresponding to the frequency point of the maximum windowed power spectrum density value.

[0079] D: For each frequency point of the maximum windowed power spectrum density value, an estimation process is performed based on the interpolated peak position, sampling frequency and frequency point of the maximum windowed power spectrum density value to determine the interharmonic frequency characteristics.

[0080] Here, for each frequency point of the maximum windowed power spectrum density value, an estimation process is performed based on the interpolated peak position, sampling frequency and frequency point of the maximum windowed power spectrum density value to determine the interharmonic frequency characteristics.

[0081] Here, although the frequency estimation based on fast Fourier transform is easy to calculate, its resolution is limited, so the interpolation method is added here to improve the accuracy of frequency estimation.

[0082] In a possible implementation manner, the interpolated peak position corresponding to the frequency point of the maximum windowed power spectrum density value is determined by the following formula:

[0083]

[0084] Wherein, m is the interpolated peak position corresponding to the frequency point of the maximum windowed power spectrum density value, PSD(K) is the windowed power spectrum density, and k is the frequency point of the maximum windowed power spectrum density value.

[0085] In a possible implementation manner, the interharmonic frequency characteristic is determined by the following formula:

[0086]

[0087] Among them, f est is the interharmonic frequency feature, k is the frequency point of the maximum windowed power spectrum density value, m is the interpolated peak position corresponding to the frequency point of the maximum windowed power spectrum density value, and f s is the sampling frequency, and N is the number of sampling points.

[0088] In this scheme, the frequency domain analysis of the sound signal to be identified is achieved by performing windowing processing, Fourier transform processing and parabolic interpolation processing on the sound signal to be identified, and the interharmonic frequency characteristics of the sound signal to be identified are obtained. In this scheme, the feature vector extracted based on the frequency domain analysis has smaller fluctuations, avoiding the technical problem in the prior art that the feature vector of the Mel-frequency cepstral coefficient feature also fluctuates greatly when the amplitude of the acoustic signal changes greatly.

[0089] S103: Input the plurality of interharmonic frequency features into a pre-trained SVM classification model, calculate the Euclidean distances between the plurality of interharmonic frequency features and both sides of the optimal hyperplane, and determine whether a drone exists in the sound signal to be identified.

[0090] In this step, multiple interharmonic frequency features are input into a pre-trained SVM classification model, and the Euclidean distances between the multiple interharmonic frequency features and both sides of the optimal hyperplane are calculated to determine whether a drone exists in the sound signal to be identified.

[0091] In a possible implementation, the SVM classification model is trained by the following steps:

[0092] a: Input the inter-sample harmonic frequency features of multiple sample sound signals under two categories into the initial SVM classification model, translate the initial classification hyperplane to the inter-sample harmonic frequency features under the two categories, and determine the inter-sample harmonic frequency features closest to the initial classification hyperplane under each category.

[0093] Here, the inter-sample harmonic frequency features of multiple sample sound signals under two categories are input into the initial SVM classification model, the initial classification hyperplane is translated to the inter-sample harmonic frequency features under the two categories, and the inter-sample harmonic frequency features closest to the initial classification hyperplane under each category are determined.

[0094] The categories include sample sound signals with and without drones, n is the number of samples, and X is i is the i-th sample, X i =(x i1 ,x i2 ,…,x ip ) T ,i=1,2,…,n, p is the dimension of input, y i To output the recognition result of whether a drone exists, in this application, 1 is generally used as the sample label for the existence of a drone, and -1 is used as the sample label for the absence of a drone.

[0095] Here, the purpose of the SVM classification model is to find an optimal classification hyperplane W in the feature space of the sample.T X+b=0, effectively separating the two types of samples (+1 and -1), where W=(w1,w2,…,w p ), b is the bias term of the hyperplane, and X is the harmonic frequency feature between samples.

[0096] b: Based on the inter-sample harmonic frequency features closest to the initial classification hyperplane under each category, a positive classification hyperplane and a negative classification hyperplane are determined; wherein the categories include the presence of a drone and the absence of a drone.

[0097] Here, a positive classification hyperplane and a negative classification hyperplane are determined according to the inter-sample harmonic frequency features closest to the initial classification hyperplane in each category.

[0098] Among them, the categories include the presence of drones and the absence of drones.

[0099] Here, the Euclidean distance d from the harmonic frequency feature between samples to the hyperplane is:

[0100]

[0101] The initial classification hyperplane is translated to the point in the two types of samples that is closest to the hyperplane to form a positive classification hyperplane and a negative classification hyperplane. The geometric interval between the positive classification hyperplane and the negative classification hyperplane is:

[0102]

[0103] c: On the basis of maximizing the geometric interval between the positive classification hyperplane and the negative classification hyperplane, the initial classification hyperplane is solved based on the Lagrange multiplier method and the dual problem solution to determine the optimal classification hyperplane.

[0104] Here, on the basis of maximizing the geometric interval between the positive classification hyperplane and the negative classification hyperplane, the initial classification hyperplane is solved according to the Lagrange multiplier method and the dual problem solution to determine the optimal classification hyperplane.

[0105] Specifically, ||W|| cannot be infinitely small, and the correct classification of the hyperplane means that the distance from point i to the hyperplane is always greater than or equal to d. So there is Multiply both sides by ||W||, simplifying to: |WX i +b|=y i (WX i +b)≥1. Therefore, the objective function of the convex quadratic programming problem can be written as shown in formula (11) and the constraints as shown in formula (12):

[0106]

[0107] yi (WX i +b)-1≥0,i=1,2,…,n (12)

[0108] Here, the Lagrange multiplier method is used to solve and construct the Lagrange function. The n inequality constraints must be added with the Lagrange function λ i ≥0,λ=(λ1,λ2,…,λ n ) T , that is:

[0109]

[0110] Among them, first find the partial derivative of the parameters:

[0111]

[0112]

[0113] Here, the problem is transformed into a dual problem and the formula (14) is substituted into the formula (13) to obtain:

[0114]

[0115] Among them, the constrained optimization problem is transformed into the objective function as shown in formula (16), and the constraint conditions are shown in formula (17) and formula (18):

[0116]

[0117] 0≤λ i ≤C,i=1,2,…,n (18)

[0118] Wherein, C≥0 in formula (18) is the penalty factor.

[0119] Here, the kernel technique in the SVM model is used to map the feature vector to a higher dimensional space, and the inner product of the sample feature vector in formula (16) is expressed as i T X j Use the kernel technique for replacement, using a cubic polynomial kernel function:

[0120] K(X i T ,X j )=(X i T X j +1) 3 (19)(23)

[0121] The final recognition result of the drone acoustic signal can be obtained by the following formula:

[0122]

[0123] in, sign(·) is the sign function, and y is the output classification result.

[0124] d: Classify the inter-sample harmonic frequency features under each category based on the optimal classification hyperplane to determine a predicted classification result.

[0125] Here, the inter-sample harmonic frequency features under each category are classified and solved according to the optimal classification hyperplane to determine the predicted classification result.

[0126] e: iteratively adjusting the initial SVM classification model based on the predicted classification result and the actual classification result corresponding to the inter-sample harmonic frequency feature to determine the SVM classification model.

[0127] Here, the initial SVM classification model is iteratively adjusted according to the predicted classification results and the actual classification results corresponding to the inter-sample harmonic frequency characteristics to determine the SVM classification model.

[0128] In a possible implementation manner, the iterative adjustment of the initial SVM classification model based on the classification result and the true classification result corresponding to the inter-sample harmonic frequency feature to determine the SVM classification model includes:

[0129] i: Determine the loss value between the classification result of the inter-sample harmonic frequency feature and the true classification result of the inter-sample harmonic frequency feature.

[0130] Here, the loss value between the classification result of the inter-sample harmonic frequency feature and the true classification result of the inter-sample harmonic frequency feature is determined.

[0131] ii: If the loss value is greater than or equal to a preset threshold, the network parameters of the initial SVM classification model are changed; if the loss value is less than the preset threshold, the change of the network parameters of the initial SVM classification model is stopped to determine the SVM classification model.

[0132] Here, if the loss value is greater than or equal to the preset threshold, the network parameters of the initial SVM classification model are changed. If the loss value is less than the preset threshold, the change of the network parameters of the initial SVM classification model is stopped to determine the SVM classification model.

[0133] In this scheme, the feature data extracted based on frequency domain analysis is put into the SVM classification model for identification. The SVM model can distinguish the drone data from the noise data very well.

[0134] This application utilizes the acoustic characteristics of drones. Compared with other acoustic feature extraction methods, it has the advantages of obvious useful information extracted and small fluctuations in feature vectors. It can improve the accuracy of drone acoustic recognition and has the following advantages. 1. The extracted feature vector has small fluctuations: Traditional acoustic feature extraction methods such as Mel-frequency cepstral coefficients only have small fluctuations in the extracted feature vectors when the distance between the drone and the microphone sensor is relatively unchanged. When the drone is flying (that is, the distance between the drone and the microphone sensor is not a relatively fixed distance), the feature vector extracted using Mel-frequency cepstral coefficients has large fluctuations, which is not conducive to subsequent classification processing. The feature vector extracted based on frequency domain analysis proposed in this scheme has small fluctuations, which is conducive to subsequent classification processing. 2. Small amount of calculation: Traditional acoustic feature extraction methods such as Mel-frequency cepstral coefficients (MFCC) and gammatone cepstral coefficients (GFCC) have large amounts of calculation because these methods involve a series of complex signal processing steps. The calculation process of MFCC includes: performing a fast Fourier transform (FFT) on the signal to convert it into an energy distribution in the frequency domain, then applying a set of Mel filters to filter the input signal, and finally performing a discrete cosine transform (DCT) on the signal energy output by each filter to obtain the MFCC parameters. These steps require a lot of calculations. The calculation process of the gammatone cepstral coefficient (GFCC) is similar to the Mel cepstral coefficient (MFCC). The feature extraction method based on frequency domain analysis does not require a series of complex signal processing steps, so the amount of calculation is small. 3. High proportion of useful information: The features extracted by the traditional acoustic feature extraction method contain a lot of irrelevant information (that is, information that does not help detection and recognition). Traditional acoustic features such as Mel cepstral coefficients contain too much information, and useful information and irrelevant information are mixed. Mel cepstral coefficients contain time domain information and frequency domain information. Time domain information is irrelevant information because the time domain information contains amplitude information. When the drone is at different distances from the microphone sensor, the amplitude will change dramatically, which is not the information we want to extract. The feature information extracted by the feature extraction method based on frequency domain analysis avoids this irrelevant information, and the proportion of useful information is high. 4. Utilize the unique properties of drone sounds: drone sounds are usually divided into aerodynamic sounds and mechanical sounds. The former originates from the rotor cutting the air and the friction between the body and the air. It is usually distributed in the low-frequency part of the sound signal and decays slowly. The latter originates from the motor and transmission parts. It is usually distributed in the high-frequency part of the sound signal and decays quickly. In addition, wind and rain sounds, as the most common background noise in the actual environment, are broadband noise signals. Their signal characteristics are significantly different from those of drone signals and are easy to distinguish in the time domain and frequency domain. Therefore, the unique properties of drone sounds can be utilized to effectively distinguish them from environmental noise.

[0135] The embodiment of the present application provides a method for drone acoustic identification based on frequency domain analysis and machine learning, the method comprising: obtaining a sound signal to be identified; performing windowing, Fourier transform and parabola interpolation processing on the sound signal to be identified, and determining multiple interharmonic frequency features in the sound signal to be identified; inputting the multiple interharmonic frequency features into a pre-trained SVM classification model, calculating the Euclidean distance between the multiple interharmonic frequency features and both sides of the optimal hyperplane, and determining whether there is a drone in the sound signal to be identified. Using sound signals to identify drones can improve the accuracy of drone identification.

[0136] See also Figure 2 , Figure 3 , Figure 2 One of the structural schematic diagrams of a drone acoustic recognition system based on frequency domain analysis and machine learning provided in an embodiment of the present application; Figure 3 The second structural diagram of a drone acoustic recognition system based on frequency domain analysis and machine learning provided in the embodiment of the present application. Figure 2 As shown in , the drone acoustic identification system 200 includes:

[0137] An acquisition module 210 is used to acquire a sound signal to be recognized;

[0138] The frequency processing module 220 is used to perform windowing processing, Fourier transform processing and parabola interpolation processing on the sound signal to be identified, so as to determine a plurality of interharmonic frequency features in the sound signal to be identified;

[0139] The determination module 230 is used to input the multiple interharmonic frequency features into a pre-trained SVM classification model, calculate the Euclidean distance between the multiple interharmonic frequency features and the two sides of the optimal hyperplane, and determine whether there is a drone in the sound signal to be identified.

[0140] Furthermore, the frequency processing module performs windowing processing, Fourier transform processing and parabola interpolation processing on the sound signal to be identified to determine multiple interharmonic frequency features in the sound signal to be identified, including:

[0141] Performing windowing processing on the sound signal to be identified to determine the sound signal to be identified after windowing;

[0142] Performing Fourier transform processing on the windowed sound signal to be identified to determine the windowed power spectrum density of the windowed sound signal to be identified;

[0143] Screening out multiple frequency points of maximum windowed power spectrum density values ​​from the windowed power spectrum density, performing parabolic interpolation processing on each frequency point of the maximum windowed power spectrum density value, and determining the interpolated peak position corresponding to the frequency point of the maximum windowed power spectrum density value;

[0144] For each frequency point of the maximum windowed power spectrum density value, an estimation process is performed based on the interpolated peak position, sampling frequency and frequency point of the maximum windowed power spectrum density value to determine the interharmonic frequency characteristics.

[0145] Furthermore, the frequency processing module determines the interpolated peak position corresponding to the frequency point of the maximum windowed power spectrum density value by the following formula:

[0146]

[0147] Wherein, m is the interpolated peak position corresponding to the frequency point of the maximum windowed power spectrum density value, PSD(K) is the windowed power spectrum density, and k is the frequency point of the maximum windowed power spectrum density value.

[0148] Furthermore, the frequency processing module determines the interharmonic frequency characteristics by the following formula:

[0149]

[0150] Among them, f est is the interharmonic frequency feature, k is the frequency point of the maximum windowed power spectrum density value, m is the interpolated peak position corresponding to the frequency point of the maximum windowed power spectrum density value, and f s is the sampling frequency, and N is the number of sampling points.

[0151] Further, such as Figure 3 As shown, the drone acoustic recognition system 200 further includes a model training module 240, which trains the SVM classification model through the following steps:

[0152] Inputting the inter-sample harmonic frequency features of the plurality of sample sound signals under two categories into the initial SVM classification model, translating the initial classification hyperplane to the inter-sample harmonic frequency features under the two categories, and determining the inter-sample harmonic frequency features under each category that are closest to the initial classification hyperplane;

[0153] Based on the inter-sample harmonic frequency features closest to the initial classification hyperplane under each category, a positive classification hyperplane and a negative classification hyperplane are determined; wherein the categories include the presence of a drone and the absence of a drone;

[0154] On the basis of maximizing the geometric interval between the positive classification hyperplane and the negative classification hyperplane, solving the initial classification hyperplane based on the Lagrange multiplier method and the dual problem solution to determine the optimal classification hyperplane;

[0155] Classify the inter-sample harmonic frequency features under each category based on the optimal classification hyperplane to determine a predicted classification result;

[0156] The initial SVM classification model is iteratively adjusted based on the predicted classification result and the actual classification result corresponding to the inter-sample harmonic frequency feature to determine the SVM classification model.

[0157] Furthermore, the model training module is used to iteratively adjust the initial SVM classification model based on the classification result and the real classification result corresponding to the inter-sample harmonic frequency feature to determine the SVM classification model:

[0158] Determining a loss value between a classification result of the inter-sample harmonic frequency feature and a true classification result of the inter-sample harmonic frequency feature;

[0159] If the loss value is greater than or equal to a preset threshold, the network parameters of the initial SVM classification model are changed; if the loss value is less than the preset threshold, the change of the network parameters of the initial SVM classification model is stopped to determine the SVM classification model.

[0160] The embodiment of the present application provides a drone acoustic identification system based on frequency domain analysis and machine learning, and the drone identification system includes: an acquisition module for acquiring a sound signal to be identified; a frequency processing module for performing windowing, Fourier transform and parabola interpolation processing on the sound signal to be identified, and determining multiple interharmonic frequency features in the sound signal to be identified; a determination module for inputting the multiple interharmonic frequency features into a pre-trained SVM classification model, calculating the Euclidean distance between the multiple interharmonic frequency features and the two sides of the optimal hyperplane, and determining whether there is a drone in the sound signal to be identified. Using sound signals to identify drones can improve the accuracy of drone identification.

[0161] See also Figure 4 , Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. Figure 4 As shown in , the electronic device 400 includes a processor 410 , a memory 420 and a bus 430 .

[0162] The memory 420 stores machine-readable instructions executable by the processor 410. When the electronic device 400 is running, the processor 410 communicates with the memory 420 via the bus 430. When the machine-readable instructions are executed by the processor 410, the above-mentioned Figure 1 The specific implementation of the steps of the drone acoustic identification method based on frequency domain analysis and machine learning in the method embodiment shown can be found in the method embodiment, and will not be repeated here.

[0163] The present application also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the computer program can execute the above-mentioned Figure 1 The specific implementation of the steps of the drone acoustic identification method based on frequency domain analysis and machine learning in the method embodiment shown can be found in the method embodiment, and will not be repeated here.

[0164] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0165] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, and the indirect coupling or communication connection of devices or units can be electrical, mechanical or other forms.

[0166] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0167] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0168] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium that is executable by a processor. Based on this understanding, the technical solution of the present application can essentially be embodied in the form of a software product, or in other words, the part that contributes to the prior art or the part of the technical solution. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0169] Finally, it should be noted that the above-described embodiments are only specific implementation methods of the present application, which are used to illustrate the technical solutions of the present application, rather than to limit them. The protection scope of the present application is not limited thereto. Although the present application is described in detail with reference to the above-mentioned embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-mentioned embodiments within the technical scope disclosed in the present application, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application. Therefore, the protection scope of the present application shall be based on the protection scope of the claims.

Claims

1. A drone acoustic recognition method based on frequency domain analysis and machine learning, characterized in that: The drone acoustic identification method comprises: Acquiring a sound signal to be recognized; Performing windowing processing, Fourier transform processing, and parabola interpolation processing on the sound signal to be identified, so as to determine a plurality of interharmonic frequency features in the sound signal to be identified; The plurality of interharmonic frequency features are input into a pre-trained SVM classification model, and the Euclidean distances between the plurality of interharmonic frequency features and both sides of the optimal hyperplane are calculated to determine whether a drone exists in the sound signal to be identified.

2. The drone acoustic identification method according to claim 1, characterized in that: The performing windowing processing, Fourier transform processing and parabola interpolation processing on the sound signal to be identified to determine a plurality of interharmonic frequency features in the sound signal to be identified includes: Performing windowing processing on the sound signal to be identified to determine the sound signal to be identified after windowing; Performing Fourier transform processing on the windowed sound signal to be identified to determine the windowed power spectrum density of the windowed sound signal to be identified; Screening out multiple frequency points of maximum windowed power spectrum density values ​​from the windowed power spectrum density, performing parabolic interpolation processing on each frequency point of the maximum windowed power spectrum density value, and determining the interpolated peak position corresponding to the frequency point of the maximum windowed power spectrum density value; For each frequency point of the maximum windowed power spectrum density value, an estimation process is performed based on the interpolated peak position, sampling frequency and frequency point of the maximum windowed power spectrum density value to determine the interharmonic frequency characteristics.

3. The drone acoustic identification method according to claim 2, characterized in that: The interpolated peak position corresponding to the frequency point of the maximum windowed power spectrum density value is determined by the following formula: Wherein, m is the interpolated peak position corresponding to the frequency point of the maximum windowed power spectrum density value, PSD(K) is the windowed power spectrum density, and k is the frequency point of the maximum windowed power spectrum density value.

4. The drone acoustic identification method according to claim 2, characterized in that: The interharmonic frequency characteristics are determined by the following formula: Among them, f est is the interharmonic frequency feature, k is the frequency point of the maximum windowed power spectrum density value, m is the interpolated peak position corresponding to the frequency point of the maximum windowed power spectrum density value, and f s is the sampling frequency, and N is the number of sampling points.

5. The drone acoustic identification method according to claim 1, characterized in that: The SVM classification model is trained by the following steps: Inputting the inter-sample harmonic frequency features of the plurality of sample sound signals under two categories into the initial SVM classification model, translating the initial classification hyperplane to the inter-sample harmonic frequency features under the two categories, and determining the inter-sample harmonic frequency features under each category that are closest to the initial classification hyperplane; Based on the inter-sample harmonic frequency features closest to the initial classification hyperplane under each category, a positive classification hyperplane and a negative classification hyperplane are determined; wherein the categories include the presence of a drone and the absence of a drone; On the basis of maximizing the geometric interval between the positive classification hyperplane and the negative classification hyperplane, solving the initial classification hyperplane based on the Lagrange multiplier method and the dual problem solution to determine the optimal classification hyperplane; Classify the inter-sample harmonic frequency features under each category based on the optimal classification hyperplane to determine a predicted classification result; The initial SVM classification model is iteratively adjusted based on the predicted classification result and the actual classification result corresponding to the inter-sample harmonic frequency feature to determine the SVM classification model.

6. The drone acoustic identification method according to claim 5, characterized in that: The iterative adjustment of the initial SVM classification model based on the classification result and the real classification result corresponding to the inter-sample harmonic frequency feature to determine the SVM classification model includes: Determining a loss value between a classification result of the inter-sample harmonic frequency feature and a true classification result of the inter-sample harmonic frequency feature; If the loss value is greater than or equal to a preset threshold, the network parameters of the initial SVM classification model are changed; if the loss value is less than the preset threshold, the change of the network parameters of the initial SVM classification model is stopped to determine the SVM classification model.

7. A drone identification system based on frequency domain analysis and machine learning, characterized in that: The drone acoustic identification system comprises: An acquisition module, used for acquiring a sound signal to be recognized; A frequency processing module, used for performing windowing processing, Fourier transform processing and parabola interpolation processing on the sound signal to be identified, so as to determine a plurality of interharmonic frequency features in the sound signal to be identified; The determination module is used to input the multiple interharmonic frequency features into a pre-trained SVM classification model, calculate the Euclidean distance between the multiple interharmonic frequency features and the two sides of the optimal hyperplane, and determine whether there is a drone in the sound signal to be identified.

8. The drone acoustic identification system according to claim 7, characterized in that: The frequency processing module performs windowing processing, Fourier transform processing and parabola interpolation processing on the sound signal to be identified to determine multiple interharmonic frequency features in the sound signal to be identified, including: Performing windowing processing on the sound signal to be identified to determine the sound signal to be identified after windowing; Performing Fourier transform processing on the windowed sound signal to be identified to determine the windowed power spectrum density of the windowed sound signal to be identified; Screening out multiple frequency points of maximum windowed power spectrum density values ​​from the windowed power spectrum density, performing parabolic interpolation processing on each frequency point of the maximum windowed power spectrum density value, and determining the interpolated peak position corresponding to the frequency point of the maximum windowed power spectrum density value; For each frequency point of the maximum windowed power spectrum density value, an estimation process is performed based on the interpolated peak position, sampling frequency and frequency point of the maximum windowed power spectrum density value to determine the interharmonic frequency characteristics.

9. An electronic device, characterized in that: include: A processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor and the memory communicate through the bus, and the machine-readable instructions are executed by the processor to execute the steps of the drone acoustic identification method based on frequency domain analysis and machine learning as described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, the steps of the drone acoustic recognition method based on frequency domain analysis and machine learning as described in any one of claims 1 to 6 are executed.

Images