Unmanned aerial vehicle recognition method, device and equipment based on voiceprint and storage medium

Through complex convolutional recurrent neural networks, the problem of drone recognition in dark and radio frequency signals is solved, and efficient drone soundprint recognition is achieved, especially in complex convolutional recurrent neural networks. After training, the recognition rate reached 96.21%.

CN118969000APending Publication Date: 2024-11-15ARMY ENG UNIV OF PLA
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202411164113.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-23
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The existing drone recognition technology has poor recognition results in dark environments and limited RF signal environments, especially for drones with pre-set cruise routes without RF signal control.

Method used

Using a drone recognition method based on voiceprint, the acoustic signals of the drone are converted into complex time spectrum features using complex convolutional recurrent neural networks. Through differentiable Fourier variation module, complex convolution module, complex length and short time memory module It is processed with a fully connected classification module to realize the identification of drone types.

Benefits of technology

Under dark environments and limited radio frequency signals, the drone type can be accurately identified, which improves the accuracy of drone recognition, especially after training in complex convolutional recurrent neural networks, the recognition rate reaches 96.21%.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118969000A_ABST
    Figure CN118969000A_ABST
Patent Text Reader

Abstract

The invention discloses a voiceprint-based unmanned aerial vehicle recognition method, device and equipment and a storage medium, and the method comprises the steps: inputting an unmanned aerial vehicle flight sound signal with an unknown label into a trained complex convolutional recurrent neural network, and obtaining an unmanned aerial vehicle classification result. According to the method, the complex convolutional recurrent neural network is innovatively used, the network combines the advantages of a convolutional network and a recurrent network, and operation is performed on the complex time-frequency features of the flight sound signals of the unmanned aerial vehicle, so that the type features of the unmanned aerial vehicle can be extracted more effectively; according to the method, the problem of low performance of an image-based and radio frequency signal-based voiceprint-based unmanned aerial vehicle recognition method caused by a dark environment or a wireless signal limited environment is solved, and a new solution is provided for monitoring, management and control of the unmanned aerial vehicle by combining a sound signal and a deep learning technology.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a method, device, equipment and storage medium for unmanned aerial vehicle identification based on voiceprint, and belongs to the technical field of sound signal processing. Background Art

[0002] With the rapid development of drone technology, its application in agriculture, mining, geology and other industries has become increasingly popular, greatly promoting the improvement of productivity in these fields. However, with the popularization of drone technology and the expansion of its application fields, we are also facing some challenges that cannot be ignored. Drones are used by criminals to carry out drug smuggling activities, which seriously endangers the security and stability of society. At the same time, drones may also interfere with civil aviation routes and navigation systems, posing a potential threat to aviation safety. In order to meet these challenges, we must strengthen the supervision of drones and ensure that drones are used within a legal and safe range. To this end, it is particularly important to introduce a drone identification system.

[0003] The existing mainstream drone recognition technologies are based on image processing and RF signal analysis. Image-based drone recognition technology mainly relies on high-definition cameras, computer vision algorithms, and image analysis technology to capture real-time images of drones flying in the air. However, dark environments will lead to poor application effects of image processing-based drone recognition technology. For example: Researchers (Yuan Yuan, Research on Complex Target Tracking Algorithms in UAV Environments [D]. Anhui University, 2023.) proposed a target tracking algorithm based on feature residual automatic spatiotemporal regularization correlation filters for tracking drones, but it is easy to lose the target in insufficient lighting. RF signal-based drone recognition technology focuses on the detection and analysis of drone communication signals. Although this type of technology can work effectively under conditions of obstructed vision or at night, it is not effective for drones with pre-set cruise routes that do not require RF signal control. For example, researchers (Bisio I, Garibotto C, Lavagetto F et al. Blind detection: Advancedtechniques for WiFi-based drone surveillance. IEEE Trans Veh Technol, 2018.) identified drones by collecting surrounding RF signals. However, since drone signals use frequency hopping output, the detection of RF signals becomes extremely complex and technically difficult, and the operation process is cumbersome. What is more difficult is that when a drone enters "silent" mode, its signal will be difficult to capture, resulting in the failure of the detection method. Summary of the invention

[0004] The purpose of the present invention is to overcome the severe challenges faced by drone identification in environments with insufficient lighting such as darkness and limited radio frequency signals, and to provide a drone identification method, device, equipment and storage medium based on voiceprint to improve the accuracy of drone identification.

[0005] The present invention provides a drone identification method based on voiceprint, comprising: Acquire drone acoustic signals; Input the drone acoustic signal into the trained complex convolutional recurrent neural network to obtain the drone type recognition result; Among them, the complex convolutional recurrent neural network includes a differentiable Fourier transform module, a complex convolution module, a complex long short-term memory module and a fully connected classification module; the differentiable Fourier transform module converts the time domain waveform characteristics of the drone sound signal into the complex time-frequency spectrum characteristics of the drone sound signal; several complex convolution modules extract the complex time-frequency spectrum characteristics of the drone sound signal to obtain deep complex features; the complex long short-term memory module extracts the deep complex features to obtain the time-series related complex features of the drone sound signal; the fully connected classification module processes the time-series related complex features of the drone sound signal to obtain a vector with the same dimension as the number of drone types to represent the recognition probability of each type of drone.

[0006] Optionally, the complex convolution module includes a complex convolution layer, a complex batch normalization layer, and a complex PReLU layer connected in sequence, and the calculation formula of the complex convolution layer is:

[0007] In the formula, represents the output of the complex convolutional layer, represents the input of the complex convolutional layer, , Respectively The real and imaginary parts of represents the convolution kernel of the complex convolutional layer, , Represent the real and imaginary parts of the convolution kernel respectively, represents the complex convolution operation, Represents the convolution operation; The calculation formula of the complex batch normalization layer is:

[0008] In the formula, represents the output of the complex batch normalization layer, , Respectively The real and imaginary parts of represents the complex batch normalization operation, Represents batch normalization operation; The calculation formula of the complex PReLU layer is:

[0009] In the formula, represents the output of the complex PReLU layer, , Respectively The real and imaginary parts of represents a complex PReLU operation, express operate.

[0010] Optionally, the plural long short-term memory module includes two sequentially connected long short-term memory network layers, and the calculation formula of the plural long short-term memory module is:

[0011]

[0012] in, represents the output of the complex long short-term memory module, represents the input of the complex long short-term memory module, , Respectively The real and imaginary parts of Indicates complex long and short-term memory operations, , Represents the long short-term memory operation of the real and imaginary parts.

[0013] Optionally, the fully connected classification module includes a flattening layer, a first fully connected layer, a first batch of normalization layers, a second fully connected layer, a second batch of normalization layers, a third fully connected layer and a softmax layer. The flattening layer, the first fully connected layer, the first batch of normalization layers, the second fully connected layer, the second batch of normalization layers, the third fully connected layer and the softmax layer are connected in sequence. The flattening layer is used to stack the elements input therein into a one-dimensional tensor, and the fully connected layer is used to perform weighted summation of the elements input therein with the weights of the complex convolutional recurrent neural network.

[0014] Optionally, the complex convolutional recurrent neural network is trained using a cross entropy loss function, and the training includes: The acoustic signals generated by the drone during flight are input into the complex convolutional recurrent neural network to obtain the same number of outputs as the number of drone types; Calculate the cross entropy loss between the complex convolutional recurrent neural network output and the true drone type label; According to the cross entropy loss, the complex convolutional recurrent neural network is optimized through back propagation and gradient descent algorithms to obtain a trained complex convolutional recurrent neural network.

[0015] Optionally, the cross entropy loss function is:

[0016] in, represents the cross entropy loss, represents the number of training samples in a training batch, Indicates the category of the drone label, express, Indicates Training sample categories Tags, Indicates the predicted The training samples are of category probability.

[0017] The present invention also provides a drone identification device, comprising: An acquisition module is used to acquire drone sound signals; The classification module inputs the drone sound signal into the trained complex convolutional recurrent neural network to obtain the drone type recognition result; Among them, the complex convolutional recurrent neural network includes a differentiable Fourier transform module, a complex convolution module, a complex long short-term memory module and a fully connected classification module; the differentiable Fourier transform module converts the time domain waveform characteristics of the drone sound signal into the complex time-frequency spectrum characteristics of the drone sound signal; several complex convolution modules extract the complex time-frequency spectrum characteristics of the drone sound signal to obtain deep complex features; the complex long short-term memory module extracts the deep complex features to obtain the time-series related complex features of the drone sound signal; the fully connected classification module processes the time-series related complex features of the drone sound signal to obtain a vector with the same dimension as the number of drone types to represent the recognition probability of each type of drone.

[0018] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the aforementioned voiceprint-based drone identification method when executing the program.

[0019] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of the aforementioned voiceprint-based drone identification method when executed by a processor.

[0020] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: the present invention is based on a complex convolutional recurrent neural network, and utilizes the uniqueness of the air cavity sound generated by the differences in size, shape, rotation speed and other characteristics of various components of the drone, and captures and identifies the sound characteristics generated by the drone during flight through deep learning technology. The complex convolutional recurrent neural network combines the advantages of convolutional neural networks and recurrent neural networks, and can simultaneously capture local and global features and timing information when processing sound signals, and by operating on the complex time-frequency characteristics of the sound signal, it can achieve efficient recognition of the drone voiceprint, so that the type of drone can still be accurately identified in dark environments and environments with limited RF signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a training and testing flow chart of a complex convolutional recurrent neural network in a drone identification method based on voiceprint provided by an embodiment of the present invention; Figure 2 It is a structural diagram of a complex convolutional recurrent neural network in a drone identification method based on voiceprint provided by an embodiment of the present invention; Figure 3 is a structural diagram of a complex convolution module in a drone identification method based on voiceprint provided in an embodiment of the present invention; Figure 4 It is a structural diagram of a plurality of long and short-term memory modules in a drone identification method based on voiceprint provided by an embodiment of the present invention; Figure 5 3 is a comparison diagram of a drone identification method based on voiceprint and a drone identification method based on image and radio frequency signal provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0022] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention. Embodiment 1:

[0023] This embodiment provides a drone identification method based on voiceprint, including: Acquire drone acoustic signals; The drone acoustic signal is input into the trained complex convolutional recurrent neural network to obtain the drone type recognition result.

[0024] The complex convolutional recurrent neural network is trained and tested using a public dataset (Sara Al-Emadi, Abdulla Al-Ali, Amr Mohammad, andAbdulaziz Al Ali. Audio based drone detection and identification using deeplearning.Using deep learning for audio-based drone detection and identification, In 2019 15th International Wireless Communications & Mobile Computing Conference (IWCMC, International Wireless Communications and Mobile Computing Conference), 2019.) divided into training sets, validation sets, and test sets. The public dataset contains two types of drones "Bebop" and "Mambo", each type of drone contains 666 audios, and each type of drone uses 66 audios as a test dataset, 60 audios as a validation dataset, and the remaining 540 audios as a training dataset. Figure 1 As shown in the figure, the training and testing process of the complex convolutional recurrent neural network is as follows: Training phase: ① Input the audio in the training data set, i.e., the drone flight sound training data, into the complex convolutional recurrent neural network to obtain the same output as the number of drone types; ② Calculate the cross entropy loss between the complex convolutional recurrent neural network output and the real drone type label; ③ Optimize the complex convolutional recurrent neural network based on cross entropy loss and through back propagation and gradient descent algorithms; ④ Repeat steps ①-③ until the performance of the complex convolutional recurrent neural network reaches the optimal level on the validation data set, and a trained complex convolutional recurrent neural network is obtained.

[0025] Testing phase: ① Input the audio in the test data set, i.e., the drone flight sound test data, into the trained complex convolutional recurrent neural network to obtain the predicted drone type recognition result.

[0026] The specific steps are as follows: Step 1: Train a complex convolutional recurrent neural network.

[0027] like Figure 2As shown in the figure, a complex convolutional recurrent neural network consisting of a differentiable Fourier transform module, a complex convolution module, a complex long short-term memory module and a fully connected classification module is constructed; the differentiable Fourier transform module converts the time domain waveform characteristics of the drone sound signal into the complex time-frequency spectrum characteristics of the drone sound signal; several complex convolution modules extract the complex time-frequency spectrum characteristics of the drone sound signal to obtain deep complex features; the complex long short-term memory module extracts the deep complex features to obtain the time-series related complex features of the drone sound signal; the fully connected classification module processes the time-series related complex features of the drone sound signal to obtain a vector with the same dimension as the number of drone types to represent the recognition probability of each type of drone.

[0028] Among them, Figure 2 and Figure 3 As shown in Figure 1, the complex convolution module includes a complex convolution layer, a complex batch normalization layer, and a complex PReLU layer connected in sequence. The calculation formula of the complex convolution layer is:

[0029] In the formula, represents the output of the complex convolutional layer, represents the input of the complex convolutional layer, , Respectively The real and imaginary parts of represents the convolution kernel of the complex convolutional layer, , Represent the real and imaginary parts of the convolution kernel respectively, represents the complex convolution operation, Represents the convolution operation; The calculation formula of the complex batch normalization layer is:

[0030] In the formula, represents the output of the complex batch normalization layer, , Respectively The real and imaginary parts of represents the complex batch normalization operation, Represents batch normalization operation; The calculation formula of the complex PReLU layer is:

[0031] In the formula, represents the output of the complex PReLU layer, , Respectively The real and imaginary parts of represents a complex PReLU operation, express operate.

[0032] like Figure 2 and Figure 4 As shown, the complex long short-term memory module includes two long short-term memory network layers connected in sequence. The calculation formula of the complex long short-term memory module is:

[0033]

[0034] in, represents the output of the complex long short-term memory module, represents the input of the complex long short-term memory module, , Respectively The real and imaginary parts of Indicates complex long and short-term memory operations, , Represents the long short-term memory operation of the real and imaginary parts.

[0035] The fully connected classification module includes a flat layer, a first fully connected layer, a first batch of normalization layers, a second fully connected layer, a second batch of normalization layers, a third fully connected layer and a softmax layer. The flat layer, the first fully connected layer, the first batch of normalization layers, the second fully connected layer, the second batch of normalization layers, the third fully connected layer and the softmax layer are connected in sequence. The flat layer is used to stack the elements input therein into a one-dimensional tensor, and the fully connected layer is used to perform weighted summation of the elements input therein with the weights of the complex convolutional recurrent neural network.

[0036] The scale of the input complex convolutional recurrent neural network data is agreed to be [B, 1, t], and the processing process of the complex convolutional recurrent neural network is as follows: (1) For the input data [B, 1, t] on the time domain waveform, after the Fourier transform module implemented by convolution, the output scale is agreed to be [B, 2, F, T].

[0037] (2) For input data with 2 channels, after a complex convolution module with a convolution kernel size of 3, a stride of [2, 1], and padding of [1, 2], the number of channels increases to 4, and an output of scale [B, 4, F / 2, T] is obtained.

[0038] (3) For input data with 4 channels, after a complex convolution module with a convolution kernel size of 3, a stride of [2, 1], and padding of [1, 2], the number of channels increases to 8, and an output of scale [B, 8, F / 4, T] is obtained.

[0039] (4) For input data with 8 channels, after a complex convolution module with a convolution kernel size of 3, a stride of [2, 1], and a padding of [1, 2], the number of channels increases to 16, and an output of scale [B, 16, F / 8, T] is obtained.

[0040] (5) For input data with 16 channels, after a complex convolution module with a convolution kernel size of 3, a stride of [2, 1], and padding of [1, 2], the number of channels increases to 32, and an output of scale [B, 32, F / 16, T] is obtained.

[0041] (6) For input data with 32 channels, after a complex convolution module with a convolution kernel size of 3, a stride of [2, 1], and padding of [1, 2], the number of channels increases to 64, and an output of scale [B, 64, F / 32, T] is obtained.

[0042] (7) For input data with 64 channels, after passing through a complex convolution module with a convolution kernel size of 3 and a stride of [2, 1], the number of channels remains at 64, and an output of scale [B, 64, F / 64, T] is obtained.

[0043] (8) For input data with 64 channels, a scale-invariant output is obtained after two complex long short-term memory modules with 64 hidden layers.

[0044] (9) For input data of scale [B, 64, F / 64, T], after passing through the fully connected classification module, the output is a tensor with the same dimension as the number of preset drone types.

[0045] Among them, B represents the batch size, t represents the time dimension size of waveform audio data, T represents the time dimension size of frequency domain audio data, and F represents the frequency dimension size of frequency domain audio data.

[0046] (10) Calculate the cross entropy loss between the network output tensor and the true label of the data, and use the back propagation and gradient descent algorithm (the learning rate is set to 0.01, and decays by 0.5 until 0.00025 when the loss increases) to optimize the complex convolutional recurrent neural network; the expression of the cross entropy loss function is:

[0047] in, represents the cross entropy loss, represents the number of audio samples in a training batch, Indicates the category of the drone label, represents the total number of drone category labels, Indicates Audio sample categories Tags, Indicates the predicted The audio samples are of the class probability.

[0048] Step 2: Test the drone type recognition effect of the complex convolutional recurrent neural network.

[0049] The audio in the test data set, i.e., the drone flight sound test data, is input into the trained complex convolutional recurrent neural network to obtain the predicted drone type recognition result.

[0050] According to the test results: the trained complex convolutional recurrent neural network can achieve a recognition rate of 96.21% for ordinary drone voiceprint audio data, indicating that the neural network has high performance and can achieve high-confidence recognition of drone voiceprints, thereby accurately identifying the type of drone.

[0051] In summary, the drone identification method based on voiceprint described in this embodiment is based on a complex convolutional recurrent neural network, which utilizes the uniqueness of the air cavity sound generated by the differences in size, shape, rotation speed and other characteristics of various drone components, and captures and identifies the sound characteristics generated by the drone during flight through deep learning technology. The complex convolutional recurrent neural network combines the advantages of convolutional neural networks and recurrent neural networks, and can simultaneously capture local and global features and timing information when processing sound signals, and by operating on the complex time-frequency characteristics of the sound signal, it can achieve efficient recognition of drone voiceprints, such as Figure 5 As shown in the figure, compared with the recognition methods based on images and radio frequency signals, the recognition method based on voiceprint can still accurately identify the type of drone in dark environments and environments with limited radio frequency signals. Embodiment 2:

[0052] This embodiment provides a drone identification device, based on the drone identification method based on voiceprint described in Embodiment 1, including: An acquisition module is used to acquire drone sound signals; The classification module inputs the drone sound signal into the trained complex convolutional recurrent neural network to obtain the drone type recognition result; Among them, the complex convolutional recurrent neural network includes a differentiable Fourier transform module, a complex convolution module, a complex long short-term memory module and a fully connected classification module; the differentiable Fourier transform module converts the time domain waveform characteristics of the drone sound signal into the complex time-frequency spectrum characteristics of the drone sound signal; several complex convolution modules extract the complex time-frequency spectrum characteristics of the drone sound signal to obtain deep complex features; the complex long short-term memory module extracts the deep complex features to obtain the time-series related complex features of the drone sound signal; the fully connected classification module processes the time-series related complex features of the drone sound signal to obtain a vector with the same dimension as the number of drone types to represent the recognition probability of each type of drone. Embodiment three:

[0053] This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps of the drone identification method based on voiceprint described in Embodiment 1 are implemented. Embodiment 4:

[0054] This embodiment 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 identification method based on voiceprint described in Embodiment 1 are implemented.

[0055] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0056] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0057] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0058] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0059] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A drone identification method based on voiceprint, characterized in that: include: Acquire drone acoustic signals; Input the drone acoustic signal into the trained complex convolutional recurrent neural network to obtain the drone type recognition result; Among them, the complex convolutional recurrent neural network includes a differentiable Fourier transform module, a complex convolution module, a complex long short-term memory module and a fully connected classification module; the differentiable Fourier transform module converts the time domain waveform characteristics of the drone sound signal into the complex time-frequency spectrum characteristics of the drone sound signal; several complex convolution modules extract the complex time-frequency spectrum characteristics of the drone sound signal to obtain deep complex features; the complex long short-term memory module extracts the deep complex features to obtain the time-series related complex features of the drone sound signal; the fully connected classification module processes the time-series related complex features of the drone sound signal to obtain a vector with the same dimension as the number of drone types to represent the recognition probability of each type of drone.

2. The drone identification method based on voiceprint according to claim 1 is characterized in that: The complex convolution module includes a complex convolution layer, a complex batch normalization layer, and a complex PReLU layer connected in sequence. The calculation formula of the complex convolution layer is: ; In the formula, represents the output of the complex convolutional layer, represents the input of the complex convolutional layer, , Respectively The real and imaginary parts of represents the convolution kernel of the complex convolutional layer, , Represent the real and imaginary parts of the convolution kernel respectively, represents the complex convolution operation, Represents the convolution operation; The calculation formula of the complex batch normalization layer is: ; In the formula, represents the output of the complex batch normalization layer, , Respectively The real and imaginary parts of represents the complex batch normalization operation, Represents batch normalization operation; The calculation formula of the complex PReLU layer is: ; In the formula, represents the output of the complex PReLU layer, , Respectively The real and imaginary parts of represents a complex PReLU operation, express operate.

3. The drone identification method based on voiceprint according to claim 1 is characterized in that: The complex long short-term memory module includes two long short-term memory network layers connected in sequence. The calculation formula of the complex long short-term memory module is: ; ; in, represents the output of the complex long short-term memory module, represents the input of the complex long short-term memory module, , Respectively The real and imaginary parts of Indicates complex long and short-term memory operations, , Represents the long short-term memory operation of the real and imaginary parts.

4. The drone identification method based on voiceprint according to claim 1 is characterized in that: The fully connected classification module includes a flat layer, a first fully connected layer, a first batch of normalization layers, a second fully connected layer, a second batch of normalization layers, a third fully connected layer and a softmax layer. The flat layer, the first fully connected layer, the first batch of normalization layers, the second fully connected layer, the second batch of normalization layers, the third fully connected layer and the softmax layer are connected in sequence. The flat layer is used to stack the elements input therein into a one-dimensional tensor, and the fully connected layer is used to perform weighted summation of the elements input therein with the weights of the complex convolutional recurrent neural network.

5. The drone identification method based on voiceprint according to claim 1 is characterized in that: The complex convolutional recurrent neural network is trained using the cross entropy loss function. The training includes: The acoustic signals generated by the drone during flight are input into the complex convolutional recurrent neural network to obtain the same number of outputs as the number of drone types; Calculate the cross entropy loss between the complex convolutional recurrent neural network output and the true drone type label; According to the cross entropy loss and the back propagation and gradient descent algorithms, the complex convolutional recurrent neural network is optimized to obtain a trained complex convolutional recurrent neural network.

6. The drone identification method based on voiceprint according to claim 5 is characterized in that: The cross entropy loss function is: ; in, represents the cross entropy loss, represents the number of training samples in a training batch, Indicates the category of the drone label, represents the total number of drone category labels, Indicates Training sample categories Tags, The predicted The training samples are of category probability.

7. A drone identification device, characterized in that: include: An acquisition module is used to acquire drone sound signals; The classification module inputs the drone sound signal into the trained complex convolutional recurrent neural network to obtain the drone type recognition result; Among them, the complex convolutional recurrent neural network includes a differentiable Fourier transform module, a complex convolution module, a complex long short-term memory module and a fully connected classification module; the differentiable Fourier transform module converts the time domain waveform characteristics of the drone sound signal into the complex time-frequency spectrum characteristics of the drone sound signal; several complex convolution modules extract the complex time-frequency spectrum characteristics of the drone sound signal to obtain deep complex features; the complex long short-term memory module extracts the deep complex features to obtain the time-series related complex features of the drone sound signal; the fully connected classification module processes the time-series related complex features of the drone sound signal to obtain a vector with the same dimension as the number of drone types to represent the recognition probability of each type of drone.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the drone identification method based on voiceprint described in any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the drone identification method based on voiceprint described in any one of claims 1 to 6 are implemented.

Citation Information

Cited By

  • Communication modulation identification method, system and device based on complex number cyclic convolutional network, and medium

    CN119696967A

  • Unmanned aerial vehicle identification method and device based on small target detection, and electronic equipment

    CN120951004A