A method for generating unmanned aerial vehicle management and control signals based on a convolutional neural network

Abstract

The application discloses a kind of unmanned aerial vehicle management and control signal generation methods based on convolutional neural network, belong to unmanned aerial vehicle communication technical field, including the following steps: S1: unmanned aerial vehicle communication signal is collected, and unmanned aerial vehicle communication signal dataset is constructed;S2: unmanned aerial vehicle management and control signal generation network model is constructed, including convolution network, feature extraction network with residual module is added and spectrum signal adjustment network;S3: with the minimum generated control unmanned aerial vehicle communication signal and real unmanned aerial vehicle communication signal as optimization goal, unmanned aerial vehicle management and control signal generation network model constructed is trained by unmanned aerial vehicle communication signal dataset, until convergence;S4: unmanned aerial vehicle management and control signal is generated by the unmanned aerial vehicle management and control signal generation network model of convergence;The application can learn the time-frequency domain mode of unmanned aerial vehicle according to historical unmanned aerial vehicle communication signal, resist the complex and unknown modulation mode of unmanned aerial vehicle.

Description

Technical Field

[0001] This invention belongs to the field of unmanned aerial vehicle (UAV) communication technology, specifically relating to a method for generating UAV control signals based on convolutional neural networks. Background Technology

[0002] In recent years, drones have been widely used in both civilian and military fields due to their low cost, high mobility, and flexible deployment. However, with the increasing prevalence of drones, both civilian and military sectors are also facing security threats. To address these threats, drone communication control has become an effective countermeasure.

[0003] Current drone communication control methods generally involve transmitting broadband noise or single-frequency signals to cover the target frequency band, forcibly blocking the communication link between the drone and the remote controller or base station, or misleading the drone's flight by forging navigation signals. This method relies on pre-set control parameters, but these pre-set control parameters have limitations and are difficult to match the complex modulation technology of modern drones. At the same time, they cannot train models using historical data to identify the time-frequency domain patterns of signals and are completely defenseless against unknown modulation methods.

[0004] Therefore, a method for generating UAV control signals based on convolutional neural networks is designed to solve the above problems. Summary of the Invention

[0005] To address the problems mentioned in the background section, this invention provides a method for generating UAV control signals based on convolutional neural networks. This method is characterized by its ability to learn the time-frequency domain patterns of UAVs from historical UAV communication signals and its resistance to complex and unknown modulation methods of UAVs.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for generating unmanned aerial vehicle (UAV) control signals based on convolutional neural networks, comprising the following steps:

[0007] S1: Collect UAV communication signals and construct a UAV communication signal dataset;

[0008] S2: Construct a network model for generating UAV control signals, including a complex convolutional network, a feature extraction network with residual modules, and a spectrum signal adjustment network;

[0009] Complex convolutional networks are used to extract the phase and amplitude features of UAV communication signals;

[0010] The feature extraction network with added residual modules extracts spectral features from the phase and amplitude features extracted by the complex convolutional network;

[0011] The spectrum signal adjustment network generates a mask matrix based on the spectrum features extracted by the feature extraction network with the residual module, and continuously adjusts the spectrum signal components.

[0012] S3: Using a dataset of drone communication signals, train the constructed drone control signal generation network model with the optimization objective of minimizing the difference between the generated control drone communication signals and the real drone communication signals until convergence;

[0013] S4: Generate drone control signals through a converged drone control signal generation network model.

[0014] Furthermore, in step S2, the specific steps for the complex convolutional network to extract the phase and amplitude features of the UAV communication signal include:

[0015] S2-1: Perform a short-time Fourier transform on the UAV communication signal to convert the time-domain signal into a complex time spectrum;

[0016] S2-2: Construct a complex convolutional network, including complex convolutional layers, ComplexPReLU activation function layers, and complex pooling layers;

[0017] S2-3: Input the complex time spectrum into the complex convolutional layer for convolution to extract phase and amplitude features;

[0018] S2-4: The phase and amplitude features processed by the complex convolutional layer are input into the ComplexPReLU activation function layer. The ComplexPReLU activation function adjusts the nonlinear mapping relationship between the real and imaginary parts of the phase and amplitude features through the learned slope parameter.

[0019] S2-5: Input the phase and amplitude features processed by the ComplexPReLU activation function into the complex pooling layer, apply max pooling operation to downsample the phase and amplitude features, and output the processed phase and amplitude features.

[0020] Furthermore, in step S2, the specific steps for the feature extraction network with added residual modules to extract spectral features from the phase and amplitude features extracted by the complex convolutional network include:

[0021] S2-6: Construct a feature extraction network with residual modules, including a normalization layer, a compound activation function, a compound convolutional layer, and a residual connection layer;

[0022] S2-7: The output phase and amplitude characteristics are input into the normalization processing layer for normalization processing;

[0023] S2-8: Input the normalized phase and amplitude features into the composite activation function layer to generate phase and amplitude features after nonlinear transformation;

[0024] S2-9: Input the phase and amplitude features processed by the composite activation function into the composite convolutional layer to extract the phase and amplitude features;

[0025] S2-10: Input the phase and amplitude features processed by the composite convolutional layer into the residual connection layer. Add the phase and amplitude features processed by the composite convolutional layer to the input phase and amplitude features to output the spectral features.

[0026] Further, in step S2, the specific steps of the spectrum signal adjustment network generating a mask matrix based on the spectrum features extracted by the feature extraction network with the residual module added, and continuously adjusting the spectrum signal components, include:

[0027] S2-11: Construct a spectrum signal adjustment network, including a feature coding network layer, a mask generation layer, and a decoding layer;

[0028] S2-12: The output spectral features are input into the feature encoding network layer to extract information from the spectral features;

[0029] S2-13: The extracted spectral feature information is input into the mask generation layer, multiplied with the corresponding weight matrix, and then the corresponding bias is added. The result is then input into the composite Sigmoid activation function to generate the corresponding mask matrix. The mask value is constrained to [0,1].

[0030] S2-14: Input the generated mask matrix into the decoding layer to convert it into the corresponding control signal.

[0031] Compared with the prior art, the beneficial effects of the present invention are:

[0032] 1. This invention constructs a UAV control signal generation network model, trains the constructed UAV control signal generation network model with UAV communication signals until convergence, and generates UAV control signals through the converged UAV control signal generation network model to control the UAV. Compared with the prior art, it can learn the time-frequency domain mode of the UAV based on historical UAV communication signals and resist the complex and unknown modulation methods of the UAV.

[0033] 2. The UAV control signal generation network model constructed in this invention includes a complex convolutional network, a feature extraction network with residual modules, and a spectrum signal adjustment network. The complex convolutional network includes complex convolutional layers, ComplexPReLU activation function layers, and complex pooling layers, which can completely preserve the phase and amplitude features of the communication signal. The feature extraction network with residual modules includes normalization processing layers, composite activation functions, composite convolutional layers, and residual connection layers, which can extract the spectrum features of the communication signal. The spectrum signal adjustment network includes feature encoding network layers, mask generation layers, and decoding layers, which can achieve precise and continuous control of spectrum components, has strong generalization ability for multi-modulation signals, and the generated control signal is highly similar to the original signal, significantly improving control efficiency and adaptability. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the implementation environment of the present invention;

[0035] Figure 2 This is a flowchart of the overall method of the present invention;

[0036] Figure 3 This is a flowchart of the complex convolutional network method of the present invention;

[0037] Figure 4 The flowchart of the feature extraction network method with residual module added to this invention is shown below;

[0038] Figure 5 This is a flowchart of the spectrum signal adjustment network method of the present invention;

[0039] Figure 6 This is a comparison diagram of the control signal generated on four different modulation signals and the original signal generated by the present invention.

[0040] Figure 7 This is a distribution diagram of the spectral signal energy of the control signal generated by the present invention on four different modulation signals. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] See appendix Figure 1 The diagram illustrates the implementation environment, which includes at least one signal transmitter and one signal receiver. The signal transmitter refers to a ground base station, other drones, or other peripheral equipment. The network module refers to the drone control signal generation network based on convolutional neural networks proposed in this invention. The signal receiver refers to the drone equipment that receives communication signals.

[0043] See appendix Figure 2 The present invention relates to a method for generating UAV control signals based on a convolutional neural network, as illustrated in the schematic diagram of the implementation environment. The method includes the following steps:

[0044] S1: Collect UAV communication signals and construct a UAV communication signal dataset;

[0045] S2: Construct a network model for generating UAV control signals, including a complex convolutional network, a feature extraction network with residual modules, and a spectrum signal adjustment network;

[0046] Complex convolutional networks are used to extract the phase and amplitude features of UAV communication signals;

[0047] The feature extraction network with added residual modules extracts spectral features from the phase and amplitude features extracted by the complex convolutional network;

[0048] The spectrum signal adjustment network generates a mask matrix based on the spectrum features extracted by the feature extraction network with the residual module, and continuously adjusts the spectrum signal components.

[0049] S3: A drone control signal generation network model is built by training a drone communication signal dataset until it converges;

[0050] S4: Generate drone control signals through a converged drone control signal generation network model.

[0051] See appendix Figure 3 In step S2, the specific steps for the complex convolutional network to extract the phase and amplitude features of the UAV communication signal include:

[0052] S2-1: Perform a short-time Fourier transform on the UAV communication signal to convert the time-domain signal into a complex time spectrum;

[0053] S2-2: Construct a complex convolutional network, including complex convolutional layers, ComplexPReLU activation function layers, and complex pooling layers;

[0054] S2-3: Input the complex time spectrum into the complex convolutional layer for convolution to extract phase and amplitude features;

[0055] S2-4: The phase and amplitude features processed by the complex convolutional layer are input into the ComplexPReLU activation function layer. The ComplexPReLU activation function learns the slope parameters. and The nonlinear mapping relationship between the real and imaginary parts of the phase and amplitude characteristics is adjusted.

[0056] S2-5: Input the phase and amplitude features processed by the ComplexPReLU activation function into the complex pooling layer, apply max pooling operation to downsample the phase and amplitude features, and output the processed phase and amplitude features.

[0057] The phase and amplitude features of UAV communication signals extracted by complex convolutional networks can be expressed as:

[0058] ;

[0059] In the formula: g(·) represents the max pooling operation; W k The weight matrix of the k-th convolutional kernel is represented by X; X represents the input signal; b kThis represents the complex bias of the k-th convolution kernel.

[0060] See appendix Figure 4 In step S2, the specific steps for the feature extraction network with added residual modules to extract spectral features from the phase and amplitude features extracted by the complex convolutional network include:

[0061] S2-6: Construct a feature extraction network with residual modules, including a normalization layer, a compound activation function, a compound convolutional layer, and a residual connection layer;

[0062] S2-7: The output phase and amplitude characteristics are input into the normalization processing layer for normalization processing;

[0063] S2-8: Input the normalized phase and amplitude features into the composite activation function layer to generate phase and amplitude features after nonlinear transformation;

[0064] S2-9: Input the phase and amplitude features processed by the composite activation function into the composite convolutional layer to extract the phase and amplitude features;

[0065] S2-10: Input the phase and amplitude features processed by the composite convolutional layer into the residual connection layer. Add the phase and amplitude features processed by the composite convolutional layer to the input phase and amplitude features to output the spectral features.

[0066] The final output expression is:

[0067] ;

[0068] In the formula: This indicates the phase and amplitude characteristics of the input; This indicates a batch normalization operation; Represents a nonlinear activation function; This indicates a convolution operation.

[0069] See appendix Figure 5 In step S2, the spectrum signal adjustment network generates a mask matrix based on the spectrum features extracted by the feature extraction network with the residual module, and the specific steps for continuously adjusting the spectrum signal components include:

[0070] S2-11: Construct a spectrum signal adjustment network, including a feature coding network layer, a mask generation layer, and a decoding layer;

[0071] S2-12: The output spectral features are input into the feature encoding network layer to extract information from the spectral features;

[0072] S2-13: Input the extracted spectral feature information into the mask generation layer, along with the corresponding weight matrix W. mask After multiplication, add the corresponding bias b. maskThe input is fed into the composite Sigmoid activation function to generate its corresponding mask matrix, with the mask values ​​constrained to [0,1].

[0073] The expression for the generated mask matrix is:

[0074] ;

[0075] In the formula: H represents the information in the spectral features; W mask b represents the weight matrix; mask Indicates the bias term;

[0076] S2-14: Input the generated mask matrix into the decoding layer to convert it into the corresponding control signal;

[0077] The expression for the converted signal is:

[0078] ;

[0079] In the formula: Represents the mask matrix; Indicates the initial spectral characteristics; This indicates the generated control signal.

[0080] Specifically, in step S3, the training objective loss function of the constructed UAV control signal generation network model is:

[0081] ;

[0082] In the formula: This represents the expected generation of the i-th control signal; This represents the i-th control signal generated; n represents the number of control signals.

[0083] A simulated UAV communication signal dataset was built using simulation software including Python and the deep learning framework PyTorch. This dataset includes UAV signals with four different modulation schemes: BPSK, QPSK, 16QAM, and OFDM. Control signals were generated for the UAV signals with these four modulation schemes. The results are shown in the appendix. Figure 6 and 7 As shown:

[0084] See appendix Figure 6 The diagram shows a comparison between the control signal generated on four different modulation signals and the original signal. It can be seen that the generated control signal exhibits smooth character characteristics and is very similar to the original signal.

[0085] See appendix Figure 7The diagram shows the frequency domain distribution of the energy of the control signal generated on four different modulation signals according to the present invention. It can be seen that the spectral energy distribution of the generated control signal is consistent with the spectral energy distribution of the original signal, indicating that the waveform of the generated control signal is very similar to the waveform of the original signal.

[0086] Simultaneously, experimental tests were conducted and compared with ConvNet, CompletesConvNet, GAN, and RNN models;

[0087] The bit error rate was used as a comparison standard. The bit error rate refers to the ratio of errors received by the receiving end to the total number of bits transmitted during communication. The higher the bit error rate, the more successful the control. The results are shown in Table 1.

[0088]

[0089] As can be seen from Table 1, the bit error rate at the receiver caused by the control signal generated by the present invention on four different signal modulation schemes is higher than that of the existing model, which demonstrates its effectiveness.

[0090] The final loss was used as the comparison standard. The smaller the final loss value, the better the quality of the control signal generated by the model. The results are shown in Table 2.

[0091]

[0092] As can be seen from Table 2, the final training loss values ​​of the present invention trained on four different signal modulation schemes are all smaller than those of existing models, indicating its effectiveness.

[0093] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for generating unmanned aerial vehicle (UAV) control signals based on convolutional neural networks, characterized in that, Includes the following steps: S1: Collect UAV communication signals and construct a UAV communication signal dataset; S2: Construct a network model for generating UAV control signals, including a complex convolutional network, a feature extraction network with residual modules, and a spectrum signal adjustment network; Complex convolutional networks are used to extract the phase and amplitude features of UAV communication signals; The feature extraction network with added residual modules extracts spectral features from the phase and amplitude features extracted by the complex convolutional network; The spectrum signal adjustment network generates a mask matrix based on the spectrum features extracted by the feature extraction network with the residual module, and continuously adjusts the spectrum signal components. S3: Using a dataset of drone communication signals, train the constructed drone control signal generation network model with the optimization objective of minimizing the difference between the generated control drone communication signals and the real drone communication signals until convergence; S4: Generate drone control signals through a converged drone control signal generation network model.

2. The method for generating UAV control signals based on convolutional neural networks according to claim 1, characterized in that: In step S2, the specific steps for the complex convolutional network to extract the phase and amplitude features of the UAV communication signal include: S2-1: Perform a short-time Fourier transform on the UAV communication signal to convert the time-domain signal into a complex time spectrum; S2-2: Construct a complex convolutional network, including complex convolutional layers, ComplexPReLU activation function layers, and complex pooling layers; S2-3: Input the complex time spectrum into the complex convolutional layer for convolution to extract phase and amplitude features; S2-4: The phase and amplitude features processed by the complex convolutional layer are input into the ComplexPReLU activation function layer. The ComplexPReLU activation function adjusts the nonlinear mapping relationship between the real and imaginary parts of the phase and amplitude features through the learned slope parameter. S2-5: Input the phase and amplitude features processed by the ComplexPReLU activation function into the complex pooling layer, apply max pooling operation to downsample the phase and amplitude features, and output the processed phase and amplitude features.

3. The method for generating UAV control signals based on convolutional neural networks according to claim 2, characterized in that: In step S2, the specific steps for the feature extraction network with added residual modules to extract spectral features from the phase and amplitude features extracted by the complex convolutional network include: S2-6: Construct a feature extraction network with residual modules, including a normalization layer, a compound activation function, a compound convolutional layer, and a residual connection layer; S2-7: The output phase and amplitude characteristics are input into the normalization processing layer for normalization processing; S2-8: Input the normalized phase and amplitude features into the composite activation function layer to generate phase and amplitude features after nonlinear transformation; S2-9: Input the phase and amplitude features processed by the composite activation function into the composite convolutional layer to extract the phase and amplitude features; S2-10: Input the phase and amplitude features processed by the composite convolutional layer into the residual connection layer. Add the phase and amplitude features processed by the composite convolutional layer to the input phase and amplitude features to output the spectral features.

4. The method for generating UAV control signals based on convolutional neural networks according to claim 3, characterized in that: In step S2, the spectral signal adjustment network generates a mask matrix based on the spectral features extracted by the feature extraction network with the residual module, and the specific steps for continuously adjusting the spectral signal components include: S2-11: Construct a spectrum signal adjustment network, including a feature coding network layer, a mask generation layer, and a decoding layer; S2-12: The output spectral features are input into the feature encoding network layer to extract information from the spectral features; S2-13: The extracted spectral feature information is input into the mask generation layer, multiplied with the corresponding weight matrix, and then the corresponding bias is added. The result is then input into the composite Sigmoid activation function to generate the corresponding mask matrix. The mask value is constrained to [0,1]. S2-14: Input the generated mask matrix into the decoding layer to convert it into the corresponding control signal.

Images