Unmanned aerial vehicle radio frequency signal recognition method based on pulse neural network

Abstract

The application discloses a kind of unmanned aerial vehicle radio frequency signal identification methods based on pulse neural network, unmanned aerial vehicle radio frequency signal is collected, and the unmanned aerial vehicle radio frequency signal of collection is preprocessed, and the unmanned aerial vehicle radio frequency signal after preprocessing is divided into training set and test set;Pulse neural network model is established according to basic module;The basic module includes ConvBNAct module, Fused-MBConv module and MBConv module, TCSJA module is used in the MBConv module, i.e.time-channel and space joint attention module;Training data set is input into pulse neural network model for training, and test data set is used to evaluate model performance, effectively reduce the time step required in training and inference process, while realizing high-precision unmanned aerial vehicle radio frequency signal identification task.

Description

Technical Field

[0001] This invention belongs to the field of radio frequency signal technology, specifically relating to a method for identifying radio frequency signals of unmanned aerial vehicles based on a pulse neural network. Background Technology

[0002] Artificial Neural Networks (ANNs) have made groundbreaking progress in various fields, among which Convolutional Neural Networks (CNNs) have been widely used in radio frequency (RF) signal recognition tasks due to their powerful feature extraction and generalization capabilities. Although ANN models have achieved remarkable results in UAV RF signal recognition tasks—for example, Basak et al. designed the RF-UAVNet model, which uses one-dimensional grouped convolutions and multi-level residual connections to achieve high-precision UAV recognition—traditional ANN models face a trade-off between computational power consumption and model accuracy in practical applications. This limits the application of traditional ANN models in low-power and resource-constrained environments.

[0003] In contrast, spiking neural networks (SNNs), as a new generation of neural networks, are inspired by the mechanisms of biological neurons. They use discrete pulse events, similar to the action potentials generated by neurons in the brain, to transmit information, thus having information processing capabilities closer to the brain. This also makes spiking neural networks have lower power consumption on dedicated architectures of neuromorphic architectures or neuromorphic hardware. In addition, the unique spatiotemporal characteristics of spiking neural networks enable them to better process sequential signals.

[0004] However, existing spiking neural network-based models are simple and difficult to handle UAV radio frequency signal identification tasks in complex scenarios. At the same time, how to make full use of the spatiotemporal characteristics of spiking neural networks to reduce the time step required for model inference and achieve algorithms with high performance, robustness and low power consumption is an important technical problem in this field. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for identifying radio frequency signals of unmanned aerial vehicles (UAVs) based on a spiking neural network.

[0006] The specific plan is as follows: A method for identifying radio frequency signals of unmanned aerial vehicles (UAVs) based on pulse neural networks includes the following steps: S1): Collect the drone radio frequency signal, preprocess the collected drone radio frequency signal, and divide the preprocessed drone radio frequency signal into training set and test set; S2): Establish a spiking neural network model based on the basic modules; the basic modules include the ConvBNAct module, the Fused-MBConv module, and the MBConv module. The MBConv module adopts a time-channel and spatial joint attention module, namely the TCSJA module. S3): Input the training dataset into the spiking neural network model for training, and use the test dataset to evaluate the model performance.

[0007] The method for acquiring radio frequency signals from drones includes the following steps: S11): Set up an RF signal acquisition device and at least four types of drones, and set the sampling frequency, signal bandwidth and IQ points of each sample of the RF signal acquisition device. S12): The radio frequency signal acquisition device collects radio frequency signals from each model of drone at different distances under 2.4GHz and 5.8GHz. At the same time, the radio frequency signal acquisition device collects radio frequency signals from each model of drone at different operating modes under 2.4GHz and 5.8GHz. If it is an indoor environment, it also collects WIFI signal and Bluetooth signal as interference signals. The different operating modes of drone include power-on, flight and hover. S13): Store the collected radio frequency signals in different folders according to different drone models.

[0008] The method for preprocessing the collected UAV radio frequency signals is as follows: P1): Enhance the radio frequency signal by superimposing WIFI and Bluetooth signals onto the collected radio frequency signal data of different drone models; P2): Perform a short-time Fourier transform on the enhanced RF signal data to obtain the time spectrum of the RF signal, and divide the time spectrum of the RF signal into a training set and a test set.

[0009] Building a spiking neural network model based on the basic modules includes the following steps: S21): Establish a ConvBNAct module, which includes a convolutional layer Conv(k×k), a batch normalization layer BN, and an IF activation neuron connected in sequence, where k represents the size of the convolutional kernel; S22): Establish the Fused-MBConv module based on the expansion coefficient; When the expansion coefficient = 1, the Fused-MBConv module includes a ConvBNAct module with a 3x3 kernel and a Dropout layer; When the expansion coefficient > 1, the main branch of the Fused-MBConv module is connected in sequence to the ConvBNAct module with a 3x3 kernel, the Conv(1×1) convolutional layer with a 1x1 kernel, the batch normalization layer BN, and the Dropout layer, where the Dropout layer is the Stochastic Depth. S23): Establish an MBConv module, on the main branch of which are sequentially connected a ConvBNAct module with a 1x1 kernel, a ConvBNAct module with a 3x3 kernel, a TCSJA module, a convolutional layer with a 1x1 kernel, a batch normalization layer, and a Dropout layer, wherein the Dropout layer is a Stochastic Depth. S24): Establish the architecture of a spiking neural network, which includes Stage 0 to Stage 7 connected sequentially, wherein... Stage 0 contains a ConvBNAct module with a stride of 2 and a kernel size of 3x3. Stage 0 has 3 input channels and 24 output channels. Stage 1 contains two Fused-MBConv modules with an expansion coefficient of 1 and a step size of 1. The number of input channels and output channels in Stage 1 are both 24. Stage 2 consists of two Fused-MBConv modules with an expansion factor of 4 and a step size of 2. Stage 2 has 24 input channels and 48 output channels. Stage 3 contains two Fused-MBConv modules with an expansion factor of 4 and a step size of 2. Stage 3 has 48 input channels and 64 output channels. Stage 4 contains two MBConv modules with an expansion factor of 4 and a step size of 2. Stage 4 has 64 input channels and 128 output channels. Stage 5 contains two MBConv modules with an expansion factor of 6 and a step size of 2. Stage 5 has 128 input channels and 160 output channels. Stage 6 contains two MBConv modules with an expansion factor of 6 and a step size of 2. Stage 6 has 160 input channels and 256 output channels. Stage 7 includes a 1x1 convolutional layer (Conv1×1), an adaptive global average pooling layer (Pooling), and a fully connected layer (FC). The fully connected layer in Stage 7 has 1280 output channels.

[0010] The IF activated neuron model receives input pulses from synaptic neurons, accumulates these pulses into a membrane potential, and then activates the membrane potential when it reaches a preset voltage threshold. When an IF-activated neuron outputs a pulse signal, the equation describing the change in membrane potential of the IF-activated neuron is: in, t represents the time step, and n represents the nth layer of the network. It is the input of the neuron. This represents the membrane potential on the neuron at time step t. It is a pulse firing state, through a step function. Determine whether the membrane potential at time step t exceeds a preset voltage threshold. and emit pulses, It is the membrane potential after reset.

[0011] The TCSJA module adjusts the pulse distribution in the spiking neural network model from the time dimension, spatial dimension, and channel dimension, respectively. The pulse distribution adjustment method includes the following steps: Q1): In the time dimension attention module, the input feature information is passed through a three-dimensional adaptive average pooling layer, and then through a three-dimensional convolutional layer with a kernel of 1x1x1, so as to reduce the time step of the input channel from timestep to timestep / 2. Next, a batch normalization layer is used to introduce non-linear characteristics, and then a three-dimensional convolutional layer with a 1x1x1 kernel is used to restore the output channel to the original time step, thus obtaining the attention weights in the time dimension. Q2): In the channel-dimensional attention module, the input feature information is pooled through a two-dimensional adaptive average pooling layer, and then the number of input channels is reduced from channels to channels / 16 through a two-dimensional convolutional layer with a 1x1 kernel. Next, a batch normalization layer is used with the ReLU activation function to introduce non-linearity, and then another two-dimensional convolutional layer with a 1x1 kernel is used to restore the number of output channels to the original number of channels, thus obtaining the channel-dimensional attention weights. Q3): In the spatial dimension attention module, the input feature information is passed through a two-dimensional convolutional layer with a kernel of 1x1, reducing the number of input channels from channels to channels / 16; Next, batch normalization is performed using the ReLU activation function to introduce non-linearity. Then, a dilated convolution operation is used to capture the spatial correlation of the input features; the dilated convolution has a 3x3 kernel with a dilation parameter d=4. Next, batch normalization and ReLU activation are performed again, and then a 2D convolutional layer with a 1x1 kernel is passed through to reduce the number of output channels to 1 in order to generate a spatial attention map, in which each pixel in the spatial attention map corresponds to a weighted value. Q4): The attention weights of the three dimensions are added together through a broadcast mechanism, and the final fusion result is non-linearly mapped through a sigmoid activation function.

[0012] Step S3) employs a gradient substitution-based backpropagation algorithm for training. The training and testing process using the Sigmoid function as the gradient substitution function includes the following steps: S31): Adjust the size of the time spectrum of the preprocessed signal; when loading the training set data, transform the image size and convert the image data into tensor format; when loading the validation set and test set, transform the image size of the input time spectrum and convert the image data into tensor format, and then use the normalize operation to normalize the image data. S32): The processed UAV RF signal time spectrum is input into the SNN model. A time dimension is added before inputting into the network to obtain a sequence with shape=[T, N, C, H, W], which is then fed into the network layer. The first ConvBNAct module in the SNN model performs the encoding work. In the sequence, T represents the time step required by the model, N represents the number of batches, C represents the number of time spectrum channels, H represents the height of the time spectrum, and W represents the width of the time spectrum. S33): During training, the SGD optimizer is used to update the model parameters, and the cosine annealing strategy is used to dynamically adjust the learning rate during training. The initial learning rate is set to 0.01, and the loss function is the cross-entropy loss function. S34): Set different time steps for training and testing, obtain experimental results under different time steps, and select the optimal time step based on model accuracy and computational power consumption.

[0013] This invention discloses a method for identifying radio frequency (RF) signals of unmanned aerial vehicles (UAVs) based on a spiking neural network. The spiking neural network model, based on a multidimensional attention mechanism, is used for UAV RF signal identification in complex electromagnetic environments, achieving the goal of UAV identification. Compared with existing technologies, the spiking neural network designed in this method can provide a low-power, high-precision UAV RF signal recognition algorithm. Compared with commonly used convolutional neural networks, the network structure of this method has the ability to fuse multi-dimensional feature information and stronger spatiotemporal modeling capabilities. By fully utilizing the spatiotemporal characteristics of spiking neural networks, feature information of UAV RF signals is extracted from multiple dimensions, selectively focusing on important inputs, reducing interference from co-frequency signals, improving the recognition efficiency of UAV RF signals, and effectively reducing the time step required in the training and inference processes, while achieving high-precision UAV RF signal recognition. Attached Figure Description

[0014] Figure 1 This is a structural diagram of the ConvBNAct module, the Fused-MBConv module, and the MBConv module.

[0015] Figure 2 This is a structural diagram of the TCSJA module.

[0016] Figure 3 This is the architecture table of a spiking neural network. Detailed Implementation

[0017] The technical solutions in the embodiments of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the implementation of the present invention, and not all of it. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0018] The purpose of this invention is to provide a spiking neural network model based on a multidimensional attention mechanism for UAV radio frequency signal recognition tasks, aiming to achieve high-precision recognition of UAV radio frequency signals in complex electromagnetic environments and reduce the power consumption of hardware resources required by the model.

[0019] To achieve the above objectives, this invention uses a gradient substitution algorithm to construct a spiking neural network model based on a multidimensional attention mechanism for UAV radio frequency signal identification tasks in complex electromagnetic environments.

[0020] A method for identifying radio frequency signals of unmanned aerial vehicles (UAVs) based on pulse neural networks includes the following steps: S1): Collect the drone radio frequency signal, preprocess the collected drone radio frequency signal, and divide the preprocessed drone radio frequency signal into training set and test set; The method for acquiring radio frequency signals from drones includes the following steps: S11): Set up an RF signal acquisition device and at least four types of drones, and set the sampling frequency, signal bandwidth and IQ points of each sample of the RF signal acquisition device. In this embodiment, the USRP B210 is used as the radio frequency signal acquisition device. The four drone models are DJI Avata, DJI FPV drone, DJI Mavic 2 drone, and DJI Mavic 3 drone. The sampling frequency is set to 50M Sample / s, the signal bandwidth is 50MHz, and the number of IQ points per sample is 10M. S12): The radio frequency signal acquisition device collects radio frequency signals from each model of drone at different distances under 2.4GHz and 5.8GHz. At the same time, the radio frequency signal acquisition device collects radio frequency signals from each model of drone at different operating modes under 2.4GHz and 5.8GHz. If it is an indoor environment, it also collects WIFI signal and Bluetooth signal as interference signals. The different operating modes of drone include power-on, flight and hover. S13): The collected radio frequency signals are stored in different folders according to different drone models. In this embodiment, the dataset contains image transmission signals with different bandwidths and low signal-to-noise ratio signals and co-channel interference signals, totaling 2440 drone radio frequency signals.

[0021] The method for preprocessing the collected UAV radio frequency signals is as follows: P1): Enhance the radio frequency signal by superimposing WIFI and Bluetooth signals onto the collected radio frequency signal data of different drone models; P2): Perform a short-time Fourier transform on the enhanced RF signal data to obtain the time spectrum of the RF signal, and divide the time spectrum of the RF signal into a training set and a test set; In this embodiment, when performing a short-time Fourier transform on the UAV radio frequency signal, a Hamming window is used as the window function to reduce artifacts caused by spectral leakage. The length of the Fourier transform is set to 1024 to obtain better spectral resolution, and the number of overlapping samples is set to 512 to reduce the deviation of the spectral estimation.

[0022] S2): Establish a spiking neural network model based on the basic modules; such as... Figures 1 to 2 As shown, the basic modules include the ConvBNAct module, the Fused-MBConv module, and the MBConv module. The MBConv module adopts a joint attention module of time-channel and space, namely the TCSJA module. The TCSJA module integrates attention mechanisms of time, channel, and space dimensions. The introduction of the TCSJA module can improve the robustness and accuracy of the model in different dimensions, thereby improving the overall model performance. Building a spiking neural network model based on the basic modules includes the following steps: S21): Establish a ConvBNAct module, which includes a convolutional layer Conv(k×k), a batch normalization layer BN, and an IF activation neuron connected in sequence, where k represents the size of the convolutional kernel; The IF activated neuron model receives input pulses from synaptic neurons, accumulates these pulses into a membrane potential, and then activates the membrane potential when it reaches a preset voltage threshold. When an IF-activated neuron outputs a pulse signal, the equation describing the change in membrane potential of the IF-activated neuron is: in, t represents the time step, and n represents the nth layer of the network. It is the input of the neuron. This represents the membrane potential on the neuron at time step t. It is a pulse firing state, through a step function. Determine whether the membrane potential at time step t exceeds a preset voltage threshold. and emit pulses, It is the membrane potential after reset.

[0023] S22): Establish the Fused-MBConv module based on the expansion coefficient; When the expansion coefficient = 1, the Fused-MBConv module includes a ConvBNAct module with a 3x3 kernel and a Dropout layer; When the expansion coefficient > 1, the main branch of the Fused-MBConv module is connected in sequence to the ConvBNAct module with a 3x3 kernel, the Conv(1×1) convolutional layer with a 1x1 kernel, the batch normalization layer BN, and the Dropout layer. The Dropout layer is Stochastic Depth, which can randomly discard part of the network structure during training to speed up the network training. S23): Establish an MBConv module, on the main branch of which are sequentially connected a ConvBNAct module with a 1x1 kernel, a ConvBNAct module with a 3x3 kernel, a TCSJA module, a convolutional layer with a 1x1 kernel, a batch normalization layer, and a Dropout layer. The Dropout layer is a Stochastic Depth layer, which can randomly discard part of the network structure during training to speed up the network training. The TCSJA module adjusts the pulse distribution in the spiking neural network model from the time dimension, spatial dimension, and channel dimension, respectively. The pulse distribution adjustment method includes the following steps: Q1): In the time dimension attention module, the input feature information is passed through a three-dimensional adaptive average pooling layer, and then through a three-dimensional convolutional layer with a kernel of 1x1x1, so as to reduce the time step of the input channel from timestep to timestep / 2. Next, a batch normalization layer is used to introduce non-linear characteristics, and then a three-dimensional convolutional layer with a 1x1x1 kernel is used to restore the output channel to the original time step, thus obtaining the attention weights in the time dimension. Q2): In the channel-dimensional attention module, the input feature information is pooled through a two-dimensional adaptive average pooling layer, and then the number of input channels is reduced from channels to channels / 16 through a two-dimensional convolutional layer with a 1x1 kernel. Next, a batch normalization layer is used with the ReLU activation function to introduce non-linearity, and then another two-dimensional convolutional layer with a 1x1 kernel is used to restore the number of output channels to the original number of channels, thus obtaining the channel-dimensional attention weights. Q3): In the spatial dimension attention module, the input feature information is passed through a two-dimensional convolutional layer with a 1x1 kernel, reducing the number of input channels from channels to channels / 16, thus reducing computational complexity and the number of channels; Next, batch normalization is performed, using the ReLU activation function to introduce non-linear characteristics. Then, an operation with dilated convolution is used to capture the spatial correlation of the input features. The dilated convolution has a 3x3 kernel with a dilation parameter d=4. The dilated convolution is used to increase the receptive field of the convolution operation, which helps the model better understand the spatial structure of the input features. Next, batch normalization and ReLU activation are performed again, and then a 2D convolutional layer with a 1x1 kernel is passed through to reduce the number of output channels to 1 in order to generate a spatial attention map, in which each pixel in the spatial attention map corresponds to a weighted value. Q4): The attention weights of the three dimensions are added together through a broadcast mechanism, and the final fusion result is non-linearly mapped through a sigmoid activation function.

[0024] S24): Establish the architecture of a spiking neural network, which includes Stage 0 to Stage 7 connected sequentially, wherein... Stage 0 contains a ConvBNAct module with a stride of 2 and a kernel size of 3x3. Stage 0 has 3 input channels and 24 output channels. Stage 1 contains two Fused-MBConv modules with an expansion coefficient of 1 and a step size of 1. The number of input channels and output channels in Stage 1 are both 24. Stage 2 consists of two Fused-MBConv modules with an expansion factor of 4 and a step size of 2. Stage 2 has 24 input channels and 48 output channels. Stage 3 contains two Fused-MBConv modules with an expansion factor of 4 and a step size of 2. Stage 3 has 48 input channels and 64 output channels. Stage 4 contains two MBConv modules with an expansion factor of 4 and a step size of 2. Stage 4 has 64 input channels and 128 output channels. Stage 5 contains two MBConv modules with an expansion factor of 6 and a step size of 2. Stage 5 has 128 input channels and 160 output channels. Stage 6 contains two MBConv modules with an expansion factor of 6 and a step size of 2. Stage 6 has 160 input channels and 256 output channels. Stage 7 includes a 1x1 convolutional layer (Conv1×1), an adaptive global average pooling layer (Pooling), and a fully connected layer (FC). The fully connected layer in Stage 7 has 1280 output channels.

[0025] S3): Input the training dataset into the spiking neural network model for training, and use the test dataset to evaluate the model performance.

[0026] Because the spiking neuron model is non-differentiable, a gradient-substitution-based backpropagation algorithm is used for training in step S3). The training and testing process using the Sigmoid function as the gradient substitution function includes the following steps: S31): Adjust the size of the time spectrum of the preprocessed signal; when loading the training set data, transform the image size and convert the image data into tensor format; when loading the validation set and test set, transform the image size of the input time spectrum and convert the image data into tensor format, and then use the normalize operation to normalize the image data. Image resizing is performed using `transforms.Resize((224,224))` to adjust the image size to 244x244. Image data is converted to tensor format using `transforms.ToTensor()`. S32): The processed UAV radio frequency signal time spectrum is input into the SNN model. A time dimension is added before inputting into the network to obtain a sequence with shape=[T, N, C, H, W], which is then fed into the network layer and encoded by the first ConvBNAct module in the SNN model. The shape=[T, N, C, H, W] is mainly used to display the shape of the input data. T represents the time step required by the model, N represents the number of batches, C represents the number of channels in the time spectrogram, H represents the height of the time spectrogram, and W represents the width of the time spectrogram. S33): During training, the SGD optimizer is used to update the model parameters, and the cosine annealing strategy is used to dynamically adjust the learning rate during training. The initial learning rate is set to 0.01, and the loss function is the cross-entropy loss function. S34): Set different time steps for training and testing, obtain experimental results under different time steps, and select the optimal time step based on model accuracy and computational power consumption. In this embodiment, the selected time step T is distributed as 2, 4, 6, 8, and 10.

[0027] This invention proposes a method for UAV radio frequency signal recognition based on a spiking neural network. The spiking neural network model, based on a multidimensional attention mechanism, is used for UAV radio frequency signal recognition in complex electromagnetic environments, achieving the goal of UAV identification. The model was trained and validated on public datasets (DroneDetectv2) and self-built datasets. The optimal time step was set to T=6. The UAV recognition accuracy reached 96.02% on the DroneDetectv2 dataset and 99.13% on the self-built dataset. Its computational power consumption is approximately 20.1% of that of traditional artificial neural network models with the same structure, greatly reducing the limitations of deployment and application in energy-constrained scenarios.

[0028] The technical means disclosed in this invention are not limited to those disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications are also considered within the scope of protection of this invention.

Claims

1. A method for identifying radio frequency signals of unmanned aerial vehicles based on a spiking neural network, characterized in that: Includes the following steps, S1): Collect the drone radio frequency signal, preprocess the collected drone radio frequency signal, and divide the preprocessed drone radio frequency signal into training set and test set; S2): Establish a spiking neural network model based on the basic modules; the basic modules include the ConvBNAct module, the Fused-MBConv module, and the MBConv module. The MBConv module adopts a time-channel and spatial joint attention module, namely the TCSJA module. S3): Input the training dataset into the spiking neural network model for training, and use the test dataset to evaluate the model performance.

2. The method for identifying UAV radio frequency signals based on a spiking neural network according to claim 1, characterized in that: The method for acquiring radio frequency signals from drones includes the following steps: S11): Set up an RF signal acquisition device and at least four types of drones, and set the sampling frequency, signal bandwidth and IQ points of each sample of the RF signal acquisition device. S12): The radio frequency signal acquisition device collects radio frequency signals from each model of drone at different distances under 2.4GHz and 5.8GHz. At the same time, the radio frequency signal acquisition device collects radio frequency signals from each model of drone at different operating modes under 2.4GHz and 5.8GHz. If it is an indoor environment, it also collects WIFI signal and Bluetooth signal as interference signals. The different operating modes of drone include power-on, flight and hover. S13): Store the collected radio frequency signals in different folders according to different drone models.

3. The UAV radio frequency signal identification method based on a spiking neural network according to claim 2, characterized in that: The method for preprocessing the collected UAV radio frequency signals is as follows: P1): Enhance the radio frequency signal by superimposing WIFI and Bluetooth signals onto the collected radio frequency signal data of different drone models; P2): Perform a short-time Fourier transform on the enhanced RF signal data to obtain the time spectrum of the RF signal, and divide the time spectrum of the RF signal into a training set and a test set.

4. The method for identifying UAV radio frequency signals based on a spiking neural network according to claim 1, characterized in that: Building a spiking neural network model based on the basic modules includes the following steps: S21): Establish a ConvBNAct module, which includes a convolutional layer Conv(k×k), a batch normalization layer BN, and an IF activation neuron connected in sequence, where k represents the size of the convolutional kernel; S22): Establish the Fused-MBConv module based on the expansion coefficient; When the expansion coefficient = 1, the Fused-MBConv module includes a ConvBNAct module with a 3x3 kernel and a Dropout layer; When the expansion coefficient > 1, the main branch of the Fused-MBConv module is connected in sequence to the ConvBNAct module with a 3x3 kernel, the Conv(1×1) convolutional layer with a 1x1 kernel, the batch normalization layer BN, and the Dropout layer, where the Dropout layer is the Stochastic Depth. S23): Establish an MBConv module, on the main branch of which are sequentially connected a ConvBNAct module with a 1x1 kernel, a ConvBNAct module with a 3x3 kernel, a TCSJA module, a convolutional layer with a 1x1 kernel, a batch normalization layer and a Dropout layer, wherein the Dropout layer is a Stochastic Depth. S24): Establish the architecture of a spiking neural network, which includes Stage 0 to Stage 7 connected sequentially, wherein... Stage 0 contains a ConvBNAct module with a stride of 2 and a kernel size of 3x3. Stage 0 has 3 input channels and 24 output channels. Stage 1 contains two Fused-MBConv modules with an expansion coefficient of 1 and a step size of 1. The number of input channels and output channels in Stage 1 are both 24. Stage 2 consists of two Fused-MBConv modules with an expansion factor of 4 and a step size of 2. Stage 2 has 24 input channels and 48 output channels. Stage 3 contains two Fused-MBConv modules with an expansion factor of 4 and a step size of 2. Stage 3 has 48 input channels and 64 output channels. Stage 4 contains two MBConv modules with an expansion factor of 4 and a step size of 2. Stage 4 has 64 input channels and 128 output channels. Stage 5 contains two MBConv modules with an expansion factor of 6 and a step size of 2. Stage 5 has 128 input channels and 160 output channels. Stage 6 contains two MBConv modules with an expansion factor of 6 and a step size of 2. Stage 6 has 160 input channels and 256 output channels. Stage 7 includes a 1x1 convolutional layer (Conv1×1), an adaptive global average pooling layer (Pooling), and a fully connected layer (FC). The fully connected layer in Stage 7 has 1280 output channels.

5. The UAV radio frequency signal identification method based on a spiking neural network according to claim 4, characterized in that: The IF activated neuron model receives input pulses from synaptic neurons, accumulates these pulses into a membrane potential, and then activates the membrane potential when it reaches a preset voltage threshold. When an IF-activated neuron outputs a pulse signal, the equation describing the change in membrane potential of the IF-activated neuron is: in, t represents the time step, and n represents the nth layer of the network. It is the input of the neuron. This represents the membrane potential on the neuron at time step t. It is a pulse firing state, through a step function. Determine whether the membrane potential at time step t exceeds a preset voltage threshold. and emit pulses, It is the membrane potential after reset.

6. The UAV radio frequency signal identification method based on a spiking neural network according to claim 4, characterized in that: The TCSJA module adjusts the pulse distribution in the spiking neural network model from the time dimension, spatial dimension, and channel dimension, respectively. The pulse distribution adjustment method includes the following steps: Q1): In the time dimension attention module, the input feature information is passed through a three-dimensional adaptive average pooling layer, and then through a three-dimensional convolutional layer with a kernel of 1x1x1, so as to reduce the time step of the input channel from timestep to timestep / 2. Next, a batch normalization layer is used to introduce non-linear characteristics, and then a three-dimensional convolutional layer with a 1x1x1 kernel is used to restore the output channel to the original time step, thus obtaining the attention weights in the time dimension. Q2): In the channel-dimensional attention module, the input feature information is pooled through a two-dimensional adaptive average pooling layer, and then the number of input channels is reduced from channels to channels / 16 through a two-dimensional convolutional layer with a 1x1 kernel. Next, a batch normalization layer is used with the ReLU activation function to introduce non-linearity, and then another two-dimensional convolutional layer with a 1x1 kernel is used to restore the number of output channels to the original number of channels, thus obtaining the channel-dimensional attention weights. Q3): In the spatial dimension attention module, the input feature information is passed through a two-dimensional convolutional layer with a kernel of 1x1, reducing the number of input channels from channels to channels / 16; Next, batch normalization is performed using the ReLU activation function to introduce non-linearity. Then, a dilated convolution operation is used to capture the spatial correlation of the input features; the dilated convolution has a 3x3 kernel with a dilation parameter d=4. Next, batch normalization and ReLU activation are performed again, and then a 2D convolutional layer with a 1x1 kernel is passed through to reduce the number of output channels to 1 in order to generate a spatial attention map, in which each pixel in the spatial attention map corresponds to a weighted value. Q4): The attention weights of the three dimensions are added together through a broadcast mechanism, and the final fusion result is non-linearly mapped through a sigmoid activation function.

7. The method for identifying UAV radio frequency signals based on a spiking neural network according to claim 1, characterized in that: Step S3) employs a gradient substitution-based backpropagation algorithm for training. The training and testing process using the Sigmoid function as the gradient substitution function includes the following steps: S31): Adjust the size of the time spectrum of the preprocessed signal; when loading the training set data, transform the image size and convert the image data into tensor format; when loading the validation set and test set, transform the image size of the input time spectrum and convert the image data into tensor format, and then use the normalize operation to normalize the image data. S32): The processed UAV RF signal time spectrum is input into the SNN model. A time dimension is added before inputting into the network to obtain a sequence with shape=[T, N, C, H, W], which is then fed into the network layer. The first ConvBNAct module in the SNN model performs the encoding work. In the sequence, T represents the time step required by the model, N represents the number of batches, C represents the number of time spectrum channels, H represents the height of the time spectrum, and W represents the width of the time spectrum. S33): During training, the SGD optimizer is used to update the model parameters, and the cosine annealing strategy is used to dynamically adjust the learning rate during training. The initial learning rate is set to 0.01, and the loss function is the cross-entropy loss function. S34): Set different time steps for training and testing, obtain experimental results under different time steps, and select the optimal time step based on model accuracy and computational power consumption.

Images