Multi-unmanned aerial vehicle radio frequency signal detection method based on artificial intelligence

Abstract

The invention relates to the technical field of unmanned aerial vehicle radio frequency signal detection, in particular to a multi-unmanned aerial vehicle radio frequency signal detection method based on artificial intelligence, and the method comprises the steps: S1, collecting the radio frequency time domain data of an unmanned aerial vehicle through a radio frequency signal end; s2, converting the time domain data into time frequency data, compressing the time frequency data, and transmitting the compressed data to a computing device; and S3, in the computing device, performing visualization processing on the compressed data, converting the data into a grey-scale map, inputting the grey-scale map into a radio frequency signal detection neural network deployed on the computing device for detection, and outputting a final detection result, thereby identifying a corresponding unmanned aerial vehicle type. The multi-unmanned-aerial-vehicle detection and identification of the radio-frequency signals are realized, and the radio-frequency signal data of the sensor end are compressed, so that the transmission efficiency from the sensor end to the neural network calculation end is greatly improved, and the number of radio-frequency signal samples which can be processed by the neural network in unit time is increased.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle (UAV) radio frequency signal detection, and more specifically, to a multi-UAV radio frequency signal detection method based on artificial intelligence. Background Art

[0002] With the rapid development of drone technology, drones are increasingly used in military and civilian fields, such as aerial photography, logistics, agriculture, etc. However, drones may be used for illegal photography, stealing confidential information or conducting other malicious activities. In addition, drones may also pose a threat to aviation safety, such as collisions with civil aircraft. Therefore, it is particularly important to effectively manage and control drones. At present, the basic detection methods for drones can be divided into visual detection, sound detection, radar detection and passive detection based on target RF signals. Visual detection is easily affected by obstructions and is sensitive to light intensity; sound detection is not suitable for use in high-noise environments; radar detection is difficult to detect low-altitude and low-speed targets; RF signal detection is easily interfered by the electromagnetic environment. Considering the advantages and disadvantages of each detection sensor, drone detection and identification based on RF signals is considered to be one of the more ideal detection methods. Drone RF signal identification detection is mainly based on the characteristics of RF signals. Drones emit RF signals during flight, and these RF signals are related to the communication module of the drone. Usually, the RF signals of different types of drones have different characteristics. By analyzing the RF signals, it is possible to effectively identify whether there are drones in the current environment and the type of drones.

[0003] However, in the prior art, artificial intelligence UAV RF signal detection methods still remain at the stage of identifying a single RF signal sample, and a single sample intelligently outputs a single classification result, and is unable to detect multiple UAV targets in a single RF signal sample. At the same time, the existing scheme requires a large amount of data to be transmitted from the sensor end to the neural network computing end, and the number of RF signal samples that can be processed per unit time is relatively low. Therefore, in order to solve the above problems, the present invention aims to provide a multi-UAV RF signal detection method based on artificial intelligence, which can realize the detection and identification of multiple UAVs for RF signals, and at the same time compress the RF signal data on the sensor end, which greatly improves the transmission efficiency from the sensor end to the neural network computing end, and increases the number of RF signal samples that can be processed by the neural network per unit time. Summary of the invention

[0004] The present invention aims to overcome at least one defect (shortcoming) of the above-mentioned prior art and provide a multi-UAV RF signal detection method based on artificial intelligence, which is used to solve the problems in the prior art that multiple UAV targets cannot be detected in a single RF signal sample and the number of RF signal samples that can be processed per unit time is low.

[0005] The technical solution adopted by the present invention is a multi-UAV radio frequency signal detection method based on artificial intelligence, the method comprising:

[0006] S1: Collect the RF time domain data of the drone through the RF signal terminal;

[0007] S2: converting the time domain data into time-frequency data and compressing the data, and then transmitting the compressed data to a computing device;

[0008] S3: In the computing device, the compressed data is visualized and converted into a grayscale image, and then the grayscale image is input into a radio frequency signal detection neural network deployed on the computing device for detection, and a final detection result is output, so as to identify the corresponding drone type.

[0009] In this application, the problem of drone RF signal detection is directly transformed into an image target detection problem, the time-frequency matrix data is visualized as a grayscale image, and the drone signal in the time-frequency image is detected through image processing, thereby achieving the purpose of multi-drone intelligent target detection. The RF signal data is also lightweight, which greatly reduces the dimension of the image data, improves the transmission efficiency between the RF signal receiving end and the neural network inference calculation unit, reduces the computational pressure of the neural network image processing, and can effectively increase the number of detections of the entire recognition and detection system per unit time.

[0010] Preferably, step S2 includes:

[0011] S21: Continuously collect several time domain data points for the UAV RF signal;

[0012] S22: performing short-time Fourier transform on the collected time domain data, transferring the center frequency component to the center of the array, and then removing the frequency component, thereby converting the time domain data into a complex matrix of time-frequency information;

[0013] S23: Perform dimensionality reduction compression on the complex matrix of time-frequency information along the frequency axis, so as to obtain compressed time-frequency graph samples.

[0014] Further preferably, in step S23, the dimension reduction compression formula is:

[0015]

[0016] in, Represents complex matrix data; f avg represents the compression operator; k represents the first-dimension subscript of the complex matrix before compression; i represents the first-dimension subscript of the complex matrix after compression; j represents the second-dimension subscript of the complex matrix after compression.

[0017] In the above steps, by using short-time Fourier transform to convert time domain data into time-frequency data and then compressing it, the transmission efficiency from the sensor end to the neural network computing end can be greatly improved, and the number of RF signal samples that the neural network can process per unit time can be increased.

[0018] Preferably, in step S3, the visualization processing includes: normalizing the time-frequency diagram samples obtained after compression and converting them into an int8 integer matrix, and then directly storing the obtained integer matrix as a grayscale image to obtain a visualized image.

[0019] Further preferably, the calculation formula for converting the integer matrix into int8 is:

[0020]

[0021] Among them, x gray (i, j) represents an integer matrix; Represents complex matrix data; Represents the absolute value of the complex matrix element to be maximized.

[0022] In this application, by converting the compressed time-frequency graph data into an integer matrix, the time-frequency data can be converted into a grayscale image through the integer matrix, thereby converting the drone RF signal detection problem into an image target detection problem. The drone model in the time-frequency graph can be detected by image processing, thereby achieving the purpose of multi-drone intelligent target detection.

[0023] Preferably, the method further comprises building and training a radio frequency signal detection neural network, the steps of which are:

[0024] S41: using the methods of steps S2 and S3 to obtain several grayscale image samples of RF signals of different types of drones, and then labeling all samples. If there is an image transmission signal, the image transmission signal of the drone is labeled; if not, the control signal of the drone is labeled and the model of the drone is marked, thereby producing a drone RF signal dataset;

[0025] S42: Use the YOLOv5s network model and introduce the attention mechanism into its model structure to improve it, so as to build a RF signal detection neural network;

[0026] S43: Input the data set into the improved RF signal detection neural network for training, and then deploy the trained RF signal detection neural network to the computing device.

[0027] The YOLO target detection scheme is used in the present invention to detect multiple targets under a single time-frequency graph sample, realize multi-target detection and recognition of drone RF signals within a fixed frequency band, and output signal frequency, bandwidth and other information.

[0028] Preferably, in step S42, the introduced attention mechanism is the SE attention mechanism, and the SE architecture replaces the original SPPF structure in YOLOv5s, so that the weight of the channel feature can be adaptively adjusted, and then the obtained channel weight vector is multiplied element by element with the original feature map to obtain the final feature output; the final feature output formula is:

[0029] output = x·s

[0030] Among them, output represents the feature output, x represents the original feature map, and s represents the channel weight vector.

[0031] In this application, in order to further improve the performance of the target detection neural network, the YOLOv5s model is also improved by introducing the SE attention mechanism to replace the original SPPF structure in YOLOv5s with the SE architecture. By explicitly modeling and recalibrating the channel features, the problem that the importance of the feature map of each channel to the final task may be different is solved. The model can adaptively learn the importance of each channel and reweight the original feature map, thereby improving the network's sensitivity to useful features and suppressing irrelevant features. When facing complex tasks, the model can capture key information more accurately, thereby significantly improving the feature expression ability of the model and improving the overall performance of the model.

[0032] Further preferably, the SE architecture includes:

[0033] S51: compress the feature map of each channel in the spatial dimension to obtain a scalar representing the global information of the channel;

[0034] S52: Learn the weight of each channel so as to reweight the original feature map and obtain a channel weight vector.

[0035] The SE architecture has good versatility and flexibility, and can be easily integrated into the existing CNN architecture without changing the overall architecture of the network. It also has high computational efficiency and can perform well even in resource-limited environments. At the same time, by adaptively adjusting the weights of channel features, the stability and accuracy of the model can be maintained, the model reasoning efficiency and reasoning performance can be improved, thereby improving the accuracy of drone RF signal detection.

[0036] More preferably, in step S51, the formula is:

[0037]

[0038] Among them, z c represents the global information scalar of the cth channel, H and W represent the height and width of the feature map respectively, x c，i，j Represents the eigenvalue of the cth channel at position (i, j).

[0039] More preferably, in step S52, a self-gating mechanism is used, wherein the mechanism includes one or more fully connected layers and a nonlinear activation function, and the formula is:

[0040] s=σ(W2δ(W1z))

[0041] Among them, s represents the learned channel weight vector, σ represents the Sigmoid activation function, δ represents ReLU or other nonlinear activation functions, and W1 and W2 represent the matrices of linear transformation.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] 1. With the help of YOLOv5s target detection neural network, the present invention transforms the problem of drone radio frequency signal detection into an image target detection problem, and then detects the drone signal in the time-frequency graph through image processing to achieve the purpose of multi-drone intelligent target detection.

[0044] 2. The present invention improves the YOLOv5s target detection model by replacing the SPPF multi-scale feature extraction fusion structure with a Squeeze-and-Excitation self-attention structure, thereby reducing the number of neural network parameters and improving the neural network reasoning performance.

[0045] 3. The present invention performs lightweight processing on the RF signal data, greatly reduces the dimension of the image data, improves the transmission efficiency between the RF signal receiving end and the neural network inference calculation unit, reduces the calculation pressure of the neural network image processing, and smoothes the noise of the time-frequency diagram data. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A schematic diagram of the detection method provided in this embodiment.

[0047] Figure 2 A schematic diagram of a grayscale image sample with annotated image transmission signals of a drone provided in this embodiment.

[0048] Figure 3 A schematic diagram of a grayscale image sample with control signals of a drone and the model of the drone marked provided in this embodiment.

[0049] Figure 4 A schematic diagram of the improved RF signal detection neural network structure provided in this embodiment. DETAILED DESCRIPTION

[0050] The drawings of the present invention are only for illustrative purposes and should not be construed as limiting the present invention. In order to better illustrate the following embodiments, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product; it is understandable to those skilled in the art that some well-known structures and their descriptions in the drawings may be omitted.

[0051] Example 1

[0052] like Figure 1 As shown, this embodiment provides a multi-UAV radio frequency signal detection method based on artificial intelligence, and the method includes:

[0053] S1: Collect the RF time domain data of the drone through the RF signal terminal;

[0054] S2: converting the time domain data into time-frequency data and compressing the data, and then transmitting the compressed data to a computing device;

[0055] S3: In the computing device, the compressed data is visualized and converted into a grayscale image, and then the grayscale image is input into a radio frequency signal detection neural network deployed on the computing device for detection, and a final detection result is output, so as to identify the corresponding drone type.

[0056] In this embodiment, a multi-UAV RF signal detection method based on artificial intelligence is proposed. The image features of the time-frequency graph of the UAV RF signal are extracted by using a neural network, and the time-frequency matrix data is visualized as a grayscale image. The UAV signal in the time-frequency graph is detected by image processing, and the UAV RF signal detection problem is directly converted into an image target detection problem, thereby achieving the purpose of multi-UAV intelligent target detection and realizing the multi-target recognition and detection function under a single time-frequency graph sample. In addition, the RF signal data is lightweight processed, which greatly reduces the dimension of the image data, improves the transmission efficiency between the RF signal receiving end and the neural network reasoning calculation unit, reduces the calculation pressure of the neural network image processing, and can effectively increase the number of detections of the entire recognition and detection system per unit time.

[0057] Preferably, in step S1, the drone needs to be started. If the drone has an image transmission function, the image transmission function needs to be turned on. Secondly, since drones usually support 2.4GHz and 5.8GHz frequency band communications, the sampling rate of the RF signal receiver at the RF signal end is set to 100MS / s, and the center frequency is 2440MHz or 5800MHz, and then the RF time domain data of the drone is collected.

[0058] Preferably, step S2 includes:

[0059] First, in step S21, 10M time domain data points are continuously collected for the UAV radio frequency signal;

[0060] In step S22, the collected time domain data is subjected to short-time Fourier transform. In this embodiment, the 10M time domain data points are subjected to Fourier transform every 8000 consecutive points, and the center frequency component is transferred to the center of the array, and then the beginning and end 800 points of the frequency component are removed, so as to convert the time domain data into a 6400×125 time-frequency information complex matrix;

[0061] Then, in step S23, in order to reduce the transmission volume between the radio frequency signal receiving end and the neural network inference end, the time-frequency information complex matrix is ​​compressed by dimensionality reduction along the frequency axis, thereby obtaining a compressed time-frequency graph sample.

[0062] In step S23, the dimension reduction compression formula is:

[0063]

[0064] in, Represents complex matrix data; f avg represents the compression operator; k represents the first dimension subscript of the complex matrix before compression; i represents the first dimension subscript of the complex matrix after compression; j represents the second dimension subscript of the complex matrix after compression; the compressed sample data is a 128×125 complex matrix, stored in a 32-bit floating point format, and the compressed RF data is 128kb.

[0065] Therefore, in the above steps, by using short-time Fourier transform to convert time domain data into time-frequency data and then compressing it, the data size of the image data can be effectively reduced, the transmission efficiency from the sensor end to the neural network computing end can be greatly improved, and the number of RF signal samples that the neural network can process per unit time can be increased.

[0066] Preferably, in step S3, the visualization processing includes: normalizing the time-frequency graph samples obtained after compression in step S2 and converting them into an int8 integer matrix, and then directly storing the obtained integer matrix as a grayscale image to obtain a visualized image, and inputting the grayscale image into a radio frequency signal detection neural network deployed on a computing device for detection, thereby obtaining the final detection result, thereby identifying the corresponding drone type.

[0067] Further preferably, the calculation formula for converting the integer matrix into int8 is:

[0068]

[0069] Among them, x gray (i, j) represents an integer matrix; Represents complex matrix data; Represents the absolute value of the complex matrix element to be maximized.

[0070] In this application, by converting the compressed time-frequency graph data into an integer matrix, the time-frequency data can be converted into a grayscale image through the integer matrix, thereby converting the drone RF signal detection problem into an image target detection problem. The drone model in the time-frequency graph can be detected by image processing, thereby achieving the purpose of multi-drone intelligent target detection.

[0071] Preferably, the method in this embodiment further includes building and training a radio frequency signal detection neural network. In order to realize radio frequency signal detection and recognition of multiple drones, this embodiment adopts a self-attention-based YOLOv5s image target detection model to realize the task, and the steps are as follows:

[0072] Step S41: First, it is necessary to collect and annotate the data set of the RF signals of different types of drones. The equipment required for data set collection includes drones and flight controllers, RF receivers, gain antennas, host computers and connecting cables, etc. The data collection process requires starting the drone, and then using the methods of steps S2 and S3 to collect and process. The specific steps are not repeated here, so as to obtain several grayscale image samples of RF signals of different types of drones, and then use LabelImg, Labelbox, VIA, RectLabel, Makesense.ai and other image annotation tools to annotate all grayscale image samples. If there is an image transmission signal, the image transmission signal of the drone is annotated. The grayscale image sample annotated with the image transmission signal of the drone is as follows: Figure 2 If not, the control signal of the drone is marked and the model of the drone is marked. The grayscale image sample marked with the control signal of the drone and the model of the drone is as follows: Figure 3 As shown, a drone radio frequency signal data set is produced; then after the labeling is completed, the data set is divided. In this embodiment, 80% of the data set is used as a training set, 10% of the data set is used as a verification set, and 10% of the data set is used as a test set. Optionally, 70% of the data set can be used as a training set, 20% of the data set can be used as a verification set, and 10% of the data set can be used as a test set.

[0073] Step S42: In this method, a lightweight YOLOv5s network model is used to realize the target detection of the UAV radio frequency signal, and an attention mechanism is introduced into its model structure to improve it, thereby building a radio frequency signal detection neural network; wherein the radio frequency signal detection neural network structure is as follows Figure 4 As shown, it mainly includes:

[0074] Input: Preprocess the input grayscale sample image, including image resizing, normalization, mosaic data enhancement, and anchor frame adjustment.

[0075] Backbone network: YOLOv5s uses a simplified version of CSPDarknet53 (commonly referred to as CSPDarknet53-tiny) as the backbone network for feature extraction. CSPDarknet53 combines the advantages of cross-stage partial connection (CSP) and Darknet network, enhancing the learning ability of the network through cross-layer connection while reducing the amount of calculation. CSPDarknet53-tiny consists of multiple residual blocks and CSP modules, which are stacked through skip connections to gradually reduce the resolution of the feature map and increase the depth of the feature map.

[0076] Neck: The neck of YOLOv5s combines the ideas of Feature Pyramid Network (FPN) and Path Aggregation Network (PANet) to enhance feature representation. FPN transfers semantic information from top to bottom, while PANet adds a bottom-up path on this basis to effectively transfer position information back to high-level feature maps, thereby enhancing feature fusion at different scales.

[0077] Head: The head of YOLOv5s consists of multiple YOLO layers, each of which is responsible for predicting the bounding box, confidence, and category of the target at a specific feature scale. The YOLO layer divides the feature map into S×S grids, and each grid is responsible for predicting the center point of the target within the grid (if the center of the target falls within the grid). At the same time, each grid predicts multiple bounding boxes, the confidence of each bounding box (indicating the probability of the existence of the target in the bounding box and the accuracy of the bounding box), and the probability that the target belongs to each category.

[0078] Post-processing: Before outputting the final detection results, YOLOv5s will perform non-maximum suppression on the predicted multiple bounding boxes to eliminate overlapping bounding boxes, thereby ensuring that each target is accurately marked by only one bounding box, and obtaining the image transmission signal or control signal of the drone and the identification detection results of the drone model.

[0079] Preferably, in step S42, the attention mechanism introduced is the SE attention mechanism. For the SPPF feature extraction structure in YOLOv5s, considering that the target features of this task do not require high-complexity multi-scale feature fusion after dimensionality reduction, in this embodiment, the self-attention structure Squeeze-and-Excitation (SE) is used to replace the original SPPF structure to improve the feature expression ability of the model.

[0080] In traditional convolutional neural networks (CNNs), convolutional layers extract local features by sliding convolution kernels on input feature maps, and construct deep feature representations by stacking multiple convolutional layers. However, this processing method often ignores the differences between different channel features, that is, the importance of feature maps of each channel to the final task may be different. The SE architecture solves this problem by explicitly modeling and recalibrating channel features. It enables the model to adaptively learn the importance of each channel and reweight the original feature map, thereby improving the network's sensitivity to useful features and suppressing irrelevant features. This mechanism enables the model to capture key information more accurately when facing complex tasks, thereby improving overall performance. Secondly, the SE module also has good versatility and flexibility, can be plug-and-play, and can be easily integrated into existing CNN architectures such as ResNet, Inception, MobileNet, etc., without changing the overall structure of the network, providing strong support for improving model performance. At the same time, the computational efficiency of the SE module is relatively high. Although it introduces additional parameters and computational complexity, the increased complexity is relatively small, and it can bring significant performance improvements, which enables the SE architecture to perform well in resource-limited environments. In addition, the SE architecture also has a certain robustness. When facing changes in data sets or task differences, the SE module can maintain the stability and accuracy of the model by adaptively adjusting the weights of channel features. This feature enables the SE architecture to perform well in a variety of application scenarios. Therefore, in this embodiment, the yolov5s model is improved by replacing the original SPPF structure in YOLOv5s with the SE architecture, which effectively reduces the number of model parameters and improves the model reasoning efficiency and reasoning performance.

[0081] Specifically, after using the SE architecture to replace the original SPPF structure in YOLOv5s, the improved RF signal detection neural network is mainly divided into the following steps:

[0082] Step S51: compress the feature map of each channel in the spatial dimension, where the spatial dimension includes height and width, so as to obtain a scalar representing the global information of the channel, which is mainly achieved through global average pooling, and the formula is:

[0083]

[0084] Among them, z c represents the global information scalar of the cth channel, H and W represent the height and width of the feature map respectively, x c，i，j Represents the eigenvalue of the cth channel at position (i, j).

[0085] Step S52: Then learn the weight of each channel so as to reweight the original feature map and obtain the channel weight vector. This is usually achieved through a self-gating mechanism, which includes one or more fully connected layers (also called bottleneck layers) and nonlinear activation functions (such as ReLU and Sigmoid), and its formula is:

[0086] s=σ(W2δ(W1z))

[0087] Among them, s represents the learned channel weight vector, σ represents the Sigmoid activation function, δ represents ReLU or other nonlinear activation functions, and W1 and W2 represent the matrices of linear transformation.

[0088] Then, the obtained channel weight vector is multiplied element by element with the original feature map to obtain the final feature output; the final feature output formula is:

[0089] output = x·s

[0090] Among them, output represents the feature output, x represents the original feature map, and s represents the channel weight vector.

[0091] Thus, a YOLOv5s UAV RF signal target detection neural network model based on SE self-attention mechanism was constructed.

[0092] Finally, in step S43, the data set prepared in step S41 is input into the improved RF signal detection neural network in step S42 for training. The trained RF signal detection neural network is then deployed to the computing device, and the RF signal receiving end and data transmission method are established. This can realize the neural network reasoning of the UAV RF signal, and ultimately detect the UAV signals in the 2400-2480MHz and 5760-5840MHz frequency bands, determine the signal bandwidth and signal frequency, and identify the corresponding UAV type.

[0093] Compared with the prior art, the present invention has the following beneficial effects:

[0094] 1. With the help of YOLOv5s target detection neural network, the present invention transforms the problem of drone radio frequency signal detection into an image target detection problem, and then detects the drone signal in the time-frequency graph through image processing to achieve the purpose of multi-drone intelligent target detection.

[0095] 2. The present invention improves the YOLOv5s target detection model by replacing the SPPF multi-scale feature extraction fusion structure with a Squeeze-and-Excitation self-attention structure, thereby reducing the number of neural network parameters and improving the neural network reasoning performance.

[0096] 3. The present invention performs lightweight processing on the RF signal data, greatly reduces the dimension of the image data, improves the transmission efficiency between the RF signal receiving end and the neural network inference calculation unit, reduces the calculation pressure of the neural network image processing, and smoothes the noise of the time-frequency diagram data.

[0097] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solution of the present invention, and are not intended to limit the specific implementation methods of the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the claims of the present invention shall be included in the protection scope of the claims of the present invention.

Claims

1. A multi-UAV radio frequency signal detection method based on artificial intelligence, characterized in that: The method comprises: S1: Collect the RF time domain data of the drone through the RF signal terminal; S2: converting the time domain data into time-frequency data and compressing the data, and then transmitting the compressed data to a computing device; S3: In the computing device, the compressed data is visualized and converted into a grayscale image, and then the grayscale image is input into a radio frequency signal detection neural network deployed on the computing device for detection, and a final detection result is output, so as to identify the corresponding drone type.

2. The method for detecting multi-UAV radio frequency signals based on artificial intelligence according to claim 1 is characterized in that: The step S2 includes: S21: Continuously collect several time domain data points for the UAV RF signal; S22: performing short-time Fourier transform on the collected time domain data, transferring the center frequency component to the center of the array, and then removing the frequency component, thereby converting the time domain data into a complex matrix of time-frequency information; S23: Perform dimensionality reduction compression on the complex matrix of time-frequency information along the frequency axis, so as to obtain compressed time-frequency graph samples.

3. The method for detecting multi-UAV radio frequency signals based on artificial intelligence according to claim 2 is characterized in that: In step S23, the dimension reduction compression formula is: in, Represents the compressed complex matrix data; f avg represents the compression operator; k represents the first-dimension subscript of the complex matrix before compression; i represents the first-dimension subscript of the complex matrix after compression; j represents the second-dimension subscript of the complex matrix after compression.

4. The method for detecting multi-UAV radio frequency signals based on artificial intelligence according to claim 3 is characterized in that: In the step S3, the visualization processing includes: normalizing the time-frequency graph samples obtained after compression and converting them into an int8 integer matrix, and then directly storing the obtained integer matrix as a grayscale image to obtain a visualized image.

5. The method for detecting multi-UAV radio frequency signals based on artificial intelligence according to claim 4 is characterized in that: The calculation formula for the integer matrix converted to int8 is: Among them, x gray (i, j) represents an integer matrix; Represents complex matrix data; Represents the absolute value of the complex matrix element to be maximized.

6. The method for detecting multi-UAV radio frequency signals based on artificial intelligence according to any one of claims 1 to 5, characterized in that: The method also includes building and training a radio frequency signal detection neural network, the steps of which are: S41: using the methods of steps S2 and S3 to obtain several grayscale image samples of RF signals of different types of drones, and then labeling all samples. If there is an image transmission signal, the image transmission signal of the drone is labeled; if not, the control signal of the drone is labeled and the model of the drone is marked, thereby producing a drone RF signal dataset; S42: Use the YOLOv5s network model and introduce the attention mechanism into its model structure to improve it, so as to build a RF signal detection neural network; S43: Input the data set into the improved RF signal detection neural network for training, and then deploy the trained RF signal detection neural network to the computing device.

7. The method for detecting multi-UAV radio frequency signals based on artificial intelligence according to claim 6 is characterized in that: In step S42, the introduced attention mechanism is the SE attention mechanism, and the SE architecture replaces the original SPPF structure in YOLOv5s, so that the weight of the channel feature can be adaptively adjusted, and then the channel weight vector is multiplied element by element with the original feature map to obtain the final feature output; the final feature output formula is: output = x·s Among them, output represents the feature output, x represents the original feature map, and s represents the channel weight vector.

8. The method for detecting multi-UAV radio frequency signals based on artificial intelligence according to claim 7 is characterized in that: The SE architecture includes: S51: compress the feature map of each channel in the spatial dimension to obtain a scalar representing the global information of the channel; S52: Learn the weight of each channel so as to reweight the original feature map and obtain a channel weight vector.

9. The method for detecting multi-UAV radio frequency signals based on artificial intelligence according to claim 8, characterized in that: In step S51, the formula is: Among them, z c represents the global information scalar of the cth channel, H and W represent the height and width of the feature map respectively, x c，i，j Represents the eigenvalue of the cth channel at position (i, j).

10. The method for detecting multi-UAV radio frequency signals based on artificial intelligence according to claim 9 is characterized in that: In step S52, a self-gating mechanism is used, which includes one or more fully connected layers and a nonlinear activation function, and its formula is: s=σ(W2δ(W1z)) Among them, s represents the learned channel weight vector, σ represents the Sigmoid activation function, δ represents ReLU or other nonlinear activation functions, and W1 and W2 represent the matrices of linear transformation.

Images