A deep learning-based unmanned aerial vehicle signal detection method and system

Abstract

The application discloses a kind of unmanned aerial vehicle signal detection method and system based on deep learning belongs to unmanned aerial vehicle signal detection field.This method first executes basic parameter definition, data reading, time-frequency transform and so on to IQ signal Generation time-frequency chart, mark unmanned aerial vehicle chart transmission signal in time-frequency chart and construct data set, train deep learning model containing multi-scale feature extraction module, SE module and deep convolution layer, again execute same processing to target signal Generation target time-frequency chart and input trained model, output time-frequency domain position information and confidence data and can be visualized.The system corresponds to the above method, through standardization process and light weight model design, improve signal feature extraction capability and anti-interference, reduce model parameter quantity, adapt edge device deployment, meet the detection demand under complex environment.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) signal detection, and in particular to a UAV signal detection method and system based on deep learning. Background Technology

[0002] With the rapid development and widespread adoption of drone technology, its applications in civilian aerial photography, logistics transportation, and agricultural plant protection are becoming increasingly profound, bringing great convenience to production and daily life. At the same time, the disorderly flight of drones in scenarios such as public safety and protection of critical facilities also poses potential risks, driving anti-drone systems to become a key technological equipment for ensuring safety. In the workflow of anti-drone systems, drone signal detection is a core preliminary step for subsequent target identification, positioning, and interception. Time-frequency maps, as an important carrier of the time-frequency domain characteristics of signals, have become a research hotspot in the extraction of drone image transmission signals. Because drone image transmission signals need to transmit high-definition images and video data in real time, they typically have wide bandwidth, wide pulse width, and continuous transmission characteristics. This is reflected in the time-frequency map as a clearly defined, regularly shaped rectangular form, providing an important basis for signal extraction. Currently, the technologies for extracting UAV signals from time-frequency maps are mainly divided into two categories: traditional algorithms and deep learning algorithms. Traditional algorithms are mostly based on grayscale differences or edge features in time-frequency maps to extract signals, while deep learning algorithms use convolutional neural networks for feature learning and target extraction. Both types of technologies have been applied to a certain extent in different scenarios, promoting the continuous development of anti-UAV system technology.

[0003] Despite the progress made in existing UAV signal detection technology, several technical challenges remain to be addressed in practical applications. First, current technologies lack standardized signal processing procedures. The lack of unified standards in parameter definition, data reading, and time-frequency transformation during time-frequency map generation leads to inconsistent time-frequency map features across different scenarios, impacting the accuracy of subsequent signal detection. Second, existing methods lack systematic design in time-frequency map annotation and dataset construction. The integration of annotation information and time-frequency map data is not standardized, resulting in insufficient generalization ability of the trained models and difficulty in adapting to UAV signal detection needs in different environments. Third, the connection between model training and inference processes is not tight enough. Inconsistencies between the time-frequency transformation processing of the signal to be detected and the training set reduce the reliability of the model's inference results. Finally, the presentation of results by existing detection methods is relatively simple, lacking intuitive visualization capabilities. This hinders staff from quickly obtaining target signal location information and confidence data, affecting the efficiency of detection work. These problems prevent existing UAV signal detection technology from meeting the comprehensive requirements of anti-UAV systems for detection accuracy, process standardization, and result intuitiveness, limiting its application effectiveness in real-world scenarios. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for detecting drone signals based on deep learning.

[0005] The objective of this invention is achieved through the following technical solution: A deep learning-based method for detecting drone signals is provided, comprising the following steps: S1. Process the IQ signal to generate a time-frequency graph. The processing steps include defining basic parameters, reading IQ data, performing time-frequency transformation, converting to decibel scale, calibrating the frequency axis, and mapping colors to generate the time-frequency graph. S2. Label the UAV image transmission signals in the time-frequency graph generated by processing IQ signals, integrate the time-frequency graph and the labeling information to construct a dataset, and use the dataset to train a preset deep learning model until the model converges; S3. Perform the same operation on the target signal as on the IQ signal to generate the target time-frequency map, and input the target time-frequency map into the trained deep learning model; S4. The deep learning model outputs the time-frequency domain location information and confidence data of the target signal to realize the visualization of the detection results.

[0006] Furthermore, step S1 includes the following sub-steps: S1.1. Define the basic parameters of sampling rate, bandwidth, and number of Fourier transform points, and derive the number of time-domain sampling points and number of frames for a single time-frequency analysis; S1.2. Read IQ data that meets the defined basic parameter requirements from the binary file in blocks; S1.3. Perform a short-time Fourier transform on the read IQ data to obtain the time-frequency domain matrix, convert the amplitude spectrum of the time-frequency domain matrix to a decibel scale and introduce a very small positive number to avoid calculation anomalies; S1.4. The frequency axis is calibrated from relative frequency to absolute frequency to generate a time-frequency plot data matrix, and the spectral data values ​​at the decibel scale are mapped to the corresponding colors to generate a time-frequency plot.

[0007] Furthermore, step S2 includes the following sub-steps: S2.1. By using annotation software, the time-frequency domain position and range of the UAV image transmission signal in the time-frequency diagram generated by processing the IQ signal are clearly defined, and the relative position information of the signal is obtained; S2.2. Divide the time-frequency graph with the acquired relative position information into a training set and a validation set according to a set ratio, and integrate them to form a dataset for model training and validation; S2.3. Construct a deep learning model based on the Yolov5s architecture. The deep learning model includes a multi-scale feature extraction module, an SE module, a detection head, and a backbone network. The backbone network uses deep convolutional layers. S2.4. Set the optimizer and learning rate variation strategy, train the deep learning model using the partitioned training set, verify the model performance using the partitioned validation set, and continue iterating until the loss function value and performance index of the deep learning model reach the set values, and obtain the trained model weight file.

[0008] Furthermore, step S3 includes the following sub-steps: S3.1. Obtain the target signal to be analyzed, and perform processing on the target signal by defining basic parameters and reading data in blocks; S3.2. Perform short-time Fourier transform, decibel conversion, frequency calibration, and color mapping operations on the processed target signal to generate a target time-frequency diagram; S3.3. Adjust the generated target time-frequency map according to the set format to ensure that the parameters of the target time-frequency map match the time-frequency parameters of the divided training set. Figure 1 To; S3.4. Input the adjusted target time-frequency map into the trained deep learning model and start the feature extraction operation of the deep learning model.

[0009] Furthermore, step S4 includes the following sub-steps: S4.1. The deep learning model performs signal recognition inference operations on the input target time-frequency map; S4.2. Output the time-frequency domain location information and confidence data corresponding to the target signal obtained by the inference operation; S4.3. Use the output time-frequency domain location information and confidence data to mark the location of the target signal in the target time-frequency map, and synchronously associate the corresponding confidence data; S4.4. The time-frequency graph of the marked target is displayed in a visual way to achieve an intuitive presentation of the detection results.

[0010] Furthermore, in step S1, the Hanning window is used to perform sliding processing on the read IQ data. The sliding step size is matched with the number of frames in the defined basic parameters. The time-domain IQ signal is mapped to a time-frequency domain matrix through Fourier transform to ensure the continuity of time-frequency analysis.

[0011] Furthermore, in step S2, the multi-scale feature extraction module includes four convolutional layers set in parallel. The four convolutional layers use convolutional kernels of different specifications and have the same stride to extract fine-grained features from the input feature map at different levels. The tensors output by the four convolutional layers are concatenated in the channel dimension. The number of channels of the concatenated tensor is adjusted by a point convolutional layer, and the concatenated features are fused with the ReLU activation function to achieve data compression.

[0012] Furthermore, in step S2, the SE module is set at the back end of the backbone network of the deep learning model. The SE module includes compression and activation operations. The compression operation compresses the spatial dimension of the input feature map into a single-channel descriptor through global average pooling. The channel descriptor represents the global distribution of each channel. The activation operation processes the channel descriptor through two fully connected layers, inserts a ReLU activation function in between, and uses the Sigmoid function to normalize the processed result to generate channel weights. The channel weights are multiplied with the original input feature map channel by channel to achieve feature recalibration, and the output is a recalibrated feature map with the same specifications as the original input feature map.

[0013] Furthermore, in step S2, the deep learning model adds a new detection head on top of the original detection head, and the parameters of the new detection head are matched with those of the original detection head; the conventional convolutional layers in the backbone network are replaced with deep convolutional layers, and the deep convolutional layers perform convolution operations on each input channel separately. The output of the deep convolutional layers is channel fused by point convolutional layers, thereby reducing the number of parameters and computational overhead of the deep learning model.

[0014] In some embodiments, a deep learning-based drone signal detection system is provided, which includes a time-frequency graph generation module, a data annotation and dataset construction module, a model training module, and a model inference module. The time-frequency graph generation module is used to perform basic parameter definition, block data reading, short-time Fourier transform, decibel conversion, frequency calibration and color mapping on the IQ signal to generate a time-frequency graph that meets the detection requirements. The data annotation and dataset construction module is used to annotate the location of UAV image transmission signals in the time-frequency map generated by annotation software, obtain the relative position information of the signals, divide the time-frequency map with relative position information into training set and validation set according to a set ratio, and integrate them to form a dataset. The model training module is used to build a deep learning model based on the Yolov5s architecture. The deep learning model includes a multi-scale feature extraction module, an SE module, a detection head, and a backbone network. The backbone network uses deep convolutional layers. An optimizer and learning rate variation strategy are set. The deep learning model is trained using the integrated training set. The model performance is verified using the integrated validation set. The iteration continues until the loss function value and performance index reach the set values. Finally, the trained model weight file is output. The model inference module performs the same processing flow as the time-frequency map generation module on the target signal to generate the target time-frequency map. After adjusting the target time-frequency map according to the set format, it inputs the deep learning model with the loaded output model weight file. It receives the position information and confidence data output by the deep learning model, marks the position in the target time-frequency map using the position information and confidence data and associates it with the confidence data. The detection results are displayed through a visualization component. The modules achieve orderly data transmission and interaction through data interfaces.

[0015] The beneficial effects of this invention are: (1) By performing standardized time-frequency transformation on the IQ signal to generate a time-frequency map, a dataset is constructed after labeling to train a deep learning model and carry out inference detection, so as to achieve accurate output of the time-frequency domain position and confidence of the UAV signal, realize the visualization of the detection results, and build a standardized UAV signal detection process. (2) By adding a multi-scale feature extraction module and a channel attention mechanism to the deep learning model, the model can fully capture the features of different levels of UAV signals, strengthen the effective signal feature expression and suppress interference, and improve the robustness of the model in signal detection in complex environments. (3) By lightweighting the backbone network of the deep learning model and optimizing the detection head structure, the number of model parameters and computational overhead are effectively reduced, making the model adaptable to the deployment requirements of resource-constrained edge devices and improving the practical application adaptability of the detection method. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the specific steps of a deep learning-based drone signal detection method. Figure 2 A flowchart illustrating the overall implementation process of drone signal detection; Figure 3 This is a diagram showing the overall structure of a deep learning model based on Yolov5s. Figure 4 Here is a flowchart illustrating the SE module structure and implementation process; Figure 5 A flowchart illustrating the structure and implementation of deep convolutional layers; Figure 6 This is a graph showing the training process and performance trends of a deep learning model. Figure 7 A visualization of drone signal detection results in a typical scenario; Figure 8 This is a visualization of drone signal detection results in scenarios with low signal-to-noise ratio and strong interference. Detailed Implementation

[0017] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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.

[0018] Example 1 In some embodiments, see Figure 1This paper provides a deep learning-based method for detecting drone signals, including the following steps: S1. Process the IQ signal to generate a time-frequency graph. The processing steps include defining basic parameters, reading IQ data, performing time-frequency transformation, converting to decibel scale, calibrating the frequency axis, and mapping colors to generate the time-frequency graph. S2. Label the UAV image transmission signals in the time-frequency graph generated by processing IQ signals, integrate the time-frequency graph and the labeling information to construct a dataset, and use the dataset to train a preset deep learning model until the model converges; S3. Perform the same operation on the target signal as on the IQ signal to generate the target time-frequency map, and input the target time-frequency map into the trained deep learning model; S4. The deep learning model outputs the time-frequency domain location information and confidence data of the target signal to realize the visualization of the detection results.

[0019] In some embodiments, step S1 includes the following sub-steps: S1.1. Define the basic parameters of sampling rate, bandwidth, and number of Fourier transform points, and derive the number of time-domain sampling points and number of frames for a single time-frequency analysis; S1.2. Read IQ data that meets the defined basic parameter requirements from the binary file in blocks; S1.3. Perform a short-time Fourier transform on the read IQ data to obtain the time-frequency domain matrix, convert the amplitude spectrum of the time-frequency domain matrix to a decibel scale and introduce a very small positive number to avoid calculation anomalies; S1.4. The frequency axis is calibrated from relative frequency to absolute frequency to generate a time-frequency plot data matrix, and the spectral data values ​​at the decibel scale are mapped to the corresponding colors to generate a time-frequency plot.

[0020] In some embodiments, step S2 includes the following sub-steps: S2.1. By using annotation software, the time-frequency domain position and range of the UAV image transmission signal in the time-frequency diagram generated by processing the IQ signal are clearly defined, and the relative position information of the signal is obtained; S2.2. Divide the time-frequency graph with the acquired relative position information into a training set and a validation set according to a set ratio, and integrate them to form a dataset for model training and validation; S2.3. Construct a deep learning model based on the Yolov5s architecture. The deep learning model includes a multi-scale feature extraction module, an SE module, a detection head, and a backbone network. The backbone network uses deep convolutional layers. S2.4. Set the optimizer and learning rate variation strategy, train the deep learning model using the partitioned training set, verify the model performance using the partitioned validation set, and continue iterating until the loss function value and performance index of the deep learning model reach the set values, and obtain the trained model weight file.

[0021] In some embodiments, step S3 includes the following sub-steps: S3.1. Obtain the target signal to be analyzed, and perform processing on the target signal by defining basic parameters and reading data in blocks; S3.2. Perform short-time Fourier transform, decibel conversion, frequency calibration, and color mapping operations on the processed target signal to generate a target time-frequency diagram; S3.3. Adjust the generated target time-frequency map according to the set format to ensure that the parameters of the target time-frequency map match the time-frequency parameters of the divided training set. Figure 1 To; S3.4. Input the adjusted target time-frequency map into the trained deep learning model and start the feature extraction operation of the deep learning model.

[0022] In some embodiments, step S4 includes the following sub-steps: S4.1. The deep learning model performs signal recognition inference operations on the input target time-frequency map; S4.2. Output the time-frequency domain location information and confidence data corresponding to the target signal obtained by the inference operation; S4.3. Use the output time-frequency domain location information and confidence data to mark the location of the target signal in the target time-frequency map, and synchronously associate the corresponding confidence data; S4.4. The time-frequency graph of the marked target is displayed in a visual way to achieve an intuitive presentation of the detection results.

[0023] In some embodiments, in step S1, the read IQ data is processed by sliding using a Hanning window, with the sliding step size matching the number of frames in the defined basic parameters. The time-domain IQ signal is mapped to a time-frequency domain matrix by Fourier transform to ensure the continuity of time-frequency analysis.

[0024] In some embodiments, in step S2, the multi-scale feature extraction module includes four convolutional layers set in parallel. The four convolutional layers use convolutional kernels of different specifications and have the same stride to extract fine-grained features from the input feature map at different levels. The tensors output by the four convolutional layers are concatenated in the channel dimension. The number of channels of the concatenated tensor is adjusted by a point convolutional layer, and the concatenated features are fused with the ReLU activation function to achieve data compression.

[0025] In some embodiments, in step S2, the SE module is set at the back end of the backbone network of the deep learning model. The SE module includes a compression operation and an activation operation. The compression operation compresses the spatial dimension of the input feature map into a single-channel descriptor through global average pooling. The channel descriptor represents the global distribution of each channel. The activation operation processes the channel descriptor through two fully connected layers, inserts a ReLU activation function in between, and uses the Sigmoid function to normalize the processed result to generate channel weights. The channel weights are multiplied with the original input feature map channel by channel to achieve feature recalibration, and the output is a recalibrated feature map with the same specifications as the original input feature map.

[0026] In some embodiments, in step S2, the deep learning model adds a new detection head to the original detection head, and the parameters of the new detection head are matched with those of the original detection head; the conventional convolutional layers in the backbone network are replaced with deep convolutional layers, and the deep convolutional layers perform convolution operations on each input channel separately, and the output of the deep convolutional layers is fused by point convolutional layers, thereby reducing the number of parameters and computational overhead of the deep learning model.

[0027] In some embodiments, a deep learning-based drone signal detection system is provided, which includes a time-frequency graph generation module, a data annotation and dataset construction module, a model training module, and a model inference module. The time-frequency graph generation module is used to perform basic parameter definition, block data reading, short-time Fourier transform, decibel conversion, frequency calibration and color mapping on the IQ signal to generate a time-frequency graph that meets the detection requirements. The data annotation and dataset construction module is used to annotate the location of UAV image transmission signals in the time-frequency map generated by annotation software, obtain the relative position information of the signals, divide the time-frequency map with relative position information into training set and validation set according to a set ratio, and integrate them to form a dataset. The model training module is used to build a deep learning model based on the Yolov5s architecture. The deep learning model includes a multi-scale feature extraction module, an SE module, a detection head, and a backbone network. The backbone network uses deep convolutional layers. An optimizer and learning rate variation strategy are set. The deep learning model is trained using the integrated training set. The model performance is verified using the integrated validation set. The iteration continues until the loss function value and performance index reach the set values. Finally, the trained model weight file is output. The model inference module performs the same processing flow as the time-frequency map generation module on the target signal to generate the target time-frequency map. After adjusting the target time-frequency map according to the set format, it inputs the deep learning model with the loaded output model weight file. It receives the position information and confidence data output by the deep learning model, marks the position in the target time-frequency map using the position information and confidence data and associates it with the confidence data. The detection results are displayed through a visualization component. The modules achieve orderly data transmission and interaction through data interfaces.

[0028] Example 2 This embodiment addresses the practical needs of UAV signal detection in anti-drone systems by using deep learning technology to accurately extract and detect UAV image transmission signals from time-frequency maps. It solves the problems of weak anti-interference ability of existing traditional algorithms and high complexity and poor robustness of deep learning algorithm models. Through a lightweight model design that enhances feature extraction capabilities, it simplifies the number of model parameters and improves feature extraction capabilities, adapts to the deployment requirements of edge devices, and meets the requirements of UAV signal detection in complex electromagnetic environments and low signal-to-noise ratio scenarios.

[0029] In some specific implementations, the overall implementation process for UAV signal detection is as follows: See Figure 2 The UAV signal detection method used in this embodiment takes the time-frequency map obtained by short-time Fourier transform of the acquired signal as the core processing object. The overall implementation process is divided into three core steps: drawing the original IQ signal time-frequency map, time-frequency map data annotation and dataset production and model training, and model inference. Each step is progressive, and the implementation results of the previous step provide the data foundation and model support for the next step. Finally, UAV signal detection based on time-frequency map is realized, and the output of target signal time-frequency domain position information and confidence data and the visualization of detection results are completed.

[0030] In practice, the process begins by using IQ signals as the processing target. A time-frequency map is generated through parameter definition, calculation, data reading, time-frequency transformation, and frequency calibration. Then, the UAV image transmission signals to be detected in the time-frequency map are manually labeled using annotation software. The labeled time-frequency maps are then integrated into a dataset, which is used to train a pre-built deep learning model. Once the model converges and its performance meets preset metrics, the trained model weight file is obtained. Finally, the signal to be analyzed undergoes the same time-frequency transformation as the training set to obtain a time-frequency map. This map is then input into the trained model for inference. The model outputs the time-frequency domain location information of the target signal and the corresponding confidence score data, and the detection results are visualized. Throughout the entire process, the short-time Fourier transform, as the core time-frequency transformation operation, is used to ensure consistency in the feature dimensions of the time-frequency maps in the training and inference stages, thus guaranteeing the accuracy of the model's detection.

[0031] In some specific implementations, the generation of the original IQ signal time-frequency diagram is carried out as follows: Generating the original IQ signal time-frequency map is a fundamental step in UAV signal detection. The process involves parameter definition, data reading, time-frequency transformation, decibel scale conversion, frequency calibration, and color mapping, ensuring that the generated time-frequency map can completely and accurately reflect the time-frequency domain characteristics of the IQ signal, providing a high-quality data source for subsequent signal annotation and model training.

[0032] 1. Definition and Calculation of Basic Parameters: First, define basic parameters such as sampling rate, bandwidth, and number of Fourier transform points. Based on the defined parameters, derive the number of time-domain sampling points and the number of frames for a single time-frequency analysis. This derivation process ensures the integrity of the frame structure in the time-frequency analysis and avoids signal feature distortion caused by frame structure mismatch.

[0033] 2. IQ Data Reading: IQ data that meets the calculation parameter requirements is read in blocks from the binary file to ensure that the read IQ data matches the basic parameters, thus ensuring the accuracy of subsequent time-frequency transformation.

[0034] 3. Short-Time Fourier Transform: Perform a short-time Fourier transform on the read IQ data to map the time-domain IQ signal into a time-frequency domain matrix. The formula for the short-time Fourier transform is: ; in, For time-domain IQ signals, For window functions, the Hanning window is used in this embodiment. The number of points in the Fourier transform. For frame indexing, representing the time dimension, For frequency indexing.

[0035] 4. Decibels scale conversion: Convert the amplitude spectrum obtained from the short-time Fourier transform to a decibel scale. The conversion formula is as follows: ; in, This is the time-frequency plot matrix after conversion to decibel scale. It is an extremely small positive number, taken from the built-in MATLAB value eps. This value is introduced to avoid abnormal situations that may occur during the calculation process.

[0036] 5. Frequency calibration and color mapping: After completing the decibel scale conversion, the frequency axis is calibrated from relative frequency to absolute frequency to generate a time-frequency graph data matrix. Finally, the spectral data values ​​of the decibel scale are mapped to the corresponding colors to generate the time-frequency graph required for subsequent detection.

[0037] After this step is completed, the resulting time-frequency diagram can clearly show the time-frequency domain distribution characteristics of the IQ signal. Due to the characteristics of wide bandwidth, wide pulse width and continuous transmission, the UAV image transmission signal presents a rectangular shape with relatively clear boundaries and regular shape in the time-frequency diagram, providing clear feature basis for subsequent signal labeling.

[0038] In some specific implementations, the overall construction and implementation of deep learning models are as follows: See Figure 3 The deep learning model constructed in this embodiment is based on the improved Yolov5s architecture. It consists of a multi-scale feature extraction module, a backbone network, a neck network, and four detection heads. The core design of the model is to improve the feature extraction capability while reducing the number of model parameters. The backbone network adds an SE module at the end, and the conventional convolutional layers in the backbone network are replaced with deep convolutional layers. At the same time, a new detection head is added on the basis of the original detection head. Each module realizes the transfer and fusion of features through tensor concatenation, upsampling, downsampling and other operations to form a complete feature extraction and detection system.

[0039] The specific implementation details of each component of the model are as follows: 1. The multi-scale feature extraction module contains four parallel convolutional layers with different kernel sizes, followed by convolutional layers with connection points, to achieve feature extraction and feature fusion at different scale levels; 2. Both CONCATENATE and CONCAT are tensor concatenation operations, which fuse multiple feature maps in the channel dimension according to feature extraction requirements; 3. DWConv stands for deep convolutional layer, which is the core structure for lightweight model design, replacing the conventional convolutional layers in the backbone network to reduce the number of parameters; 4. C3 is the basic layer of the model, consisting of ConvBNSiLU and BottleNeck. ConvBNSiLU is a concatenation of a convolutional layer, a normalization layer, and a SiLU activation function. BottleNeck is two concatenated ConvBNSiLU layers with added residual connections, which enables deep feature extraction and alleviates the gradient vanishing problem. 5. SSPF is the feature extraction layer of the model. It works in conjunction with the Maxpooling layer to reduce the spatial dimension of the feature map through downsampling operations, thereby increasing the receptive field of the model. 6. The SE-layer is a compressed excitation module, located at the end of the backbone network, which achieves feature recalibration by dynamically adjusting the channel weights; 7. Unsample is an upsampling layer that works in conjunction with downsampling to adjust the dimension of the feature map and fuse features from deep and shallow layers. 8. Detect is the detection head. The four detection heads correspond to feature maps of different dimensions, enabling the detection of UAV signals of different sizes in the time-frequency map.

[0040] The entire model is built around the time-frequency domain characteristics of UAV image transmission signals. The modules work together to ensure that the model can fully extract the signal features, and the lightweight design reduces the number of parameters, making it suitable for the deployment requirements of resource-constrained devices in anti-drone systems.

[0041] In some specific implementations, the multi-scale feature extraction module is implemented as follows: The multi-scale feature extraction module is the core module for improving the model's feature extraction capability. It is set at the front end of the deep learning model. It captures fine-grained features of different scales of UAV image transmission signals through four parallel convolutional layers with different kernel sizes. Then, it achieves feature fusion and data compression through point convolutional layers. The parameter settings of each convolutional layer in the module are strictly matched with the time-frequency domain features of the UAV signal. The specific parameters are shown in Table 1. The implementation of all convolutional layers follows the adaptation principle of convolutional kernel size, stride, and padding to ensure the dimensionality matching of the output feature map and achieve distortion-free feature stitching and fusion.

[0042] Table 1. Parameters of each convolutional layer in the multi-scale feature extraction module Convolutional layer Convolutional kernel size Step size Padding Number of input / output channels CONV1 （2，2） （2，2） （0，0） 3 / 16 CONV2 （4，4） （2，2） （1，1） 3 / 16 CONV3 （8，8） （2，2） （3，3） 3 / 16 CONV4 （16，16） （2，2） （7，7） 3 / 16 CONV5 （1，1） （1，1） （0，0） 64 / 32 The implementation process of this module is as follows: First, the time-frequency image is input into four parallel convolutional layers CONV1, CONV2, CONV3, and CONV4. The four convolutional layers use convolutional kernels of 2×2, 4×4, 8×8, and 16×16 respectively, with a stride of 2×2 for each layer. The padding parameters are adapted to the size of the convolutional kernels to ensure that the number of input and output channels for each convolutional layer is 3 / 16. This enables independent extraction of features at different scales from the time-frequency image. By using the differentiated receptive field design of convolutional kernels of different sizes, the collaborative optimization of feature extraction is achieved, allowing the model to simultaneously capture both the detailed and overall features of the UAV signal.

[0043] After feature extraction from the parallel convolutional layers, the tensors output by the four convolutional layers are concatenated along the channel dimension. The concatenated tensor is then input into the point convolutional layer CONV5. CONV5 uses a 1×1 convolutional kernel with a stride of 1×1 and 0×0 padding. The number of input / output channels is 64 / 32. Simultaneously, the non-linear activation function ReLU is used to fuse the concatenated features, achieving data compression and non-linear transformation of features. This reduces the computational overhead of subsequent model processing and makes the fused features more suitable for the detection requirements of UAV signals.

[0044] The implementation of the multi-scale feature extraction module enables deep learning models to break through the receptive field limitations of a single convolutional kernel and fully extract multi-scale features of UAV image transmission signals in the time-frequency map, laying the foundation for subsequent deep feature extraction. At the same time, through feature fusion and data compression of point convolutional layers, the model's lightweight requirements are taken into account while enriching feature information.

[0045] In some specific implementations, the structure and implementation of the SE module are as follows: See Figure 4 The SE module is a channel attention mechanism module derived from SE-Net. It is placed in the back end of the backbone network of the deep learning model. Its core structure includes two parts: compression and activation. It enhances the feature representation ability by dynamically adjusting the channel weights, thereby achieving the effects of noise suppression and signal enhancement. The implementation of the module does not change the size of the input feature map, but only recalibrates the features in the channel dimension, so that the model can extract effective signal features more accurately and improve the detection robustness in interference environments.

[0046] The specific implementation process of the SE module is as follows: given an input feature map, after mapping condition processing, a feature map of the corresponding specification is obtained. First, a compression operation is performed on the feature map. The compression is performed by global average pooling to compress the spatial dimension of the input feature map into a channel descriptor, obtain the global receptive field and characterize the global distribution of each channel, so that the model can judge the importance of each channel feature from a global perspective.

[0047] In this model, the SE-layer consists of a pooling layer, two fully connected layers, and a ReLU non-linear activation function. After compression, activation is performed: channel weights are generated through the two fully connected layers and the ReLU non-linear activation function to learn the non-linear dependencies between channels. Then, the channel weights are normalized using the Sigmoid function to map the weight values ​​to a reasonable range. Finally, the normalized weights are multiplied with the original feature map to achieve feature recalibration, and the output size is the same as the original input feature map. Figure 1 The recalibrated feature map.

[0048] The compression process of the SE module extracts channel-level information. The subsequent multiplication of feature maps with channel weights enhances the channel features corresponding to the effective signals and suppresses the channel features corresponding to interference noise. This process, by dynamically adjusting the feature channel weights, allows the model to focus on the effective features of the UAV image transmission signal, improving the targeting and effectiveness of feature extraction.

[0049] In some specific implementations, the lightweight design of deep convolutional layers is implemented as follows: See Figure 5To reduce the number of model parameters while maintaining model performance, this embodiment replaces the conventional convolutional layers in the backbone network with depthwise convolutional layers. The depthwise convolutional layer uses depthwise separable convolution as its core structure and consists of two parts: a depthwise convolutional layer and a pointwise convolutional layer. By decoupling the spatial convolution and channel convolution of conventional convolution, the number of parameters in the convolutional layer is significantly reduced. The calculation of the relevant parameters is shown in Table 2, where K is the kernel size, C... in For the number of input channels, C out G represents the number of output channels, and G represents the number of groups in the grouped convolution. The calculation methods for the number of parameters differ significantly for different convolution types. The calculation of the number of parameters for deep convolutional layers is only related to the kernel size and the number of input channels, which is the key to achieving lightweight models.

[0050] Table 2. Comparison of Parameter Quantities of Conventional Convolutional Layers, Grouped Convolutional Layers, and Deep Convolutional Layers Convolution type Parameter calculation Regular convolution <![CDATA[K*K*C in *C out ]]> Grouped convolution (G groups) <![CDATA[K*K*C in *C out / G]]> Deep convolutional layers <![CDATA[K*K*C in ]]> The specific implementation process of the deep convolutional layer is as follows: First, the deep convolution operation is performed. The deep convolution performs a separate convolution operation on each channel of the input feature map. Each convolution kernel corresponds to only one input channel, realizing the independent extraction of spatial features of each channel. After the deep convolution is completed, the point convolution operation is performed. The point convolution uses a 1×1 convolution kernel to perform cross-channel convolution operation on the output of the deep convolution, realizing the fusion of features of each channel and the adjustment of the number of output channels.

[0051] The implementation of depthwise separable convolution splits a single operation of conventional convolution into two independent operations: depthwise convolution and pointwise convolution. This achieves a lightweight model design by reducing the number of parameters without sacrificing feature extraction capabilities. Replacing conventional convolutional layers with depthwise convolutional layers in the backbone network significantly reduces the overall computational cost of the model, enabling efficient inference operations on resource-constrained edge devices. Simultaneously, the combination of depthwise and pointwise convolutions maintains the model's spatial feature extraction and channel fusion capabilities, ensuring that the model's detection performance remains unaffected.

[0052] In some specific implementations, time-frequency chart annotation and dataset construction are carried out as follows: After generating the original IQ signal time-frequency map, the UAV image transmission signals to be detected in the time-frequency map are labeled and a dataset is constructed to provide data support for the training of deep learning models. In this embodiment, the time-frequency map is manually labeled using labeling software to clarify the position and range of the signal in the time-frequency domain and obtain the relative position information of the signal. The labeling process strictly follows the rectangular shape characteristics of the UAV image transmission signal in the time-frequency map, which has relatively clear boundaries and regular shape, to ensure the accuracy of the labeled information.

[0053] After labeling a single time-frequency image, the time-frequency image is integrated with the corresponding signal annotation information. The integrated time-frequency image is then divided into a training set and a validation set according to a predetermined ratio, constructing a dataset for subsequent model training and validation. During dataset construction, it is ensured that the time-frequency images in the training and validation sets maintain consistency in signal features and scale distribution to avoid overfitting or underfitting problems caused by uneven data distribution. Simultaneously, the scale and diversity of the dataset fully match the training needs of deep learning models, covering UAV image transmission signal time-frequency images under different environments and signal-to-noise ratios, thus improving the model's generalization ability.

[0054] In some specific implementations, the training of deep learning models is carried out as follows: See Figure 6 The pre-built deep learning model is trained using the constructed dataset. The training process aims to achieve model convergence and meet preset performance metrics. By setting reasonable optimizers and learning rate change strategies, the model parameters are iteratively optimized. At the same time, the model performance is monitored in real time using the validation set to ensure that the model has good detection accuracy and generalization ability.

[0055] In this embodiment, the time-frequency map generated by the UAV signal sample is used as the training data. The labeled time-frequency map is divided into a training set and a validation set according to a set ratio. During the model training and testing process, the optimizer selects stochastic gradient descent, sets an initial learning rate, and adopts a cosine annealing learning rate variation strategy to allow the model to quickly approach the optimal parameter range in the early stage of training and achieve fine parameter adjustment in the later stage of training, thereby improving the convergence accuracy of the model.

[0056] During training, the model's various loss functions, average precision, precision, and recall are monitored in real time. The loss functions include confidence loss, bounding box loss, and class loss. Confidence loss measures the difference between the model's predicted bounding box and the ground truth bounding box. Bounding box loss measures the difference between the model's predicted bounding box and the ground truth bounding box. Class loss is always 0 because there is only one label. Average precision includes the average precision when the intersection-union ratio (IU) threshold is 0.5 and the average precision in the range of IU threshold from 0.5 to 0.95. Precision and recall are used to measure the model's detection accuracy and comprehensiveness.

[0057] As the number of training rounds increases, the values ​​of various loss functions of the model gradually decrease and tend to stabilize. The average precision, accuracy, and recall gradually increase and reach the preset indicators, indicating that the difference between the model's prediction results and the actual results is constantly decreasing. The model can fully learn the time-frequency domain features of the UAV image transmission signal. After the model converges and its performance meets the preset indicators, the trained model weight file is obtained and used for subsequent model inference and detection.

[0058] In some specific implementations, the inference detection of deep learning models is implemented as follows: Model inference is the final step in UAV signal detection. The signal to be analyzed undergoes the same time-frequency transformation process as the training set to obtain the time-frequency map to be detected. This map is then input into the trained deep learning model for inference. The model extracts and fuses features through a multi-scale feature extraction module, backbone network, SE module, and neck network. Finally, the detection head outputs the time-frequency domain location information of the target signal and the corresponding confidence data, and visualizes the detection results.

[0059] See Figure 7 The trained model was applied to the detection of UAV signals in the time-frequency map. The model was able to accurately mark the location of the UAV image transmission signal in the time-frequency map and mark the confidence data at the corresponding location. The detection results show that the model can accurately identify the UAV signal in the time-frequency map, and the confidence data corresponding to each signal is maintained at a high level, indicating that the model has good accuracy in detecting UAV signals.

[0060] See Figure 8 The model was applied to UAV signal detection in low signal-to-noise ratio and strong interference scenarios. In such cases, the target signal features in the time-frequency image are easily masked by interference signals, which places high demands on the model's feature extraction and anti-interference capabilities. Through noise suppression by the SE module and fine-grained feature capture by the multi-scale feature extraction module, the model can still accurately detect the UAV image transmission signal in the time-frequency image, accurately mark the signal position, and output the corresponding confidence data. This shows that the model has good robustness in complex environments and can meet the complex application needs of actual anti-drone scenarios.

[0061] This embodiment realizes deep learning-based UAV signal detection through a complete technical process. From the generation of time-frequency graphs of the original IQ signal to the construction of datasets, the construction and training of deep learning models, and finally the model inference detection, the implementation of each step strictly matches the time-frequency domain characteristics of UAV image transmission signals. At the same time, it is designed with specific considerations for model lightweighting and feature extraction capability improvement.

[0062] The addition of multi-scale feature extraction and SE modules significantly improves the model's feature extraction and anti-interference capabilities, enabling the model to fully capture the multi-scale features of UAV signals and achieve effective signal feature enhancement and noise feature suppression in interference environments. The lightweight design of deep convolutional layers greatly reduces the number of model parameters and computational overhead, allowing the model to be deployed on resource-constrained edge devices and adapting to the deployment requirements of anti-UAV systems.

[0063] Based on actual detection results, the model can accurately detect UAV image transmission signals in time-frequency maps under normal scenarios, strong interference, and low signal-to-noise ratio conditions. It outputs accurate time-frequency domain location information and confidence data. The visualization of the detection results makes them more intuitive, meeting the practical needs of UAV signal detection in anti-UAV systems. The implementation of this detection method solves the problems of weak anti-interference capabilities of traditional algorithms and the complexity and poor robustness of deep learning algorithm models in existing technologies. It provides an efficient, accurate, and easily deployable technical solution for UAV signal detection in anti-UAV systems.

[0064] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A deep learning-based method for detecting unmanned aerial vehicle (UAV) signals, characterized in that, Includes the following steps: S1. Process the IQ signal to generate a time-frequency graph. The processing steps include defining basic parameters, reading IQ data, performing time-frequency transformation, converting to decibel scale, calibrating the frequency axis, and mapping colors to generate the time-frequency graph. S2. Label the UAV image transmission signal in the time-frequency graph generated by processing the IQ signal, integrate the time-frequency graph and the labeling information to construct a dataset, and use the dataset to train the preset deep learning model until the model converges; S3. Perform the same operation on the target signal as on the IQ signal to generate the target time-frequency map, and input the target time-frequency map into the trained deep learning model; Step S2 includes the following sub-steps: S2.

1. By using annotation software, the time-frequency domain position and range of the UAV image transmission signal in the time-frequency diagram generated by processing the IQ signal are clearly defined, and the relative position information of the signal is obtained; S2.

2. Divide the time-frequency graph with the acquired relative position information into a training set and a validation set according to a set ratio, and integrate them to form a dataset for model training and validation; S2.

3. Construct a deep learning model based on the Yolov5s architecture. The deep learning model includes a multi-scale feature extraction module, an SE module, a detection head, and a backbone network. The backbone network uses depthwise separable convolutional layers. The multi-scale feature extraction module contains four parallel convolutional layers with different kernel sizes, which are placed at the front end of the deep learning model, before the backbone network. The tensors output by the four convolutional layers are concatenated along the channel dimension, and the concatenated tensor is passed through a point convolutional layer. The SE module is located at the back end of the backbone network of a deep learning model; S2.

4. Set the optimizer and learning rate variation strategy, train the deep learning model using the partitioned training set, verify the model performance using the partitioned validation set, and continue iterating until the loss function value and performance index of the deep learning model reach the set values, and obtain the trained model weight file. S4. The deep learning model outputs the time-frequency domain location information and confidence data of the target signal to realize the visualization of the detection results.

2. The deep learning-based UAV signal detection method according to claim 1, characterized in that, Step S1 includes the following sub-steps: S1.

1. Define the basic parameters of sampling rate, bandwidth, and number of Fourier transform points, and derive the number of time-domain sampling points and number of frames for a single time-frequency analysis; S1.

2. Read IQ data that meets the defined basic parameter requirements from the binary file in blocks; S1.

3. Perform a short-time Fourier transform on the read IQ data to obtain the time-frequency domain matrix, convert the amplitude spectrum of the time-frequency domain matrix to a decibel scale and introduce a very small positive number to avoid calculation anomalies; S1.

4. The frequency axis is calibrated from relative frequency to absolute frequency to generate a time-frequency plot data matrix, and the spectral data values ​​at the decibel scale are mapped to the corresponding colors to generate a time-frequency plot.

3. The UAV signal detection method based on deep learning according to claim 1, characterized in that, Step S3 includes the following sub-steps: S3.

1. Obtain the target signal to be analyzed, and perform processing on the target signal by defining basic parameters and reading data in blocks; S3.

2. Perform short-time Fourier transform, decibel conversion, frequency calibration, and color mapping operations on the processed target signal to generate a target time-frequency diagram; S3.

3. Adjust the generated target time-frequency map according to the set format to ensure that the parameters of the target time-frequency map are consistent with the time-frequency map of the training set formed by the partitioning; S3.

4. Input the adjusted target time-frequency map into the trained deep learning model and start the feature extraction operation of the deep learning model.

4. The UAV signal detection method based on deep learning according to claim 1, characterized in that, Step S4 includes the following sub-steps: S4.

1. The deep learning model performs signal recognition and inference operations on the input target time-frequency map; S4.

2. Output the time-frequency domain location information and confidence data corresponding to the target signal obtained from the inference operation; S4.

3. Use the output time-frequency domain location information and confidence data to mark the location of the target signal in the target time-frequency map, and synchronously associate the corresponding confidence data; S4.

4. The time-frequency graph of the marked target is displayed in a visual way to achieve an intuitive presentation of the detection results.

5. The deep learning-based UAV signal detection method according to claim 2, characterized in that, In step S1, the Hanning window is used to perform sliding processing on the read IQ data. The sliding step size is matched with the number of frames in the defined basic parameters. The time-domain IQ signal is mapped to a time-frequency domain matrix through Fourier transform to ensure the continuity of time-frequency analysis.

6. The deep learning-based UAV signal detection method according to claim 1, characterized in that, In step S2, the multi-scale feature extraction module includes four parallel convolutional layers. The four convolutional layers use convolutional kernels of different sizes and have the same stride to extract fine-grained features from the input feature map at different levels. The ReLU activation function is used to fuse the concatenated features to achieve data compression.

7. The UAV signal detection method based on deep learning according to claim 1, characterized in that, In step S2, the SE module is set at the back end of the backbone network of the deep learning model. The SE module includes compression and activation operations. The compression operation compresses the spatial dimension of the input feature map into a single-channel descriptor through global average pooling. The channel descriptor represents the global distribution of each channel. The activation operation processes the channel descriptor through two fully connected layers, inserts a ReLU activation function in between, and uses the Sigmoid function to normalize the processed result to generate channel weights. The channel weights are multiplied with the original input feature map channel by channel to achieve feature recalibration, and the output is a recalibrated feature map with the same specifications as the original input feature map.

8. The deep learning-based UAV signal detection method according to claim 1, characterized in that, In step S2, the deep learning model adds a new detection head to the existing detection head, and the parameters of the new detection head are matched with those of the existing detection head. The conventional convolutional layers in the backbone network are replaced with depthwise separable convolutional layers. The depthwise convolutional layers perform convolution operations on each input channel separately, and the output of the depthwise convolutional layers is fused by point convolutional layers, thereby reducing the number of parameters and computational overhead of the deep learning model.

9. A deep learning-based unmanned aerial vehicle (UAV) signal detection system, characterized in that, It includes a time-frequency graph generation module, a data annotation and dataset construction module, a model training module, and a model inference module; The time-frequency graph generation module is used to perform basic parameter definition, block data reading, short-time Fourier transform, decibel conversion, frequency calibration and color mapping on the IQ signal to generate a time-frequency graph that meets the detection requirements. The data annotation and dataset construction module is used to annotate the location of UAV image transmission signals in the time-frequency map generated by annotation software, obtain the relative position information of the signals, divide the time-frequency map with relative position information into training set and validation set according to a set ratio, and integrate them to form a dataset. The model training module is used to build a deep learning model based on the Yolov5s architecture. The deep learning model includes a multi-scale feature extraction module, an SE module, a detection head, and a backbone network. The backbone network uses deep separable convolutional layers. An optimizer and learning rate variation strategy are set. The deep learning model is trained using the integrated training set. The model performance is verified using the integrated validation set. The iteration continues until the loss function value and performance index reach the set values. Finally, the trained model weight file is output. The multi-scale feature extraction module contains four parallel convolutional layers with different kernel sizes, which are placed at the front end of the deep learning model, before the backbone network. The tensors output by the four convolutional layers are concatenated along the channel dimension, and the concatenated tensor is passed through a point convolutional layer. The SE module is placed at the back end of the backbone network of the deep learning model. The model inference module performs the same processing flow as the time-frequency map generation module on the target signal to generate the target time-frequency map. After adjusting the target time-frequency map according to the set format, it inputs the deep learning model with the loaded output model weight file. It receives the position information and confidence data output by the deep learning model, marks the position in the target time-frequency map using the position information and confidence data and associates it with the confidence data. The detection results are displayed through a visualization component. The modules achieve orderly data transmission and interaction through data interfaces.

Images