Unmanned aerial vehicle classification method based on time-frequency characteristics of image transmission signals

Abstract

The invention discloses an unmanned aerial vehicle classification method based on time-frequency characteristics of image transmission signals. The method comprises the following steps: S1, acquiring image transmission signals of an unmanned aerial vehicle in an ISM frequency band with the bandwidth of 25MHz; s2, performing data preprocessing on the image transmission signal of the unmanned aerial vehicle to obtain a time-frequency graph; and S3, inputting the time-frequency graph into the trained OfficientNet, and outputting a final result of classification and identification of the model of the unmanned aerial vehicle. An experiment result shows that the trained OfficientNet neural network has a good classification effect. For unmanned aerial vehicle image transmission signals with the signal-to-noise ratio of-10dB or above, a noisy unmanned aerial vehicle data set is used for training, remarkable classification effect improvement is observed on a test set, especially under the condition of the signal-to-noise ratio of-10dB or above, the classification accuracy reaches 96% or above, and the method has high application prospects in the field of civil unmanned aerial vehicle signal monitoring.

Description

Technical Field

[0001] The present invention belongs to the field of image signal detection, and in particular relates to a method for classifying drones based on time-frequency characteristics of image transmission signals. Background Art

[0002] In recent years, the technical threshold and manufacturing cost of unmanned aerial vehicles (UAVs) have continued to decrease, making UAVs applicable to many purposes, such as toys and games, photography, high-risk operations, remote sensing mapping, meteorological monitoring, agricultural and forestry plant protection, industrial safety, geological exploration, etc. The gradual opening of my country's low-altitude airspace and the positive policy support of the country for the development of the UAV field have enabled my country's civilian UAV industry to develop rapidly.

[0003] While drones bring us convenience, they also pose a threat to personal privacy, national security and airspace management. [2] The frequent incidents of illegal drones invading no-fly zones are mainly due to the small size of drones, which makes it difficult to quickly and accurately monitor and locate them. Therefore, in order to effectively implement airspace supervision, it is of great practical significance to develop third-party equipment to detect drones. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the present invention provides a drone classification method based on the time-frequency characteristics of image transmission signals, which solves the problem of large errors in the prior art in identifying drone image transmission signals in complex environments.

[0005] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is: a method for classifying drones based on the time-frequency characteristics of image transmission signals, comprising the following steps: S1, collects the UAV image transmission signal in the ISM band with a bandwidth of 25MHz; S2, preprocessing the UAV image transmission signal to obtain a time-frequency diagram; S3. Input the time-frequency graph into the trained EfficientNet and output the final result of the classification and recognition of the drone model.

[0006] Further: S2 comprises the following sub-steps: S21. Using a power spectrum estimation method to perform data cropping on the collected UAV image transmission signal, retaining a valid part of the UAV image transmission signal according to bandwidth characteristics; S22, performing time domain windowing processing on the cropped UAV image transmission signal; S23. Use short-time Fourier transform to extract the time domain and frequency domain features of the drone image transmission signal, normalize the amplitude and output the time-frequency diagram through color space mapping.

[0007] Further: In S23, the time domain and frequency domain features are extracted by short-time Fourier transform, and the short-time Fourier transform of the drone image transmission signal is The specific expression is:

[0008] In the formula, For the UAV image transmission signal, is the window function, f is the frequency, t For time, a is the integration variable, j Is an imaginary unit.

[0009] Further: In S3, the hybrid scaling method used by EfficientNet scales the network depth, width and input image resolution, specifically:

[0010] In the formula, d is the network depth, is the width, r is the input image resolution, is a preset adjustment factor used to uniformly scale network depth, width, and input image resolution parameters. To scale the network depth parameter, is the scaling width parameter, The input image resolution parameter for scaling, Represents a constraint.

[0011] Further: In S3, EfficientNet includes a first convolutional layer, a first depth-separable convolutional module, a second depth-separable convolutional module, a third depth-separable convolutional module, a fourth depth-separable convolutional module, a fifth depth-separable convolutional module, a sixth depth-separable convolutional module, a seventh depth-separable convolutional module, a second convolutional layer, a global average pooling layer and a fully connected layer, which are connected in sequence.

[0012] Furthermore, the first to seventh depth-separable convolution modules have the same structure, and all include: The first ordinary convolution, depth convolution, SE attention submodule, second ordinary convolution and Dropout layer are connected in sequence. The output of the Dropout layer is residually connected with the input feature of the first ordinary convolution to obtain the output result of the depth-wise separable convolution module.

[0013] Further: In S3, the method of outputting the final result of classification recognition includes the following steps: S31, extracting features of the time-frequency graph through a first convolutional layer to obtain a first feature graph; S32, inputting the first feature map into the first to seventh depth-separable convolution modules in sequence, capturing the scale and complexity features of the first feature map, and obtaining a second feature map; S33. Input the second feature map into the second convolutional layer, the global average pooling layer and the fully connected layer in sequence to obtain the final result of classification recognition.

[0014] Further: In S32, the first to seventh depth-separable convolution modules process input features in the same way, specifically: S321, increasing the dimension of the input feature through the first ordinary convolution; S322, input the dimension-upgraded input features into the deep convolution and SE attention submodules in sequence, adaptively adjust the weights of the channels in the input features, and obtain an intermediate feature map; S323, inputting the intermediate feature map into the second ordinary convolution, reducing the dimension of the intermediate feature map to the output dimension, and inputting the intermediate feature map after dimension reduction into the Dropout layer to reduce the complexity and calculation amount of the model; S324. Perform a residual connection on the output result of the Dropout layer and the input feature to obtain the output result of the processed input feature.

[0015] Further: In S3, the method for obtaining the trained EfficientNet includes the following steps: A1. Set the parameters of the neural network of the EfficientNet model: The parameters of the EfficientNet model are =1.2, =1.1, =1.15; The resolution of the first convolutional layer is 224x224, the number of channels is 32, and the number of layers is 1; The first depth-wise separable convolutional module has a resolution of 112x112, 16 channels, and 1 layer. The second depth-wise separable convolutional module has a resolution of 112x112, 24 channels, and 2 layers; The third depth-wise separable convolutional module has a resolution of 56x56, 40 channels, and 3 layers; The fourth depth-wise separable convolutional module has a resolution of 28x28, 80 channels, and 3 layers; The fifth depth-wise separable convolutional module has a resolution of 14x14, 112 channels, and 3 layers; The sixth depth-wise separable convolutional module has a resolution of 14x14, 192 channels, and 4 layers; The seventh depth-wise separable convolutional module has a resolution of 7x7, 320 channels, and 1 layer. The second convolutional layer, global average pooling layer, and fully connected layer have a resolution of 7x7, 1280 channels, and 1 layer. A2. Use drone I / Q data as positive samples and Wifi signal I / Q data as negative samples, add -10dB~10dB Gaussian white noise, and form a drone image transmission signal data set according to the time-frequency graph; A3. Divide the drone image transmission signal dataset into training set, test set and validation set in a ratio of 6:2:2. Set the training parameters as follows: initial learning rate of 0.001, weight decay of 0.00001, batch-size of 16, epoch of 5, image resolution of 224×224, and complete EfficientNet training.

[0016] The beneficial effects of the present invention are: (1) The present invention provides a method for classifying drones based on the time-frequency characteristics of image transmission signals. In order to solve the problem of large errors in identifying drone image transmission signals in complex environments caused by the rapid development of drones, such as illegal reconnaissance, airspace violation, and drone collision, the collected drone image transmission signals are converted into time-frequency graphs, and high-quality time-frequency graphs are constructed. EfficientNet is used for training, and significant performance improvements are achieved on the training set and validation set. Finally, experiments were conducted on six types of drones, including the DJI Mavic series and Mini 4 Pro, using the drone image transmission signal dataset constructed in this paper. The recognition and classification accuracy rate reached more than 96%, demonstrating the high application prospects of this method in the field of civil drone signal monitoring.

[0017] (2) The evaluation results on the test set verified the effectiveness of the proposed method under high and low signal-to-noise ratio conditions, especially under the condition of -10dB and above signal-to-noise ratio, the EfficientNet model still maintained a high recognition accuracy. The experimental results show that this method can realize the accurate recognition of UAV image transmission signals in complex environments and has high practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 The present invention is a flow chart of a method for classifying drones based on the time-frequency characteristics of image transmission signals.

[0019] Figure 2 This is a principle block diagram of the drone image transmission signal classification method of the present invention.

[0020] Figure 3 This is a time-frequency diagram of the image transmission signal of the Mavic 3 drone of the present invention.

[0021] Figure 4This is a structural diagram of the depth-separable convolution module of the present invention.

[0022] Figure 5 This is the data preprocessing process of the present invention.

[0023] Figure 6 This is a graph showing changes in the model loss value and accuracy of the present invention.

[0024] Figure 7 It is a thermal diagram of the test results of the model of the present invention. DETAILED DESCRIPTION

[0025] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0026] Embodiment 1: like Figure 1 As shown, in one embodiment of the present invention, a method for classifying drones based on time-frequency characteristics of image transmission signals includes the following steps: S1, collects the UAV image transmission signal in the ISM band with a bandwidth of 25MHz; S2, preprocessing the UAV image transmission signal to obtain a time-frequency diagram; S3. Input the time-frequency graph into the trained EfficientNet and output the final result of the classification and recognition of the drone model.

[0027] In this embodiment, the drone image transmission signal classification method designed by the present invention includes the following steps: first, the drone image transmission signal is collected, then the time domain and frequency domain characteristics are associated and converted into a time-frequency diagram as the input of the network, then, the classification and recognition are performed through EfficientNet, and finally the final result is output. The principle block diagram is as follows: Figure 2 shown.

[0028] In S1, a software defined radio (SDR) or other radio signal acquisition equipment is used to collect the ISM band drone image transmission signal with a bandwidth of 25 MHz.

[0029] The S2 comprises the following sub-steps: S21. Using a power spectrum estimation method to perform data cropping on the collected UAV image transmission signal, retaining a valid part of the UAV image transmission signal according to bandwidth characteristics; S22, performing time domain windowing processing on the cropped UAV image transmission signal; S23. Use short-time Fourier transform to extract the time domain and frequency domain features of the drone image transmission signal, normalize the amplitude and output the time-frequency diagram through color space mapping.

[0030] In S23, the time domain and frequency domain features are extracted by short-time Fourier transform, and the short-time Fourier transform of the drone image transmission signal is The specific expression is:

[0031] In the formula, For the UAV image transmission signal, is the window function, f is the frequency, t For time, a is the integration variable, j Is an imaginary unit.

[0032] In this embodiment, short-time Fourier transform is a signal processing method widely used in time-frequency analysis. Short-time Fourier transform combines information in the time domain and frequency domain, and can represent the characteristics of the signal in both time and frequency dimensions, which is very effective for non-stationary signals. Short-time Fourier transform uses the local stationary characteristics of the signal to perform sliding interception of the signal with a window function of a certain length, and then performs Fourier transform on the sliding intercepted signal. The graphics generated by short-time Fourier transform are formed into pseudo-color time-frequency diagrams through color mapping, Figure 3 The time-frequency diagram of a frame of the image transmission signal of DJI Mavic 3 is shown.

[0033] In S3, EfficientNet uses a hybrid scaling method to scale the network depth, width, and input image resolution, specifically:

[0034] In the formula, d is the network depth, is the width, r is the input image resolution, is a preset adjustment factor used to uniformly scale network depth, width, and input image resolution parameters. To scale the network depth parameter, is the scaling width parameter, The input image resolution parameter for scaling, Represents a constraint.

[0035] In S3, EfficientNet includes a first convolutional layer, a first depth-wise separable convolutional module, a second depth-wise separable convolutional module, a third depth-wise separable convolutional module, a fourth depth-wise separable convolutional module, a fifth depth-wise separable convolutional module, a sixth depth-wise separable convolutional module, a seventh depth-wise separable convolutional module, a second convolutional layer, a global average pooling layer and a fully connected layer, which are connected in sequence.

[0036] like Figure 4 As shown, in this embodiment, the depthwise separable convolution module is an inverted linear bottleneck layer with depthwise separable convolution, and the first to seventh depthwise separable convolution modules have the same structure, and all include: The first ordinary convolution, depth convolution, SE attention submodule, second ordinary convolution and Dropout layer are connected in sequence. The output of the Dropout layer is residually connected with the input feature of the first ordinary convolution to obtain the output result of the depth-wise separable convolution module.

[0037] In S3, the method of outputting the final result of classification recognition includes the following steps: S31, extracting features of the time-frequency graph through a first convolutional layer to obtain a first feature graph; S32, inputting the first feature map into the first to seventh depth-separable convolution modules in sequence, capturing the scale and complexity features of the first feature map, and obtaining a second feature map; S33. Input the second feature map into the second convolutional layer, the global average pooling layer and the fully connected layer in sequence to obtain the final result of classification recognition.

[0038] In S32, the first to seventh depth-separable convolution modules process input features in the same way, specifically: S321, increasing the dimension of the input feature through the first ordinary convolution; S322, input the dimension-upgraded input features into the deep convolution and SE attention submodules in sequence, adaptively adjust the weights of the channels in the input features, and obtain an intermediate feature map; S323, inputting the intermediate feature map into the second ordinary convolution, reducing the dimension of the intermediate feature map to the output dimension, and inputting the intermediate feature map after dimension reduction into the Dropout layer to reduce the complexity and calculation amount of the model; S324. Perform a residual connection on the output result of the Dropout layer and the input feature to obtain the output result of the processed input feature.

[0039] In this embodiment, the principle of the depthwise separable convolution module is: First, the feature is increased in dimension through a 1×1 ordinary convolution, and then through a k×k Depthwise Conv convolution (the specific values ​​of k are 3×3 and 5×5), and then the weights of the channels in the feature map are adaptively adjusted through an SE attention module. Next, a 1×1 ordinary convolution is used to reduce the feature dimension to the output dimension, and finally the Dropout layer is used to reduce the complexity and computation of the model, and then a residual connection is made with the input feature to output the result.

[0040] In S3, the method for obtaining the trained EfficientNet includes the following steps: A1. Set the parameters of the neural network of the EfficientNet model: The parameters of the EfficientNet model are =1.2, =1.1, =1.15; The resolution of the first convolutional layer is 224x224, the number of channels is 32, and the number of layers is 1; The first depth-wise separable convolutional module has a resolution of 112x112, 16 channels, and 1 layer. The second depth-wise separable convolutional module has a resolution of 112x112, 24 channels, and 2 layers; The third depth-wise separable convolutional module has a resolution of 56x56, 40 channels, and 3 layers; The fourth depth-wise separable convolutional module has a resolution of 28x28, 80 channels, and 3 layers; The fifth depth-wise separable convolutional module has a resolution of 14x14, 112 channels, and 3 layers; The sixth depth-wise separable convolutional module has a resolution of 14x14, 192 channels, and 4 layers; The seventh depth-wise separable convolutional module has a resolution of 7x7, 320 channels, and 1 layer. The second convolutional layer, global average pooling layer, and fully connected layer have a resolution of 7x7, 1280 channels, and 1 layer. A2. Use drone I / Q data as positive samples and Wifi signal I / Q data as negative samples, add -10dB~10dB Gaussian white noise, and form a drone image transmission signal data set according to the time-frequency graph; In this embodiment, I / Q data of drones of various manufacturers and models can be collected, and the time-frequency diagram generated by S2 can be used to form a drone image transmission signal data set.

[0041] A3. Divide the drone image transmission signal dataset into training set, test set and validation set in a ratio of 6:2:2. Set the training parameters as follows: initial learning rate of 0.001, weight decay of 0.00001, batch-size of 16, epoch of 5, image resolution of 224×224, and complete EfficientNet training.

[0042] In this embodiment, the optimal parameter values ​​are: =1.2, =1.1, =1.15, the specific parameters of EfficientNet are shown in Table 1: Table 1 EfficientNet network parameters

[0043] As can be seen from Table 1, EfficientNet is mainly composed of 16 depth-separable convolutions, 2 convolution layers, 1 global average pooling layer, and 1 fully connected layer. The first layer is a convolution layer with a convolution kernel size of 3×3, which is used to extract signal features. The 2nd to 8th layers repeatedly stack depth-separable convolution modules to capture features of various scales and complexities. The last layer consists of a 1×1 convolution layer, an average pooling layer, and a fully connected layer. The 1 or 6 after the depth-separable convolution module in Table 1 is the multiplication factor n, indicating that the first 1×1 convolution layer in the depth-separable convolution module will expand the number of channels of the input matrix to n times. k3×3 or k5×5 represents the convolution kernel size used by the depthwise convolution (Depthwise Conv) in the depth-separable convolution module.

[0044] Compared with other deep learning frameworks, EfficientNet is more concise and is mainly built on the deep separable convolutional structural unit to form the core of the network. This simplified structure helps reduce model complexity and reduce computational load.

[0045] Embodiment 2: This embodiment is an experimental case provided for Embodiment 1.

[0046] Step 1: Build a signal acquisition platform: In this embodiment, the present invention uses the software-defined SDR (radio) device USRP B210 to collect raw data of the drone image transmission signal. USRP B210 is a dual-channel, full-duplex, wide-band SDR device with a frequency range from 70 MHz to 6 GHz, which can meet our collection requirements for detecting drone image transmission signals.

[0047] Feiteng Pi is a single-board computer independently developed in my country. It is equipped with a Feiteng series processor. It has powerful computing power and low power consumption characteristics, and is suitable for computing-intensive and power-sensitive application scenarios.

[0048] The present invention builds a hardware system based on the SDR receiving module USRP and Feitengpai, connects the Feitengpai board and USRP through USB3.0, and connects an external screen to monitor the UAV radio signal.

[0049] Step 2: Data collection: Based on the software radio data acquisition platform built in the previous article, the data of the drone image transmission signal is collected. First, the sampling rate and center frequency of the USRP are configured so that it can obtain the radio signal emitted by the drone. According to the characteristics of the drone image transmission signal, the center frequency is set in the ISM band and the bandwidth is set to 25 MHz. Through the driver interface, the IQ signal is read into the computer memory for subsequent processing. The experimental target drones include DJI MAVIC series, DJI Mini4 Pro, DJI M30, Dahua X650, and Daotong LITE. Since WiFi signals and drone image transmission signals also use OFDM modulation and have strong confusion, we also collected WiFi signals as a counterexample category and added them to the data set. The transmission parameters of the above drones are shown in Table 2.

[0050] Table 2 UAV image transmission signal parameters used in this project

[0051] Step 3: Data preprocessing: The collected raw data of the UAV image transmission signal is converted into a time-frequency diagram and color mapped to generate a time-frequency diagram. The specific conversion process is as follows: Figure 5 As shown. First, the data collected by USRP is segmented and the original data of varying lengths from 20ms to 50ms is intercepted. In order to avoid spectrum leakage and obtain a clearer time-frequency diagram, the cropped original data will be windowed in the time domain. In order to enhance the anti-interference ability of the model, Gaussian white noise of -10dB to 10dB is added before the time-frequency diagram conversion. The data after the time-frequency diagram conversion is a two-dimensional matrix, and the input of the neural network algorithm is an image, so it needs to be converted. First, numerical normalization is performed to standardize the two-dimensional matrix to between 0 and 1. Finally, the image of the time-frequency diagram is output through color space mapping.

[0052] Step 4: Construction of drone image transmission data set: The present invention collects I / Q data of several drones shown in Table 2, and collects I / Q data of WiFi routers for comparison, and finally converts all I / Q data into time-frequency diagrams. 1500 data are collected for each drone and WiFi source, and a total of 10500 time-frequency diagrams are obtained. Each time-frequency diagram is scaled to 224×224 and marked into 7 categories (6 drones and Wifi categories) to form a drone image transmission data set. According to the requirements of the deep learning algorithm, the data set is divided into a training set, a validation set, and a test set in a ratio of 6:2:2.

[0053] Step 5: Model training environment configuration, as shown in Table 3: Table 3 Experimental environment

[0054] The initial learning rate of the training is 0.001, the weight decay is 0.00001, the batch-size is 16, the epoch is 5, and the image resolution is 224×224.

[0055] Step 6. Evaluation indicators: In order to objectively evaluate the performance of the network model, the experiment uses precision, recall, and F1 score as evaluation indicators to evaluate the performance of the algorithm. Precision refers to the proportion of positive samples predicted correctly by the algorithm in the total positive samples, Recall refers to the proportion of positive samples predicted correctly in the total samples, and F1 score (F1-Score) is a statistical method that combines precision and recall. The specific calculation formula is as follows:

[0056]

[0057]

[0058] In the formula, TP represents the number of positive samples predicted correctly by the model, FP represents the number of positive samples predicted incorrectly by the model, and FN represents the number of negative samples predicted incorrectly by the model. The F1 score is Set to 1, combining the results of precision and recall. The higher the value, the better the model effect.

[0059] Step 7: Model training: The EfficientNet model is trained using the constructed drone image transmission signal dataset. Figure 6The figure shows the changes in the loss value and the recognition and classification accuracy of the training set and validation set during the network model training process. It can be seen that as the training progresses, the overall trend of the loss value of the training set is decreasing. After 5 cycles of training, the loss value dropped from 1.2 to 0.25; at the same time, the recognition and classification accuracy increased from 55% to 91%. The loss value of the validation set dropped from 0.5 to 0.087; at the same time, the recognition and classification accuracy increased from 83% to 97%.

[0060] Step 8: Performance training: 1) Test set performance test: The recognition effect of this algorithm is verified on the test set data. There are 300 time-frequency diagrams of each radiation source in the test set, and a total of 2100 time-frequency diagrams. Through the reasoning of the model, the heat map of each drone model recognition in the test set is as follows: Figure 7 As shown: From the test results of the above model, the recognition accuracy of various models on the test set exceeded 91%. The F1 score on the test set was 0.955, the precision was 97%, and the recall rate was 95.1%. The overall recognition and classification effect was satisfactory.

[0061] 2) Low signal-to-noise ratio test: When the drone is far away from the monitoring equipment, the signal-to-noise ratio of the image transmission signal is low, and the recognition difficulty is high. In order to verify the recognition effect of the system on drones under low signal-to-noise ratio conditions, we added Gaussian white noise with signal-to-noise ratios of -15dB, -10dB, and -5dB to the drone image transmission signal, and sent the drone signal mixed with white noise to the model for testing. The recognition results are shown in Table 4. Table 4 shows that when the signal-to-noise ratio is -10dB or above, the model's recognition ability for drone image transmission signals does not change significantly, and the F1 score remains above 98%. When the signal-to-noise ratio drops to -15dB, the recognition rate of several DJI drones drops significantly, and the recognition rate of other drones does not change significantly. This shows that the network model trained using the time-frequency graph of drone signals can recognize various types of drones with high accuracy under low signal-to-noise ratio conditions.

[0062] Table 4 Recognition results

[0063] In summary, this paper proposes a method for detecting and identifying UAV image transmission signals based on software defined radio and deep learning. In terms of signal acquisition, a hardware platform with USRP B210 and Feitengpai as the core was built, and a variety of UAV image transmission signals and WiFi signals were successfully collected. In terms of data preprocessing, a high-quality data set was constructed through the time-frequency graph conversion method. In terms of model training, the EfficientNet model was used, and significant performance improvements were achieved on the training set and validation set. The evaluation results on the test set verified the effectiveness of the proposed method under high and low signal-to-noise ratio conditions, especially under the signal-to-noise ratio of -10dB and above, the model maintained a high recognition accuracy. The experimental results show that this method can realize accurate recognition of UAV image transmission signals in complex environments and has high practical application value.

[0064] The beneficial effects of the present invention are as follows: the present invention provides a method for classifying drones based on the time-frequency characteristics of image transmission signals. In order to solve the problem of large errors in identifying drone image transmission signals in complex environments caused by the rapid development of drones, the present invention converts the collected drone image transmission signals into time-frequency graphs, constructs high-quality time-frequency graphs, uses EfficientNet for training, and achieves significant performance improvements on training sets and validation sets. Finally, experiments were conducted on six types of drones, including the DJI Mavic series and Mini 4 Pro, using the drone image transmission signal dataset constructed in this paper. The recognition and classification accuracy reached more than 96%, demonstrating the high application prospects of this method in the field of civilian drone signal monitoring.

[0065] In the description of the present invention, it is necessary to understand that the orientation or positional relationship indicated by the terms "center", "thickness", "upper", "lower", "horizontal", "top", "bottom", "inner", "outer", "radial", etc. is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. In addition, the terms "first", "second", and "third" are used only for descriptive purposes, and cannot be understood as indicating or implying the relative importance or the number of implicitly specified technical features. Therefore, the features defined by "first", "second", and "third" may explicitly or implicitly include one or more of the features.

Claims

1. A drone classification method based on the time-frequency characteristics of image transmission signals, characterized in that: The following steps are involved: S1, collects the UAV image transmission signal in the ISM band with a bandwidth of 25MHz; S2, preprocessing the UAV image transmission signal to obtain a time-frequency diagram; S3. Input the time-frequency graph into the trained EfficientNet and output the final result of the classification and recognition of the drone model.

2. The method for classifying drones based on the time-frequency characteristics of image transmission signals according to claim 1, characterized in that: The S2 comprises the following sub-steps: S21. Using a power spectrum estimation method to perform data cropping on the collected UAV image transmission signal, retaining a valid part of the UAV image transmission signal according to bandwidth characteristics; S22, performing time domain windowing processing on the cropped UAV image transmission signal; S23. Use short-time Fourier transform to extract the time domain and frequency domain features of the drone image transmission signal, normalize the amplitude and output the time-frequency diagram through color space mapping.

3. The method for classifying drones based on the time-frequency characteristics of image transmission signals according to claim 2 is characterized in that: In S23, the time domain and frequency domain features are extracted by short-time Fourier transform, and the short-time Fourier transform of the drone image transmission signal is The specific expression is: In the formula, For the UAV image transmission signal, is the window function, f is the frequency, t For time, a is the integration variable, j Is an imaginary unit.

4. The method for classifying drones based on the time-frequency characteristics of image transmission signals according to claim 1, characterized in that: In S3, EfficientNet uses a hybrid scaling method to scale the network depth, width, and input image resolution, specifically: In the formula, d is the network depth, is the width, r is the input image resolution, is a preset adjustment factor used to uniformly scale network depth, width, and input image resolution parameters. To scale the network depth parameter, is the scaling width parameter, The input image resolution parameter for scaling, Represents a constraint.

5. The method for classifying drones based on the time-frequency characteristics of image transmission signals according to claim 4 is characterized in that: In the S3, EfficientNet includes a first convolutional layer, a first depth-separable convolutional module, a second depth-separable convolutional module, a third depth-separable convolutional module, a fourth depth-separable convolutional module, a fifth depth-separable convolutional module, a sixth depth-separable convolutional module, a seventh depth-separable convolutional module, a second convolutional layer, a global average pooling layer and a fully connected layer, which are connected in sequence.

6. The method for classifying drones based on the time-frequency characteristics of image transmission signals according to claim 5, characterized in that: The first to seventh depth-separable convolution modules have the same structure and include: The first ordinary convolution, depth convolution, SE attention submodule, second ordinary convolution and Dropout layer are connected in sequence. The output of the Dropout layer is residually connected with the input feature of the first ordinary convolution to obtain the output result of the depth-wise separable convolution module.

7. The method for classifying drones based on the time-frequency characteristics of image transmission signals according to claim 6, characterized in that: In S3, the method of outputting the final result of classification recognition includes the following steps: S31, extracting features of the time-frequency graph through a first convolutional layer to obtain a first feature graph; S32, inputting the first feature map into the first to seventh depth-separable convolution modules in sequence, capturing the scale and complexity features of the first feature map, and obtaining a second feature map; S33. Input the second feature map into the second convolutional layer, the global average pooling layer and the fully connected layer in sequence to obtain the final result of classification recognition.

8. The method for classifying drones based on the time-frequency characteristics of image transmission signals according to claim 7, characterized in that: In S32, the first to seventh depth-separable convolution modules process input features in the same way, specifically: S321, increasing the dimension of the input feature through the first ordinary convolution; S322, input the dimension-upgraded input features into the deep convolution and SE attention submodules in sequence, adaptively adjust the weights of the channels in the input features, and obtain an intermediate feature map; S323, inputting the intermediate feature map into the second ordinary convolution, reducing the dimension of the intermediate feature map to the output dimension, and inputting the intermediate feature map after dimension reduction into the Dropout layer to reduce the complexity and calculation amount of the model; S324. Perform a residual connection on the output result of the Dropout layer and the input feature to obtain the output result of the processed input feature.

9. The method for classifying drones based on the time-frequency characteristics of image transmission signals according to claim 5, characterized in that: In S3, the method for obtaining the trained EfficientNet includes the following steps: A1. Set the parameters of the neural network of the EfficientNet model: The parameters of the EfficientNet model are =1.2, =1.1, =1.15; The resolution of the first convolutional layer is 224x224, the number of channels is 32, and the number of layers is 1; The first depth-wise separable convolutional module has a resolution of 112x112, 16 channels, and 1 layer. The second depth-wise separable convolutional module has a resolution of 112x112, 24 channels, and 2 layers; The third depth-wise separable convolutional module has a resolution of 56x56, 40 channels, and 3 layers; The fourth depth-wise separable convolutional module has a resolution of 28x28, 80 channels, and 3 layers; The fifth depth-wise separable convolutional module has a resolution of 14x14, 112 channels, and 3 layers; The sixth depth-wise separable convolutional module has a resolution of 14x14, 192 channels, and 4 layers; The seventh depth-wise separable convolutional module has a resolution of 7x7, 320 channels, and 1 layer. The second convolutional layer, global average pooling layer, and fully connected layer have a resolution of 7x7, 1280 channels, and 1 layer. A2. Use drone I / Q data as positive samples and Wifi signal I / Q data as negative samples, add -10dB~10dB Gaussian white noise, and form a drone image transmission signal data set according to the time-frequency graph; A3. Divide the drone image transmission signal dataset into training set, test set and validation set in a ratio of 6:2:

2. Set the training parameters as follows: initial learning rate of 0.001, weight decay of 0.00001, batch-size of 16, epoch of 5, image resolution of 224×224, and complete EfficientNet training.

Images