Cross-domain unmanned aerial vehicle identification method based on deep learning

Abstract

The invention provides a cross-domain unmanned aerial vehicle identification method based on deep learning, and the method mainly comprises the steps: carrying out the preprocessing of unmanned aerial vehicle radio frequency data collected at different times, and generating a time-frequency atlas; decomposing the original time-frequency spectrum into a plurality of sub-spectrums by using a multi-resolution analysis technology; applying a lightweight and multi-scale neural network architecture, and extracting radio frequency features in sub-spectrums; a feature enhancement method is adopted to enrich and optimize the extracted features; and finally carrying out label classification. According to the method, data redundancy in a complex spectrum is effectively reduced, and accurate and reliable unmanned aerial vehicle identification is realized. Through combination of multi-scale analysis, a lightweight network architecture and a feature enhancement technology, the robustness and generalization ability of the model are significantly improved, the method has wide application potential, and an efficient and reliable technical scheme is provided for unmanned aerial vehicle identification in a complex electromagnetic environment in the fields of unmanned aerial vehicle management, smart city monitoring and the like.

Description

Technical Field

[0001] The present invention relates to a cross-domain UAV identification technology in a complex electromagnetic environment, and a cross-domain UAV identification method based on deep learning. Background Art

[0002] In the modern era of digital interconnection, drones are widely deployed in multiple emerging markets due to their flexibility, and their applications range from cinematography, agricultural monitoring, remote surveillance to entertainment activities. However, the popularity of drones also brings a series of safety risks, including the threat to civil aviation safety when flying at inappropriate locations and times, and the violation of privacy and security by unauthorized airspace activities. Therefore, it is particularly important to strengthen the effective monitoring and management of drones.

[0003] Traditional drone identification technologies, such as visual or sound recognition, face many limitations in complex environments, such as obstructed sight lines, poor lighting conditions, or background noise. In contrast, RF fingerprinting technology is not affected by these external factors and can work stably under various environmental conditions. RFID methods use the uniqueness of RF signals emitted during the wireless transmission of drones. These signals form unique "fingerprints" due to slight differences in components during the manufacturing process. Even small differences are enough to distinguish different drones. By accurately analyzing these unique RF signals, it is possible to not only identify the drone in flight, but also track down the specific operator or owner.

[0004] RF fingerprint recognition technology improves the accuracy and processing speed of device identification by adopting deep learning models such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), or transformers. These models are able to learn complex RF signal features from a large amount of training data and effectively generalize on unseen data. Nevertheless, RF fingerprint recognition technology based on deep learning also has obvious limitations, the most important of which is its high demand for computing and storage resources. Deep learning models usually require high-performance processors and graphics processing units (GPUs) to handle large-scale data and complex calculations, and also require a large amount of storage space to save training data and model parameters. In addition, the acquisition and preprocessing process of RF fingerprints is complicated, and the high dependence on data quality and acquisition environment may limit its application in dynamic or changing environments. In addition, the generalization ability of the model is also one of the challenges, especially when the device is aging or the hardware is replaced, the model may need to be retrained or adjusted to maintain efficient recognition results. These challenges need to be further addressed and optimized in future research to promote the widespread application of RF fingerprint recognition technology based on deep learning.

[0005] Many variables in the acquisition of drone RF signals, such as acquisition time, distance between the receiver and the drone, and type of receiver used, can lead to significant domain shifts between training data and test data. That is, even with advanced methods such as deep learning, models often find it difficult to maintain good recognition performance when faced with this domain shift problem. To address this challenge, researchers have proposed a series of cross-domain adaptive methods. First, adversarial training mechanisms can be used to train a domain simultaneously on the basis of the main task, so that the feature extraction network can learn more generalizable representations. Another strategy is to perform unsupervised domain alignment, which enables the model to better adapt to new data environments by minimizing the distribution differences between the training domain and the test domain. In addition, methods that combine transfer learning and meta-learning also show good cross-domain generalization capabilities and can quickly adapt to new deployment environments using limited target domain data.

[0006] Lightweight multi-scale modules play an important role in terms of complexity and computational complexity. Traditional residual blocks have large number of parameters and need to calculate multiple convolutions of different scales in parallel, which is computationally intensive. Lightweight residual blocks can use multi-scale convolution mode to improve computational efficiency by using convolution kernels of different sizes for parallel calculation. Depthwise separable convolution is used to replace standard convolution, and the standard convolution layer is decomposed into depthwise convolution and pointwise convolution. This decomposition further reduces computational complexity and model parameters. In addition, by dynamically adjusting the number of output channels of each branch of the module to balance model complexity and performance, it can be flexibly configured according to the needs of different tasks. In general, lightweight residual modules provide a feasible solution for deployment in resource-constrained environments such as mobile devices and edge devices by significantly reducing the number of parameters and computational complexity, and fully balance the complexity and performance of the model.

[0007] The multi-resolution analysis method can effectively reduce the number of parameters and computational complexity of deep learning models by performing one-dimensional wavelet transforms in both row and column directions. This process first performs a one-dimensional wavelet transform on the rows of the input image to obtain low-frequency subbands and high-frequency subbands; then a one-dimensional wavelet transform is performed on the columns of these subbands to further generate subbands of four different frequency bands. This decomposition mechanism reduces the resolution of the spectrum to half of the original in both the horizontal and vertical directions, thereby greatly reducing the number of parameters and floating-point calculations required for the model. Compared with the direct use of standard two-dimensional convolution, 2DDWT not only has lower complexity, but also can better capture the multi-scale features of the image, providing an effective means for building efficient deep learning models. This feature extraction method based on wavelet analysis has been widely used in many computer vision tasks and provides important support for the deployment of deep neural networks in resource-constrained scenarios.

[0008] The current deep learning models have an increasing demand for computing and storage resources, which poses a severe challenge to their deployment on mobile terminals and edge devices. These resource-constrained devices usually have limited computing power and small memory capacity, and cannot support complex deep neural networks. At the same time, training large-scale data sets will also cause extremely high computing and storage overheads, resulting in long training times and even overfitting problems. These problems highlight the importance of model compression and lightweighting. Through various technical means, such as pruning, quantization, knowledge distillation, etc., the model parameters and computational complexity can be significantly reduced, making it more suitable for deployment on mobile terminals and edge devices. At the same time, the rational use of methods such as data enhancement and transfer learning can alleviate the problem of redundant information caused by large amounts of data, improve the generalization ability of the model, and reduce the risk of overfitting. Summary of the invention

[0009] Purpose of the invention: The purpose of the present invention is to provide a cross-domain UAV identification method based on deep learning. By utilizing the existing theoretical and technical methods in deep learning, a multi-resolution analysis method is provided to reduce the input spectrum information in view of the high complexity and large computational complexity of the deep neural network model. A feature enhancement method is provided to improve the robustness and generalization of the model in cross-domain UAV RF fingerprint identification in view of the problem of domain offset between training data and test data.

[0010] In order to solve the above problems, the present invention discloses a cross-domain drone identification method based on deep learning, which is characterized by comprising the following steps:

[0011] Step 1: By preprocessing the drone RF data collected from different time periods, the time-frequency spectrum of the drone RF signal can be generated. This preprocessing step mainly includes two key operations: first, normalize the original signal to eliminate the amplitude difference caused by different acquisition conditions; then use short-time Fourier transform to convert the time domain signal into time-frequency domain representation to highlight the unique spectrum characteristics of drone RF.

[0012] Step 2: Using the multi-resolution analysis method, the original spectrum can be decomposed hierarchically. Specifically, the spectrum is decomposed and downsampled from the frequency axis (row direction) and the time axis (column direction). Decomposition can decompose the spectrum into multiple sub-bands with different frequency resolutions.

[0013] Step 3: A lightweight multi-scale neural network architecture is used. This architecture can capture key information in the spectrogram at different time-frequency resolution levels, thereby providing richer and more effective feature representation for subsequent pattern recognition tasks.

[0014] Step 4: Feature enhancement technology is used to expand the features extracted by the multi-scale neural network in step 3, and the key information of the features is analyzed and optimized to enhance the expressiveness of the features;

[0015] Step 5: Use the label classification network to classify drone labels, so as to achieve accurate identification of different drones.

[0016] Furthermore, the preprocessing of the drone radio frequency data collected at different time periods described in step 1 includes normalization and short-time Fourier transform, specifically including:

[0017] Step 1-1, the time domain data set is the time-frequency data set:

[0018]

[0019] where y i is the One-hot encoding of the corresponding label. RMS is used to normalize the input I / Q data, and the I / Q data is divided by the input sample root mean square, so that the signal amplitude is between [-1,1]. This processing method can eliminate the difference in input signal amplitude, make feature data of different scales comparable, improve numerical stability, speed up model convergence, and facilitate subsequent signal analysis and processing. The RMS normalization calculation process is as follows:

[0020]

[0021] Among them, x i is a complex value normalized by RMS, I i +jQ i is the input I / Q data, is the square of the input I / Q data amplitude, N is the number of samples, where I i and Q i Represent the real and imaginary parts of the original time domain signal at the i-th moment, j is the imaginary unit, I n and Q n Represent the real and imaginary parts of the original time domain signal at the nth moment respectively.

[0022] Step 1-2, use short-time Fourier transform (STFT) to convert the normalized I / Q data into a spectrum diagram:

[0023]

[0024] Among them, X iis the time-frequency diagram after STFT conversion, x(n) is the value of the normalized I / Q data at time point n, ω(n-mR) is the Hamming window function, which divides the continuous signal into a series of short-time windows for local Fourier transform. R represents the number of overlapping samples between adjacent windows, w represents the window function, and m is the index of the window. N is the number of normalized I / Q samples, is the discrete angular frequency, describing the frequency component in the frequency domain, and k is the index of the frequency bin, with the value of k = 0, 1, 2, ..., N-1. The original (6000, 2)-dimensional sample signal is converted into a (102, 45)-dimensional time-frequency image using Fourier transform. A data set of 2982 time-frequency images of 7 identical drones is generated from the time-frequency image. The short-time Fourier transform performs short-time windowing and Fourier transform on the continuous-time signal to obtain the time-frequency domain representation of the signal, which is convenient for analyzing the spectral characteristics of the time-varying signal.

[0025] Furthermore, the original spectrum is decomposed by the multi-resolution analysis method described in step 2, and the original spectrum is decomposed and downsampled in two steps from the row direction and the column direction in sequence, and the resolution of the original spectrum is reduced from two directions at the same time, specifically including:

[0026] Step 2-1, perform a one-dimensional discrete wavelet transform on each row of the image. This process usually involves two key steps: first, a decomposition operation to separate the data into high-frequency and low-frequency components; second, a downsampling operation to reduce the number of data points to simplify the information, thereby optimizing subsequent processing efficiency and data storage. This method effectively extracts and compresses key features from the original image.

[0027] Decomposition: By using the Haar wavelet basis function, we can effectively decompose the row data of the image into two main parts: high-frequency part (detail information) and low-frequency part (approximate information). The high-frequency part mainly contains the detailed information of the image, such as edges and texture changes, which are fast-changing features in the image; while the low-frequency part represents the approximate information of the image, that is, the slowly changing or flat areas in the image, reflecting the basic structure of the image. Decomposing the frequency helps to analyze and process the features of the image. The decomposition operation is as follows:

[0028]

[0029] Among them, X(·) is the time-frequency diagram after short-time Fourier transform, h[n] is the low-pass filter (LPF), g[n] is the high-pass filter (HPF), A[k] is the decomposed low-frequency information, and D[k] is the decomposed high-frequency detail information.

[0030] Downsampling: In order to effectively reduce the amount of data and improve processing efficiency, downsampling is usually performed after the signal decomposition process. Specifically, one sample point is selected for retention for every two samples. This can not only reduce the storage requirements of the data, but also reduce the computational burden in subsequent processing steps, while retaining the important features of the signal to a certain extent.

[0031] Step 2-2, after completing the one-dimensional discrete wavelet transform of all rows in step 2-1, the same one-dimensional discrete wavelet transform is performed on each column. This method of transforming rows first and then columns allows us to decompose the data in detail from two dimensions. Through this combination of row and column transformations, the original time-frequency map is effectively decomposed into four different sub-bands: each sub-band captures frequency and time domain information in different directions, and these sub-bands include low-frequency-low-frequency sub-band (LL (Low-Low Sub-band)), which contains the main structural information of the image, low-frequency-high-frequency sub-band (LH (Low-High Sub-band)), which contains horizontal detail information of the image, such as horizontal edges, high-frequency-low-frequency sub-band (HL (High-Low Sub-band)), which contains vertical detail information of the image, such as vertical edges, and high-frequency-high-frequency sub-band (HH (High-High Sub-band)), which contains diagonal detail information of the image. Such decomposition helps us understand and process different characteristics of the signal from multiple perspectives, improving the accuracy and efficiency of data analysis. These four sub-bands each carry different aspects of the signal information, among which the LL sub-band contains most of the energy and main information, while the other three sub-bands contain more details and edge information.

[0032] Through the multi-resolution analysis method, we effectively decomposed the original time-frequency diagram, reducing its resolution by half in both the row and column directions. Specifically, the original time-frequency diagram dimension is (102, 45). Through this process, the diagram is carefully decomposed into four small sub-spectra, and the dimension of each sub-spectra is changed to (51, 23). This dimensionality reduction not only reduces the burden of data processing, but also makes storage and further analysis more efficient. Each sub-spectrum contains a part of the information in the original time-frequency diagram, which is subdivided into four different frequency and direction combinations, allowing more accurate analysis and processing of local features of the signal. This method is very useful in image compression, signal processing, and other applications that require efficient data decomposition. It can significantly reduce the complexity of the data while maintaining key information.

[0033] Furthermore, in step 3, a lightweight multi-scale neural network architecture is used to extract the RF features of the decomposed spectrum graph and capture the key information in the spectrum graph, specifically including:

[0034] A neural network based on a lightweight multi-scale Inception-Net architecture is used to extract the features of the decomposed spectrum. The network structure captures features at various levels through convolution kernels of different sizes, effectively enhancing the model's ability to process details and global information. This method allows the network to reduce parameters and computation while maintaining high sensitivity to key features, greatly improving the accuracy and efficiency of drone RFID. Feature extraction process:

[0035]

[0036] Among them, X in is the decomposed spectrum obtained in step 2-2, f C onv is the convolution operation, It is a stack of two convolution operations. is a stacking operation of two residual blocks, ν i are the extracted features.

[0037] Furthermore, the feature enhancement method described in step 4 is used to improve the feature effect extracted by the neural network. This method not only enhances the expressiveness of the features extracted from the data, but also significantly improves the expressiveness and accuracy of the features by deeply analyzing and optimizing the key information in these features. This enhancement method enables the network to more effectively recognize and process complex data patterns, thereby improving the overall model performance, including:

[0038] Step 4-1, first calculate the first-order moment and second-order moment of the feature:

[0039]

[0040] Among them, z c The features extracted by the multi-scale network in step 3 have the dimension (B, C, H, W), where B is the number of samples in each batch, C is the number of channels, H and W are the height and width of the feature map, respectively, and h and w are the indices of the h-th row and w-th column in the feature map, respectively. is the first-order moment of the feature map pixel calculated channel by channel, s c It is the second-order moment of the feature map pixel calculated channel by channel.

[0041] In step 4-2, the first-order moments and second-order moments of different eigenvalues ​​are linearly combined by using the convex combination method. The convex combination method can comprehensively consider multiple statistical characteristics and optimize the overall feature expression effect by accurately adjusting the contribution ratio of each eigenvalue. In this way, complex changes in the data can be more effectively captured and described, thereby providing richer and more accurate information input for subsequent analysis and model training. This convex combination method not only enhances the model's sensitivity to subtle differences in data, but also improves the flexibility and adaptability of feature processing, specifically:

[0042] v=γs k +(1-γ)s m

[0043]

[0044] The linear combination of the means and standard deviations of two different features generates a mixed mean and standard deviation, where s k and m They are the features z k and z m The second moment of and They are the features z k and z m In addition, γ ~ beta (α, α), α is a hyperparameter and in general, α ∈ [0.1, 0.3, 0.5, 0.7, 0.9], γ ∈ [0, 1].

[0045] In step 4-3, based on the calculation of the mixed first-order moment and second-order moment in step 4-2, adaptive normalization technology is used to combine other features with these statistics. In this way, we generate new features that not only inherit the basic properties of the original data, but also increase the expressiveness and discrimination of the data by combining with the mixed mean and standard deviation. The adaptive normalization process dynamically adjusts and standardizes each feature, so that the relative importance and contribution of the features are optimized. This comprehensive and innovative feature processing method provides the model with more refined and information-rich input, which helps to improve the accuracy and efficiency of the model in recognizing complex patterns. The specific feature generation is as follows:

[0046]

[0047] Among them, v is the scaling factor used to adjust the scale of the normalized values, and μ is the offset used to adjust the mean of the normalized values.

[0048] Furthermore, the label classification network described in step 5 is used to classify the drone labels, and then the network is continuously trained and optimized using the back-propagation algorithm. In this way, the network is able to gradually adjust its internal parameters to reduce the prediction error during the classification process. This process involves calculating the loss function in each iteration, and then adjusting the network weights based on the loss results, so that the model is more accurate in future predictions. The model is trained to distinguish and classify up to 7 different types of drones. This function is extremely critical to improving the application flexibility and operational efficiency of the system, allowing the model to effectively identify and respond to different drone characteristics in various environments. In addition, the accurate drone identification capability also greatly enhances the safety and reliability of the drone monitoring system, providing strong technical support for drone traffic management and airspace safety.

[0049] The purpose of model training is to construct a mapping function Minimize the overall loss. In practical applications, since only some samples are known, the true data distribution is difficult to calculate accurately. Therefore, we use the empirical distribution based on the training data set to approximate the true sample distribution. Convert the optimization problem to minimize the expected loss on the training set. The specific operations are as follows:

[0050]

[0051] Among them, L{f c [f d (f g (X))],y} is the difference between the model prediction and the true label, L ce represents the cross entropy loss function, N is the number of samples, D t is the input sample set. stf(X i ) is the abbreviation of subject to, which means "under the following conditions". g 、f d and f c They are feature extraction operation, data enhancement and label classification operation. Represents the dataset D t Random sampling in .

[0052] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0053] 1. The present invention utilizes two-dimensional discrete wavelet transform to decompose the time-frequency graph after short-time Fourier transform, and effectively subdivides the time-frequency graph into four sub-spectrum graphs. The resolution of the sub-spectrum graph in both time and frequency dimensions is half of that of the original time-frequency graph. This decomposition not only significantly reduces the amount of input information, but also reduces the data complexity of neural network processing, thereby optimizing the computational efficiency of the network. Through this method, it is possible to reduce the computational burden and increase the processing speed while maintaining important information, which is particularly important for real-time signal processing and large-scale data analysis. In addition, the application of two-dimensional discrete wavelet transform also enhances the feature extraction capability of the signal, making the model more accurate and efficient when processing various signals.

[0054] 2. In order to improve the efficiency and robustness of the model, we used a lightweight multi-scale neural network to construct a feature extraction network. The network uses sub-spectral graphs obtained through two-dimensional discrete wavelet transform for model training, which significantly reduces the amount of calculation, by about 70%. In addition, after the feature extraction is completed, we further use feature enhancement technology to generate new features, which not only enhances the generalization ability of the model, but also significantly improves the performance of the model in various test environments. Specifically, after using feature enhancement, the average performance of Experiment 1 reached 90.85%, while the average performance of Experiment 2 was improved to 96.68%. This method effectively combines the computational efficiency of lightweight networks and the data expressiveness of feature enhancement, making the model more accurate and stable when processing complex data. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a system flow diagram of a cross-domain drone identification method based on deep learning provided by an embodiment of the present invention.

[0056] Figure 2 It is a spectrum decomposition diagram based on two-dimensional discrete wavelet transform provided by an embodiment of the present invention.

[0057] Figure 3 It is an overall framework diagram of the model provided by an embodiment of the present invention.

[0058] Figure 4 It is a comparison diagram of simulation parameters provided by an embodiment of the present invention.

[0059] Figure 5 It is a comparison chart of simulation calculation amount provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0060] The present invention aims to provide a cross-domain UAV identification method based on deep learning. This method comprehensively utilizes the existing theories and technologies in the field of deep learning. In view of the high complexity and large computational complexity of traditional deep neural network models, a multi-resolution analysis method is introduced to reduce the complexity of input spectrum information. In addition, considering the possible domain offset problem between training data and test data, the present invention also adopts a feature enhancement technology to enhance the robustness and generalization ability of the model when processing cross-domain UAV radio frequency fingerprint identification. Through these innovative methods, the present invention not only improves the accuracy of UAV identification, but also greatly improves the scope of application and practicality of the model, so that it can effectively identify a variety of UAVs in different application environments, and provides strong technical support for the safety supervision and management of UAVs.

[0061] Technical solution: A method for extracting electromagnetic environment features of a UAV system based on deep learning, characterized by comprising the following steps:

[0062] Step 1: By preprocessing the drone RF data collected from different time periods, the time-frequency spectrum of the drone RF signal can be generated. This preprocessing step mainly includes two key operations: first, normalize the original signal to eliminate the amplitude difference caused by different acquisition conditions; then use short-time Fourier transform to convert the time domain signal into time-frequency domain representation to highlight the unique spectrum characteristics of drone RF.

[0063] Step 2: Using the multi-resolution analysis method, the original spectrum can be decomposed in layers. Specifically, the spectrum is decomposed and downsampled from the frequency axis (row direction) and the time axis (column direction). Decomposition can decompose the spectrum into multiple sub-bands with different frequency resolutions.

[0064] Step 3: A lightweight multi-scale neural network architecture is used. This architecture can capture key information in the spectrogram at different time-frequency resolution levels, thereby providing richer and more effective feature representation for subsequent pattern recognition tasks.

[0065] Step 4: Feature enhancement technology is used to expand the features extracted by the multi-scale neural network in step 3, and the key information of the features is analyzed and optimized to enhance the expressiveness of the features;

[0066] Step 5: Use the label classification network to classify drone labels, so as to achieve accurate identification of different drones.

[0067] Further, the preprocessing of the drone radio frequency data collected in different time periods described in step 1 is specifically as follows: the time-frequency diagram of the drone radio frequency signal is generated by the preprocessing method described in step 1, specifically including:

[0068] Step 1-1, the time domain data set is the time-frequency data set:

[0069]

[0070] where y i is the One-hot encoding of the corresponding label. RMS is used to normalize the input I / Q data, and the I / Q data is divided by the input sample root mean square, so that the signal amplitude is between [-1,1]. This processing method can eliminate the difference in input signal amplitude, make feature data of different scales comparable, improve numerical stability, speed up model convergence, and facilitate subsequent signal analysis and processing. The RMS normalization calculation process is as follows:

[0071]

[0072] Among them, x i is a complex value normalized by RMS, I i +jQ i is the input I / Q data, is the square of the input I / Q data amplitude, N is the number of samples, where I i and Q i Represent the real and imaginary parts of the original time domain signal at the i-th moment, j is the imaginary unit, I n and Q n Represent the real and imaginary parts of the original time domain signal at the nth moment respectively.

[0073] Step 1-2, use short-time Fourier transform (STFT) to convert the normalized I / Q data into a spectrum diagram:

[0074]

[0075] Among them, X i is the time-frequency diagram after STFT conversion, x(n) is the value of the normalized I / Q data at time point n, ω(n-mR) is the Hamming window function, which divides the continuous signal into a series of short-time windows for local Fourier transform. R represents the number of overlapping samples between adjacent windows, ω represents the window function, and m is the index of the window. N is the number of normalized I / Q samples, is the discrete angular frequency, describing the frequency component in the frequency domain, and k is the index of the frequency bin, with the value of k = 0, 1, 2, ..., N-1. The original (6000, 2)-dimensional sample signal is converted into a (102, 45)-dimensional time-frequency image using Fourier transform. A data set of 2982 time-frequency images of 7 identical drones is generated from the time-frequency image. The short-time Fourier transform performs short-time windowing and Fourier transform on the continuous-time signal to obtain the time-frequency domain representation of the signal, which is convenient for analyzing the spectral characteristics of the time-varying signal.

[0076] Furthermore, the original spectrum is decomposed by the multi-resolution analysis method described in step 2, and the original spectrum is decomposed and downsampled in two steps from the row direction and the column direction in sequence, and the resolution of the original spectrum is reduced from two directions at the same time, specifically including:

[0077] Step 2-1, perform a one-dimensional discrete wavelet transform on each row of the image. This process usually involves two key steps: first, a decomposition operation to separate the data into high-frequency and low-frequency components; second, a downsampling operation to reduce the number of data points to simplify the information, thereby optimizing subsequent processing efficiency and data storage. This method effectively extracts and compresses key features from the original image.

[0078] Decomposition: By using the Haar wavelet basis function, we can effectively decompose the row data of the image into two main parts: high-frequency part (detail information) and low-frequency part (approximate information). The high-frequency part mainly contains the detailed information of the image, such as edges and texture changes, which are fast-changing features in the image; while the low-frequency part represents the approximate information of the image, that is, the slowly changing or flat areas in the image, reflecting the basic structure of the image. Decomposing the frequency helps to analyze and process the features of the image. The decomposition operation is as follows:

[0079]

[0080] Among them, X(·) is the time-frequency diagram after short-time Fourier transform, h[n] is the low-pass filter (LPF), g[n] is the high-pass filter (HPF), A[k] is the decomposed low-frequency information, and D[k] is the decomposed high-frequency detail information.

[0081] Downsampling: In order to effectively reduce the amount of data and improve processing efficiency, downsampling is usually performed after the signal decomposition process. Specifically, one sample point is selected for retention for every two samples. This can not only reduce the storage requirements of the data, but also reduce the computational burden in subsequent processing steps, while retaining the important features of the signal to a certain extent.

[0082] Step 2-2, after completing the one-dimensional discrete wavelet transform of all rows in step 2-1, the same one-dimensional discrete wavelet transform is performed on each column. This method of transforming rows first and then columns allows us to decompose the data in detail from two dimensions. Through this combination of row and column transformations, the original time-frequency map is effectively decomposed into four different sub-bands: each sub-band captures frequency and time domain information in different directions, including low-frequency-low-frequency sub-band (LL), low-frequency-high-frequency sub-band (LH), high-frequency-low-frequency sub-band (HL), and high-frequency-high-frequency sub-band (HH). Such decomposition helps us understand and process different characteristics of the signal from multiple perspectives, improving the accuracy and efficiency of data analysis. These four sub-bands each carry different aspects of the signal information, among which the LL sub-band contains most of the energy and main information, while the other three sub-bands contain more details and edge information.

[0083] Through the multi-resolution analysis method, we effectively decomposed the original time-frequency diagram, reducing its resolution by half in both the row and column directions. Specifically, the original time-frequency diagram dimension is (102, 45). Through this process, the diagram is carefully decomposed into four small sub-spectra, and the dimension of each sub-spectra is changed to (51, 23). This dimensionality reduction not only reduces the burden of data processing, but also makes storage and further analysis more efficient. Each sub-spectrum contains a part of the information in the original time-frequency diagram, which is subdivided into four different frequency and direction combinations, allowing more accurate analysis and processing of local features of the signal. This method is very useful in image compression, signal processing, and other applications that require efficient data decomposition. It can significantly reduce the complexity of the data while maintaining key information.

[0084] Furthermore, in step 3, a lightweight multi-scale neural network architecture is used to extract the RF features of the decomposed spectrum graph and capture the key information in the spectrum graph, specifically including:

[0085] A neural network based on a lightweight multi-scale Inception-Net architecture is used to extract the features of the decomposed spectrum. The network structure captures features at various levels through convolution kernels of different sizes, effectively enhancing the model's ability to process details and global information. This method allows the network to reduce parameters and computation while maintaining high sensitivity to key features, greatly improving the accuracy and efficiency of drone RFID. Feature extraction process:

[0086]

[0087] Among them, X in is the decomposed spectrum obtained in step 2-2, f C onv is the convolution operation, It is a stack of two convolution operations. is a stacking operation of two residual blocks, ν i are the extracted features.

[0088] Furthermore, the feature enhancement method described in step 4 is used to improve the feature effect extracted by the neural network. This method not only enhances the expressiveness of the features extracted from the data, but also significantly improves the expressiveness and accuracy of the features by deeply analyzing and optimizing the key information in these features. This enhancement method enables the network to more effectively recognize and process complex data patterns, thereby improving the overall model performance, including:

[0089] Step 4-1, first calculate the first-order moment and second-order moment of the feature:

[0090]

[0091] Among them, z c The features extracted by the multi-scale network in step 3 have the dimension (B, C, H, W), where B is the number of samples in each batch, C is the number of channels, H and W are the height and width of the feature map, respectively, and h and w are the indices of the h-th row and w-th column in the feature map, respectively. is the first-order moment of the feature map pixel calculated channel by channel, s c It is the second-order moment of the feature map pixel calculated channel by channel.

[0092] In step 4-2, the first-order moments and second-order moments of different eigenvalues ​​are linearly combined by using the convex combination method. The convex combination method can comprehensively consider multiple statistical characteristics and optimize the overall feature expression effect by accurately adjusting the contribution ratio of each eigenvalue. In this way, complex changes in the data can be more effectively captured and described, thereby providing richer and more accurate information input for subsequent analysis and model training. This convex combination method not only enhances the model's sensitivity to subtle differences in data, but also improves the flexibility and adaptability of feature processing, specifically:

[0093] ν=γs k +(1-γ)s m

[0094]

[0095] The linear combination of the means and standard deviations of two different features generates a mixed mean and standard deviation, where s k and m They are the features z k and z m The second moment of and They are the features zk and z m In addition, γ ~ beta (α, α), α is a hyperparameter and in general, α ∈ [0.1, 0.3, 0.5, 0.7, 0.9], γ ∈ [0, 1].

[0096] In step 4-3, based on the calculation of the mixed first-order moment and second-order moment in step 4-2, adaptive normalization technology is used to combine other features with these statistics. In this way, we generate new features that not only inherit the basic properties of the original data, but also increase the expressiveness and discrimination of the data by combining with the mixed mean and standard deviation. The adaptive normalization process dynamically adjusts and standardizes each feature, so that the relative importance and contribution of the features are optimized. This comprehensive and innovative feature processing method provides the model with more refined and information-rich input, which helps to improve the accuracy and efficiency of the model in recognizing complex patterns. The specific feature generation is as follows:

[0097]

[0098] Among them, ν is the scaling factor used to adjust the scale of the normalized values, and μ is the offset used to adjust the mean of the normalized values.

[0099] Furthermore, the label classification network described in step 5 is used to classify the drone labels, and then the network is continuously trained and optimized using the back-propagation algorithm. In this way, the network is able to gradually adjust its internal parameters to reduce the prediction error during the classification process. This process involves calculating the loss function in each iteration and then adjusting the network weights based on the loss results, so that the model is more accurate in future predictions.

[0100] The purpose of model training is to construct a mapping function Minimize the overall loss. In practical applications, since only some samples are known, the true data distribution is difficult to calculate accurately. Therefore, we use the empirical distribution based on the training data set to approximate the true sample distribution. Convert the optimization problem to minimize the expected loss on the training set. The specific operations are as follows:

[0101]

[0102] Among them, L{f c [f d (f g (X))],y} is the difference between the model prediction and the true label, L ce represents the cross entropy loss function, N is the number of samples, D t is the input sample set. stf(X i) is the abbreviation of subject to, which means "under the following conditions". g 、f d and f c They are feature extraction operation, data enhancement and label classification operation. Represents the dataset D t Random sampling in .

[0103] Furthermore, the cross-domain drone identification system model optimization process based on deep learning can be described in detail as follows:

[0104]

[0105]

[0106] Finally, the calculated gradient information is fed back to the optimization algorithm. The present invention adopts the stochastic gradient descent SGD algorithm. The parameters of the model are updated according to the gradient. By continuously iterating the process of forward propagation-backward propagation-parameter update, the model can be gradually optimized in the direction of minimizing the loss function.

[0107] Among them, z l ,z ul They refer to the electromagnetic environment characteristics of the UAV system extracted by the neural network during the forward propagation process.

[0108] The server uses a GeForce GTX 2080Ti graphics processor to perform calculations. First, the original (6000, 2)-dimensional I / Q signal is converted into a (102, 45)-dimensional time-frequency image. The two-dimensional discrete wavelet transform is further used to decompose the time-frequency image into four sub-spectral images with dimensions of (51, 23), and the time-frequency image is sent to Python for subsequent processing. The environment in Python is the TORCH 1.10.2 deep learning framework. The model was tested using the sklear0.24.2 tool. The training dataset used contains 2765 time-frequency images of 7 drones of the same model. Each drone includes 395 time-frequency images. The dimension of each time-frequency image is 51×23. The specific parameter settings are shown in the following table.

[0109]

[0110]

[0111] The accuracy of different methods on the drone dataset is shown in the following table:

[0112]

[0113] In summary, the present invention proposes a cross-domain UAV system identification method based on deep learning. Aiming at the problem of insufficient model generalization ability in actual cross-domain RF fingerprint recognition tasks, this method adopts the following innovative points:

[0114] The two-dimensional discrete wavelet transform is used to decompose the time-frequency diagram, reduce the resolution of the spectrum, reduce the model calculation complexity, and the decomposition is conducive to analyzing and processing the local characteristics of the signal;

[0115] A lightweight multi-scale neural network was constructed to extract UAV RF features from different levels for learning, which enhanced the model's ability to process details and global information, captured key information in the spectrum, and improved the model's generalization performance.

[0116] The expressiveness of the features extracted from the data is enhanced through feature enhancement methods. In-depth analysis and optimization of the key information in these features significantly improve the expressiveness of the features and enhance the robustness of the model in cross-domain recognition tasks.

[0117] This method makes full use of the powerful advantages of deep learning in feature representation and pattern recognition, and cleverly applies multi-scale networks to enhance the learning ability of the model. Compared with the traditional method based on manually designed features, this method not only has a higher level of automation, but also shows better robustness and adaptability in complex electromagnetic environments. Through continuous optimization and iteration, this cross-domain recognition technology can effectively solve the problem of insufficient model generalization ability encountered by UAV systems in electromagnetic environment perception, and provide more reliable electromagnetic environment information support for future UAV applications.

Claims

1. A cross-domain drone identification method based on deep learning, characterized in that: The following steps are involved: Step 1: Preprocess the UAV RF data collected from different time periods to generate a time-frequency diagram of the UAV RF signal; the preprocessing includes signal normalization and short-time Fourier transform; Step 2: Decompose the original spectrum using a multi-resolution analysis method, and perform two operations: decomposing and downsampling the spectrum in the row direction and column direction respectively; Step 3: A lightweight multi-scale neural network architecture is used to extract the decomposed spectrogram features and capture the key information in the spectrogram; Step 4: The feature enhancement method is used to enhance the features extracted by the neural network, analyze and optimize the key information of the features, and enhance the expressiveness of the features; Step 5: Use the label classification network to classify drone labels, so as to achieve accurate identification of different drones.

2. According to the deep learning-based cross-domain drone identification method of claim 1, it is characterized in that: In step 1, the time-frequency diagram of the drone RF signal is generated by a preprocessing method, which specifically includes: Step 1-1, use the root mean square RMS to normalize the input I / Q data: Among them, x i is a complex number normalized by RMS, I i +jQ i is the input I / Q data, is the square of the input I / Q data amplitude, N is the number of samples, where I i and Q i Represent the real and imaginary parts of the original time domain signal at the i-th moment, j is the imaginary unit, I n and Q n Respectively represent the real and imaginary parts of the original time domain signal at the nth moment; Step 1-2, use short-time Fourier transform (STFT) to convert the normalized I / Q data into a time-frequency diagram: Among them, X i is the time-frequency diagram after STFT conversion, x(n) is the value of the normalized I / Q data at time point n, ω(n-mR) is the Hamming window function, which divides the continuous signal into a series of short-time windows for local Fourier transform, R represents the number of overlapping samples between adjacent windows, w represents the window function, m is the index of the window, N is the number of normalized I / Q samples, is the discrete angular frequency, describing the frequency component in the frequency domain, k is the index of the frequency bin, and its value is k=0,1,2,...,N-1; the original (6000, 2)-dimensional sample signal is converted into a (102, 45)-dimensional time-frequency image by Fourier transform, and a data set containing 7 time-frequency images of the same UAV is generated from the time-frequency image. The short-time Fourier transform performs short-time windowing and Fourier transform on the continuous-time signal to obtain the time-frequency domain representation of the signal, which is convenient for analyzing the spectral characteristics of the time-varying signal.

3. According to the cross-domain drone identification method based on deep learning in claim 2, it is characterized in that: In step 2, a multi-resolution analysis method is used to decompose the original spectrum, and the original spectrum is decomposed and downsampled in two steps from the row direction and the column direction, and the resolution of the original spectrum is reduced from two directions at the same time, specifically including: Step 2-1, performing a one-dimensional discrete wavelet transform on each row of the image, including two operations: decomposition and downsampling; Decomposition: Use the Haar wavelet basis function to decompose the line data into high-frequency and low-frequency parts: Where X(·) is the time-frequency diagram after short-time Fourier transform, h[n] is the low-pass filter LPF, g[n] is the high-pass filter HPF, A[k] is the decomposed low-frequency information, and D[k] is the decomposed high-frequency detail information; Downsampling: To reduce the amount of data, the signal is downsampled after decomposition, that is, one sample is taken for every two samples; Step 2-2, after all the row transformations are completed in step 2-1, the same one-dimensional discrete wavelet transform is performed on each column; after the row transformation and column transformation, the original time-frequency graph is decomposed into 4 sub-bands: Low frequency - low frequency subband LL, contains the main structural information of the image; The low-frequency-high-frequency subband LH contains the horizontal detail information of the image, such as horizontal edges; The high-low frequency subband HL contains vertical detail information of the image, such as vertical edges; High frequency - high frequency subband HH, contains diagonal detail information of the image; After multi-resolution analysis, the resolution of the original time-frequency diagram is reduced by half in both row and column directions, and the original time-frequency diagram with a dimension of (102, 45) is decomposed into four sub-spectra with a dimension of (51, 23).

4. According to the deep learning-based cross-domain drone identification method of claim 3, it is characterized in that: In step 3, a lightweight multi-scale neural network architecture is used to extract the RF features of the decomposed spectrogram and capture the key information in the spectrogram, including: A neural network based on a lightweight multi-scale Inception-Net architecture is used to extract the features of the decomposed spectrum graph; the network structure captures features at various levels through convolution kernels of different sizes, effectively enhancing the model's ability to process details and global information; this method allows the network to reduce parameters and computational complexity while maintaining high sensitivity to key features, greatly improving the accuracy and efficiency of drone radio frequency identification; feature extraction process: Among them, X in is the decomposed spectrum obtained in step 2-2, f C onv is the convolution operation, It is a stack of two convolution operations. is a stacking operation of two residual blocks, v i are the extracted features.

5. According to the deep learning-based cross-domain drone identification method of claim 4, it is characterized in that: In step 4, feature enhancement methods are used to enhance the features extracted by the neural network, analyze and optimize the key information of the features, and enhance the expressiveness of the features, including: Step 4-1, first calculate the statistical value of the feature: Among them, z c The features extracted by the multi-scale network in step 3 have the dimension (B, C, H, W), where B is the number of samples in each batch, C is the number of channels, H and W are the height and width of the feature map, respectively, and h and w are the indices of the hth row and wth column in the feature map, respectively; is the first-order moment of the feature map pixel calculated channel by channel, s c It is the second-order moment of the feature map pixel calculated channel by channel; Step 4-2, linearly combine the first-order moments and second-order moments of different eigenvalues ​​calculated in step 4-1 by convex combination, and realize linear fusion of the first-order moments and second-order moments of different eigenvalues; this method comprehensively considers multiple statistical attributes, and enhances the expressiveness of the overall feature by finely adjusting the weight ratio of each eigenvalue; this strategy effectively depicts the complex dynamics in the data, and provides rich and accurate information input for further data analysis and model training; the use of this convex combination method not only improves the model's perception of subtle data changes, but also increases the flexibility and adaptability of feature processing, specifically: n=γs k +(1-γ)s m The first-order moment and second-order moment of two different features are linearly combined to generate mixed first-order moment and second-order moment, where s k and m They are the features z k and z m The second moment of and They are the features z k and z m The first moment of z k and z m Represent two different feature maps, γ~beta(α,α), α is a hyperparameter, α∈[0.1,0.3,0.5,0.7,0.9], γ∈[0,1]; Step 4-3, based on the mixed moment calculated in step 4-2, adaptive normalization is used to combine other features with the mixed moment to generate new features: Among them, v is the scaling factor used to adjust the scale of the normalized values, and μ is the offset used to adjust the mean of the normalized values.

6. The cross-domain drone identification method based on deep learning according to claim 5 is characterized in that: In step 5, the label classification network is used to classify the drone labels. Subsequently, the network is continuously trained and optimized using the back propagation algorithm. The optimization problem is converted into minimizing the expected loss on the training set. The specific operations are as follows: Among them, L{f c [f d (f g (X))],y} is the difference between the model prediction and the true label, L ce represents the cross entropy loss function, N is the number of samples, D t is the input sample set; stf(X i ) is the abbreviation of subject to, which means "under the following conditions"; g 、f d and f c They are feature extraction operation, data enhancement and label classification operation; Represents the dataset D t Random sampling in y i is the One-hot encoding of the corresponding label.

Images