Unmanned aerial vehicle radio frequency signal recognition method and device based on improved capsule network model, equipment and medium

Abstract

The application discloses an unmanned aerial vehicle radio frequency signal recognition method and device based on an improved capsule network model, equipment and a medium, relates to the technical field of unmanned aerial vehicle recognition, and comprises the following steps: collecting an unmanned aerial vehicle radio frequency original signal, generating a radio frequency time-frequency graph through short-time Fourier transform and logarithmic transformation; then inputting the radio frequency time-frequency graph into the improved capsule network model, outputting an unmanned aerial vehicle recognition result through multi-scale feature extraction, spatial relationship capture and classification processing. Through the combination of the improved Inception structure and the capsule network, the adaptability of the model to complex electromagnetic environments is enhanced, the accuracy and stability of the unmanned aerial vehicle radio frequency signal recognition are improved, large-size time-frequency graph input is adapted, multi-scale features are comprehensively extracted, and spatial relationships are retained.

Description

Technical Field

[0001] This invention relates to the field of drone identification technology, and in particular to a method, apparatus, device, and medium for drone radio frequency signal identification based on an improved capsule network model. Background Technology

[0002] Currently, existing approaches to UAV radio frequency signal recognition mainly focus on two main directions: feature selection and deep learning model improvement. In terms of feature processing, methods such as short-time Fourier transform, discrete Fourier transform, and wavelet transform are commonly used to convert the raw radio frequency signal into a time-frequency graph or spectrum graph to preserve the signal's time and frequency domain features. Regarding model applications in UAV radio frequency signal recognition, existing solutions include multi-size convolutional neural networks (CNNs), basic capsule networks, and capsule networks based on the traditional Inception structure. Among these, multi-size CNNs extract multi-scale features by stacking multi-size convolutional kernels, basic capsule networks preserve feature space relationships through dynamic routing mechanisms, and capsule networks based on the traditional Inception structure combine multi-scale parallel feature extraction with the advantages of capsule networks to attempt to improve recognition performance.

[0003] Current approaches still have significant limitations: First, in the feature extraction process, traditional CNN models are prone to losing spatial hierarchical information of the signal due to pooling operations, and single or simply combined convolutional kernels are insufficient to fully capture the local details and global correlation features of RF time-frequency maps, resulting in insufficient feature representation capabilities. Second, the single-layer convolutional layers of the original capsule network have limitations when processing large-size RF time-frequency maps, easily leading to an excessive number of primary capsules and affecting network computational efficiency. Third, the traditional Inception structure combined with capsule networks has not undergone targeted optimization of the Inception module, leaving room for improvement in the parameter efficiency of feature extraction and the accuracy of multi-scale fusion. Moreover, in long-distance, low signal-to-noise ratio scenarios, the recognition accuracy of these models is easily and significantly reduced due to signal path attenuation, environmental scattering, and electromagnetic interference, making them unsuitable for complex real-world application scenarios.

[0004] Therefore, there is an urgent need for a method for identifying radio frequency signals of drones to improve the accuracy and stability of drone radio frequency signal identification. Summary of the Invention

[0005] The main objective of this application is to provide a method, apparatus, and medium for identifying radio frequency signals of unmanned aerial vehicles (UAVs) based on an improved capsule network model, aiming to solve the technical problem of how to improve the accuracy and stability of UAV radio frequency signal identification.

[0006] To achieve the above objectives, this application proposes a method for identifying UAV radio frequency signals based on an improved capsule network model, comprising:

[0007] Collect raw radio frequency signals from drones at different distances;

[0008] Short-time Fourier transform and logarithmic transform are performed on the raw radio frequency signal of the UAV to obtain the radio frequency time-frequency diagram;

[0009] The radio frequency time-frequency map is input into the improved capsule network model for feature extraction and classification to obtain the UAV recognition result. The improved capsule network model includes a 7×7 convolutional layer, an improved Inception feature extraction module, a primary capsule layer, and a digital capsule layer. The improved Inception feature extraction module includes three improved Inception structures.

[0010] In one embodiment, the step of inputting the radio frequency time-frequency map into an improved capsule network model for feature extraction and classification to obtain the UAV identification result includes:

[0011] The radio frequency time-frequency map is input into a 7×7 convolutional layer of the improved capsule network model for preliminary feature extraction to obtain an initial feature map;

[0012] The initial feature map is processed through the improved Inception structure of the improved Inception feature extraction module to obtain a fused feature map;

[0013] The multi-scale features are input into the primary capsule layer to convert the multi-scale features into a set of vector capsules.

[0014] The vector capsule set is processed by a dynamic routing mechanism to obtain feature space information;

[0015] The feature space information is input into the digital capsule layer for classification processing, and the UAV recognition result is output.

[0016] In one embodiment, the step of processing the initial feature map through the improved Inception structure of the improved Inception feature extraction module to obtain the fused feature map includes:

[0017] The initial feature map is input into the first parallel path, the second parallel path, the third parallel path, and the fourth parallel path of the improved Inception structure. The first parallel path is configured with a 1×1 convolutional layer, the second parallel path is configured with a 1×1 convolutional layer and a 3×3 convolutional layer in sequence, the third parallel path is configured with a 1×1 convolutional layer and two consecutive 3×3 convolutional layers in sequence, and the fourth parallel path is configured with a 3×3 pooling layer and a 1×1 convolutional layer in sequence.

[0018] The initial feature map is reduced in dimensionality and reorganized by a 1×1 convolutional layer in the first parallel path to obtain the first-dimensional feature.

[0019] The initial feature map is compressed by channel compression through a 1×1 convolutional layer in the second parallel path, and then local detail-related features are captured by a 3×3 convolutional layer to obtain the second-dimensional features.

[0020] The initial feature map is filtered by a 1×1 convolutional layer in the third parallel path, and then global correlation features are captured by two consecutive 3×3 convolutional layers to obtain the third dimension feature.

[0021] The initial feature map is downsampled and compressed using a 3×3 pooling layer in the fourth parallel path, and then channel optimization is performed using a 1×1 convolutional layer to obtain the fourth dimension feature.

[0022] The first dimension feature, the second dimension feature, the third dimension feature and the fourth dimension feature are respectively subjected to batch normalization processing to obtain the standardized first dimension feature, the standardized second dimension feature, the standardized third dimension feature and the standardized fourth dimension feature;

[0023] The standardized first-dimensional features, standardized second-dimensional features, standardized third-dimensional features, and standardized fourth-dimensional features are fused using a channel splicing operation to obtain a fused feature map.

[0024] In one embodiment, the step of converting the fused feature map into a set of vector capsules input to the primary capsule layer includes:

[0025] Obtain the number of capsules in the primary capsule layer and the capsule dimension corresponding to each capsule. The number of capsules is determined based on the spatial size calculation result of the convolution operation and the number of capsule types. The capsule dimension is a fixed value. The spatial size calculation result of the convolution operation refers to the product of the height and width of the output feature map after the input feature map is processed by the convolution layer. The number of capsule types is the number of capsule types that exist in parallel at each spatial location in the primary capsule layer.

[0026] The multi-scale features of the fused feature map are mapped using a convolutional layer of a preset size to obtain a multi-channel feature map.

[0027] Based on the capsule dimension of the primary capsule layer, the multi-channel feature map is dimensionally split to obtain feature blocks corresponding to the number of capsules;

[0028] Each feature block is subjected to vector transformation to obtain an initial vector corresponding to the number of capsules;

[0029] Each initial vector is normalized and compressed using a nonlinear activation function to obtain a normalized vector capsule;

[0030] Integrate all the normalized vector capsules to obtain a vector capsule set.

[0031] In one embodiment, the step of processing the vector capsule set through a dynamic routing mechanism to obtain feature space information includes:

[0032] Obtain the preset number of target capsules and vector dimensions of the digital capsule layer, and initialize the association weight matrix between the target capsules and each vector capsule in the vector capsule set;

[0033] The initial association weight between each vector capsule and each target capsule in the vector capsule set is calculated based on the association weight matrix.

[0034] Perform softmax normalization on all initial association weights corresponding to the same target capsule to obtain the coupling coefficient;

[0035] The corresponding vector capsules in the vector capsule set are weighted and summed according to the coupling coefficient to obtain the original input vector of each target capsule;

[0036] The original input vector is nonlinearly transformed by a preset compression function to obtain the intermediate output vector of the target capsule;

[0037] Calculate the vector similarity between the intermediate output vector and the corresponding vector capsule, and iteratively update the coupling coefficient based on the similarity to obtain the change in the coupling coefficient, until a preset number of iterations is reached or the change in the coupling coefficient is less than a preset threshold, to obtain the optimized target capsule output vector;

[0038] By integrating all the optimized target capsule output vectors, feature space information containing spatial relationships between features is obtained.

[0039] In one embodiment, the step of inputting the feature space information into a digital capsule layer for classification processing and outputting the UAV identification result includes:

[0040] Obtain preset parameters for the digital capsule layer, wherein the preset parameters include the number of digital capsules and the vector dimension of each digital capsule, and the number of digital capsules corresponds to the number of drone categories set in the identification task;

[0041] The feature space information is assigned features according to the vector dimension to obtain category-related features corresponding to each digital capsule;

[0042] The category-related features of each digital capsule are linearly transformed by the feature mapping matrix of the digital capsule layer to obtain the category feature vector corresponding to each digital capsule.

[0043] Calculate the magnitude of the feature vector for each category and obtain the predicted probability value for the corresponding UAV category, wherein the magnitude value is positively correlated with the predicted probability value;

[0044] Integrate the predicted probability values ​​of all drone categories to generate a category predicted probability set;

[0045] The category with the highest predicted probability value is selected from the predicted probability set to obtain the drone identification result.

[0046] In one embodiment, the step of performing short-time Fourier transform and logarithmic transform on the raw radio frequency signal of the UAV to obtain a radio frequency time-frequency diagram includes:

[0047] The raw radio frequency signal of the drone is sliced ​​into multiple continuous signal segments.

[0048] A Hamming window is applied to each of the signal segments to obtain windowed signal segments.

[0049] Perform a local Fourier transform on each of the signal segments to obtain the corresponding local Fourier transform result;

[0050] The windowing and local Fourier transform operations are repeated on the unprocessed signal regions in all the signal segments according to the preset step size sliding Hamming window to obtain the full local Fourier transform results;

[0051] All the full local Fourier transform results are concatenated in chronological order to form a two-dimensional time-frequency matrix and obtain the joint distribution characteristics of the signal;

[0052] A logarithmic transformation is performed on the joint distribution characteristics to map the values ​​of the two-dimensional time-frequency matrix to a preset dynamic range, thereby obtaining a radio frequency time-frequency diagram.

[0053] Furthermore, to achieve the above objectives, this application also proposes a drone radio frequency signal identification device based on an improved capsule network model, wherein the drone radio frequency signal identification device based on the improved capsule network model includes:

[0054] The acquisition module is used to acquire raw radio frequency signals from drones at different distances;

[0055] The processing module is used to perform short-time Fourier transform and logarithmic transform on the raw radio frequency signal of the UAV to obtain the radio frequency time-frequency diagram;

[0056] The result module is used to input the radio frequency time-frequency map into the improved capsule network model for feature extraction and classification to obtain the UAV identification result. The improved capsule network model includes a 7×7 convolutional layer, an improved Inception feature extraction module, a primary capsule layer and a digital capsule layer. The improved Inception feature extraction module includes three improved Inception structures.

[0057] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the UAV radio frequency signal identification method based on the improved capsule network model described above.

[0058] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the UAV radio frequency signal identification method based on the improved capsule network model described above.

[0059] This application acquires raw radio frequency (RF) signals from UAVs, generates RF time-frequency maps through short-time Fourier transform and logarithmic transform, and then inputs these maps into an improved capsule network model. Through multi-scale feature extraction, spatial relationship capture, and classification processing, the model outputs UAV recognition results. By combining the improved Inception structure with the capsule network, the model's adaptability to complex electromagnetic environments is enhanced, improving the accuracy and stability of UAV RF signal recognition. It is also adapted to large-size time-frequency map inputs, comprehensively extracting multi-scale features while preserving spatial relationships. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 This is a flowchart illustrating the first embodiment of the UAV radio frequency signal identification method based on the improved capsule network model of this application;

[0062] Figure 2 This is a block diagram of the improved capsule network model structure of the first embodiment of the UAV radio frequency signal identification method based on the improved capsule network model of this application;

[0063] Figure 3 This is a radio frequency time-frequency diagram of various types of UAVs in a real flight scenario d1 of the first embodiment of the UAV radio frequency signal identification method based on the improved capsule network model of this application;

[0064] Figure 4 This is a time-frequency diagram of radio frequency signals of various types of UAVs in a real flight d3 scenario of the first embodiment of the UAV radio frequency signal identification method based on the improved capsule network model of this application;

[0065] Figure 5 This is a flowchart illustrating the second embodiment of the UAV radio frequency signal identification method based on the improved capsule network model of this application;

[0066] Figure 6 This is a schematic diagram of the module structure of the UAV radio frequency signal identification device based on the improved capsule network model of this application;

[0067] Figure 7 This is a schematic diagram of the device structure of the hardware operating environment involved in the UAV radio frequency signal identification method based on the improved capsule network model in the embodiments of this application.

[0068] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0069] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0070] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0071] Currently, existing approaches to UAV radio frequency signal identification mainly focus on two main directions: feature selection and deep learning model improvement. In terms of feature processing, methods such as short-time Fourier transform, discrete Fourier transform, and wavelet transform are commonly used to convert the raw radio frequency signal into a time-frequency graph or spectrum graph to preserve the signal's time-domain and frequency-domain features. Regarding model application, existing solutions include multi-size convolutional neural networks (CNNs), basic capsule networks, and capsule networks based on the traditional Inception structure. Among these, multi-size CNNs extract multi-scale features by stacking multi-size convolutional kernels, basic capsule networks preserve feature space relationships through dynamic routing mechanisms, and capsule networks based on the traditional Inception structure combine multi-scale parallel feature extraction with the advantages of capsule networks to attempt to improve recognition performance.

[0072] Current approaches still have significant limitations: First, in the feature extraction process, traditional CNN models are prone to losing spatial hierarchical information of the signal due to pooling operations, and single or simply combined convolutional kernels are insufficient to fully capture the local details and global correlation features of RF time-frequency maps, resulting in insufficient feature representation capabilities. Second, the single-layer convolutional layers of the original capsule network have limitations when processing large-size RF time-frequency maps, easily leading to an excessive number of primary capsules and affecting network computational efficiency. Third, the traditional Inception structure combined with capsule networks has not undergone targeted optimization of the Inception module, leaving room for improvement in the parameter efficiency of feature extraction and the accuracy of multi-scale fusion. Moreover, in long-distance, low signal-to-noise ratio scenarios, the recognition accuracy of the aforementioned models is easily and significantly reduced due to signal path attenuation, environmental scattering, and electromagnetic interference, making them unsuitable for complex real-world application scenarios.

[0073] Therefore, this application proposes a method for identifying UAV radio frequency signals based on an improved capsule network model to solve the above problems.

[0074] Based on the above, this application also provides a method for identifying UAV radio frequency signals based on an improved capsule network model, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the UAV radio frequency signal identification method based on the improved capsule network model of this application. In this embodiment, the UAV radio frequency signal identification method based on the improved capsule network model includes steps S10 to S30:

[0075] Step S10: Collect raw radio frequency signals from the drone at different distances.

[0076] It should be noted that the raw radio frequency signal of a drone refers to the unprocessed electromagnetic signal transmitted by the drone through the communication link during flight. This signal contains the drone's communication commands, status data, and other information, and is the core data foundation for identifying drone types. This signal will exhibit unique frequency, amplitude, and other characteristics depending on the drone model and flight status.

[0077] Understandably, the first step is to select a suitable experimental site, ensuring the environment can meet the requirements of drone flight and signal acquisition at different distances. Then, RF signal acquisition equipment such as a Universal Software Radio Peripheral (USRP) is set up, with appropriate parameters such as sampling frequency, center frequency, and receiver gain configured to ensure accurate capture of the drone's RF signals. Next, the drone is controlled to fly sequentially within preset distance ranges (e.g., 20-40 meters, 40-80 meters, 80-150 meters), collecting a sufficient number of raw RF signal samples within each range. The purpose of this is to obtain the characteristic variations of the drone's RF signals at different distances, providing diverse data support for subsequent model training and validation, thereby improving the model's ability to recognize drone signals and its generalization performance in different distance scenarios.

[0078] Step S20: Perform short-time Fourier transform and logarithmic transform on the raw radio frequency signal of the UAV to obtain the radio frequency time-frequency diagram.

[0079] It should be noted that the Short-Time Fourier Transform (SFT) is a signal processing method that decomposes a non-stationary signal into a two-dimensional distribution in time and frequency. By windowing, a long signal is divided into multiple short time intervals, and a Fourier Transform is performed on each short time interval, thus simultaneously reflecting the signal's time-domain and frequency-domain characteristics. The Logarithmic Transform (RTT) is a processing method that performs logarithmic operations on the joint time-frequency distribution of the signal. Its purpose is to enhance the discernibility of signal features, compress the dynamic range of the signal, and make the energy distribution of the signal more intuitive and easier to read, facilitating the extraction of effective features by subsequent models.

[0080] Further, step S20 includes: slicing the raw UAV radio frequency signal into multiple continuous signal segments; applying a Hamming window to each signal segment to obtain windowed signal segments; performing a local Fourier transform on each signal segment to obtain the corresponding local Fourier transform result; repeatedly applying the windowing and local Fourier transform operations to the unprocessed signal regions in all signal segments by sliding the Hamming window according to a preset step size to obtain the full local Fourier transform result; concatenating all full local Fourier transform results in chronological order to form a two-dimensional time-frequency matrix and obtain the joint distribution characteristics of the signal; performing a logarithmic transform on the joint distribution characteristics to map the values ​​of the two-dimensional time-frequency matrix to a preset dynamic range to obtain the radio frequency time-frequency diagram. It should be noted that the Hamming window is a commonly used window function whose amplitude smoothly transitions from both ends to the middle, effectively reducing spectral leakage during signal segmentation and making the spectral characteristics of each signal segment closer to reality. It is widely used in the segmentation processing of non-stationary signals. The windowed signal segment refers to a short-segment signal obtained by segmenting the raw UAV RF signal using a Hamming window. Each signal segment is considered a stationary signal. Local Fourier Transform is performed separately on each signal segment, acquiring the frequency distribution characteristics of each short-segment signal and reflecting the frequency domain information of the signal within that time period. Joint distribution features are feature data obtained by superimposing the local Fourier Transform results of all signal segments, integrating the frequency domain characteristics of the signal at different time segments, and fully presenting the joint time-frequency distribution of the signal. The RF time-frequency plot is two-dimensional image data obtained after logarithmic transformation, intuitively presenting the energy distribution of the RF signal in different time and frequency dimensions. It can be understood that the raw UAV RF signal is first sliced ​​into multiple continuous signal segments. The purpose of this is to decompose the long signal into multiple shorter signal segments, facilitating subsequent local processing and feature extraction. The length of each signal segment is usually set according to the signal characteristics and analysis requirements. Secondly, a Hamming window is applied to each signal segment to obtain the windowed signal segment. The Hamming window reduces abrupt changes at signal edges, thereby reducing spectral leakage and making spectral analysis more accurate. Windowed signal segments better reflect the signal's characteristics within a local time window. Next, a local Fourier transform is performed on each windowed signal segment to obtain the corresponding local Fourier transform result. The local Fourier transform captures the characteristics of each signal segment in the frequency domain, reflecting the energy distribution of the signal at different frequency components. This step aims to transform the signal from the time domain to the frequency domain, providing a foundation for subsequent time-frequency analysis. Then, the Hamming window slides according to a preset step size, repeating the windowing and local Fourier transform operations on the unprocessed signal regions in all signal segments to obtain the full local Fourier transform result. By using a sliding window, it ensures that every part of the signal is processed, thus obtaining the frequency domain characteristics of the signal at different time points.The preset step size determines the window sliding interval; a smaller step size results in higher time resolution but also increases computational complexity. Next, all full-local Fourier transform results are concatenated chronologically to form a two-dimensional time-frequency matrix. Each row of this matrix represents the frequency domain characteristics at a given time point, and each column represents the variation of a frequency component across different time points. This method yields the joint distribution characteristics of the signal, i.e., a comprehensive representation of the signal in both time and frequency. Finally, a logarithmic transform is performed on the joint distribution characteristics, mapping the values ​​of the two-dimensional time-frequency matrix to a preset dynamic range to obtain the radio frequency time-frequency plot. The logarithmic transform compresses the dynamic range, enhancing the characteristics of low-energy signals while suppressing excessively strong high-energy signals.

[0081] Step S30: Input the radio frequency time-frequency map into the improved capsule network model for feature extraction and classification to obtain the UAV recognition result.

[0082] It should be noted that, as Figure 2 The diagram shows the framework of the improved capsule network model. The improved capsule network model includes 7×7 convolutional layers, an improved Inception feature extraction module, primary capsule layers, and digital capsule layers. The improved Inception feature extraction module comprises three improved Inception structures. The improved Inception feature extraction module is a feature extraction unit optimized from the traditional Inception structure. The 7×7 convolutional layer refers to a module that uses a 7×7 convolutional layer to perform convolution operations on the input data. The improved Inception structure is a feature extraction unit containing multiple parallel convolutional paths, each path configured with convolutional layers of different sizes. This allows for the simultaneous capture of local detail features and global correlation features of the input data. Stacking multiple units enhances the depth and comprehensiveness of feature extraction. The primary capsule layer contains 32 capsules of 8-dimensional vectors, converting the feature maps extracted by convolution into a set of vector capsules, preparing information for the dynamic routing mechanism. The digital capsule layer contains capsules of 16-dimensional vectors corresponding to the number of drone categories. The modulus of each capsule represents the predicted probability of the corresponding category, responsible for outputting the final classification result.

[0083] Understandably, the preprocessed radio frequency (RF) time-frequency map is first input into the improved capsule network model, which is specifically designed for efficient feature extraction and fusion of RF signals. Initial feature extraction is performed on the input RF map using a 7×7 convolutional layer. The 7×7 convolutional layer quickly captures global features in the RF map using large-size convolutional kernels, providing a foundation for subsequent feature extraction. Next, the output of the 7×7 convolutional layer is passed to the improved Inception feature extraction module. This module consists of three stacked improved Inception structures, each containing multiple parallel convolutional paths configured with different kernel sizes (e.g., 1×1, 3×3, etc.). Through these parallel paths, the model can simultaneously extract local details and global structural features, enhancing the expressive power of the features and generating a fused feature map. Then, the fused feature map is fed into the primary capsule layer. The primary capsule layer transforms the two-dimensional feature map into vector capsules, each containing a set of feature vectors that better represent the spatial relationships and hierarchical structure between features. Subsequently, the set of vector capsules output from the primary capsule layer enters the digital capsule layer. In the digital capsule layer, a dynamic routing mechanism iteratively updates the connection weights between capsules, thereby achieving further feature fusion and classification. The core of this dynamic routing mechanism is to adaptively adjust the connection strength between capsules based on the similarity and correlation between features, enabling the model to more accurately identify the feature patterns of different drone signals. Finally, the magnitude of the vector capsules output by the digital capsule layer is used to represent the predicted probability of different drone categories. By comparing these magnitudes, the model can determine the drone category most likely corresponding to the input signal, thus outputting the final drone identification result. Through these steps, the improved capsule network model can efficiently complete the classification task of drone radio frequency signals, significantly improving the accuracy and efficiency of drone identification.

[0084] Specifically, the flight test used the USRP X310 software radio platform to collect the UAV's radio frequency signals; the matching antenna was the LW2450, supporting the 2.4~2.48GHz and 4.9~5.9GHz frequency bands, with a gain of 3dBi, adapting to the common communication frequency band requirements of UAVs. The USRP X310 transmitted the collected raw IQ data to the Lenovo GMOD G529 server in real time via the PCIe high-speed interface; the server's processor was an Intel(R) Core(TM) i9-13900k, equipped with GUN Radio software.

[0085] The experiment was conducted using a USRP X310 with a sampling frequency of 100 MS / s, a center frequency of 2.44 GHz, and a receiver gain of 80 dB. Four drone models were used for the actual flight tests: DJI Phantom 4 Pro, DJI Matrice 4, DJI Mavic 3, and Autel EVO Lite, labeled as category a, b, c, and d respectively. During flight, RF signals were collected at seven different distances: 20–40 m, 40–80 m, 80–150 m, 300–350 m, 500–550 m, 1000–1100 m, and 2000–2100 m, forming the drone dataset. Table 1 shows the drone data distance labeling comparison table.

[0086] Table 1. Comparison of Distance Markers for Data from Actual UAV Data Collection

[0087]

[0088] After performing short-time Fourier transform and logarithmic transform processing on the radio frequency signals of the actual sampling UAVs, taking the distance scenarios d1 and d3 as examples, the radio frequency time-frequency diagrams of different types of UAVs are as follows: Figure 3 The radio frequency time-frequency diagrams of various types of drones in the actual flight d1 scenario shown are as follows: Figure 4 The table shows the radio frequency (RF) time-frequency plots of various types of UAVs in the actual flight scenario d3. At close range (d1), the RF signal time-frequency characteristics of various UAVs differ significantly, providing a reliable feature basis for accurate model classification. At long range (d3), the RF signal is affected by path attenuation and environmental scattering, resulting in a significant decrease in the signal-to-noise ratio at the receiver. The overall energy of various signals attenuates, and some signal time-frequency characteristics become blurred and details are lost, significantly increasing the difficulty of classification. Multi-size CNNs, Basic Capsule Networks (BLCs), and Inception-based Capsule Networks (Inception Capsule Networks) models are selected as comparative models to evaluate the performance advantages of the model in this embodiment. The four models are used to train and test the RF time-frequency plots of actual flight UAVs at all distances. The UAV classification evaluation metrics for each model are shown in Table 2, which presents the classification evaluation metrics for UAVs at distances d1 to d7.

[0089] Table 2. Evaluation Indicators for UAV Classification in Distance Scenarios of Four Models (d1~d7)

[0090]

[0091] As shown in the table above, this model consistently ranked first in all seven test scenarios across all four core metrics: accuracy, precision, recall, and F1 score. Its advantages were particularly pronounced in mid-to-long-distance scenarios: in the d5 scenario, its F1 score reached 0.8485, far exceeding the second-place model's 0.5563, a performance lead of over 52%. Simultaneously, this model demonstrated excellent robustness, achieving perfect performance (F1=1.0000) in multiple scenarios including d1, d2, and d4. In contrast, the performance of the three comparison models (InceptionCapNet, BasicCapNet, and multi-size CNN) showed significant fluctuations and degradation with increasing distance. This fully demonstrates that this model not only boasts the best overall performance but, more importantly, possesses extremely strong environmental adaptability and reliability, effectively solving the core problem of performance degradation in complex scenarios encountered by existing technologies.

[0092] This embodiment acquires raw radio frequency (RF) signals from a UAV, generates an RF time-frequency map through short-time Fourier transform and logarithmic transform, and then inputs it into an improved capsule network model. Through multi-scale feature extraction, spatial relationship capture, and classification processing, the UAV recognition result is output. The combination of the improved Inception structure and capsule network enhances the model's adaptability to complex electromagnetic environments, improves the accuracy and stability of UAV RF signal recognition, adapts to large-size time-frequency map input, comprehensively extracts multi-scale features, and preserves spatial relationships.

[0093] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 5 The UAV radio frequency signal identification method based on the improved capsule network model, step S50, further includes steps S201 to S205:

[0094] Step S201: Input the radio frequency time-frequency map into the 7×7 convolutional layer of the improved capsule network model for preliminary feature extraction to obtain the initial feature map.

[0095] Specifically, the radio frequency (RF) time-frequency image is first fed into a 7×7 convolutional layer in the improved capsule network model. The core operation of the 7×7 convolutional layer is to perform a sliding convolution operation on the time-frequency image using a 7×7-sized convolutional layer, extracting a large range of feature information from the time-frequency image through the convolution operation. This is done because a large-size convolutional layer can cover a larger receptive field, quickly capturing global features in the time-frequency image and providing a basic feature representation for subsequent feature extraction. After the convolution operation, a feature map, i.e., the initial feature map, is obtained. This initial feature map contains preliminary feature information extracted from the original time-frequency image, which reflects some basic patterns and structures in the time-frequency image. The generation of the initial feature map is the foundation for subsequent multi-scale feature extraction and fusion, providing the model with a preliminary understanding of the input data.

[0096] Step S202: The initial feature map is processed by the improved Inception structure of the improved Inception feature extraction module to obtain the fused feature map.

[0097] Specifically, the initial feature map is first input into the first, second, third, and fourth parallel paths of the improved Inception structure. Then, a 1×1 convolutional layer in the first parallel path performs dimensionality reduction and feature reorganization on the initial feature map to obtain the first-dimensional feature. Next, a 1×1 convolutional layer in the second parallel path performs channel compression on the initial feature map, followed by a 3×3 convolutional layer to capture local detail-related features, resulting in the second-dimensional feature. Finally, a 1×1 convolutional layer in the third parallel path performs feature filtering on the initial feature map, followed by two consecutive 3×3 convolutional layers to capture global related features, resulting in the third-dimensional feature. The initial feature map is then downsampled and compressed using a 3×3 pooling layer in the fourth parallel path, followed by channel optimization using a 1×1 convolutional layer to obtain the fourth-dimensional feature. Then, the first, second, third, and fourth-dimensional features are batch normalized to obtain standardized first, second, third, and fourth-dimensional features, respectively. Finally, a channel concatenation operation is used to fuse the standardized first, second, third, and fourth-dimensional features to obtain a fused feature map.

[0098] It should be noted that the first parallel path is configured with a 1×1 convolutional layer, the second parallel path is configured with a 1×1 convolutional layer and a 3×3 convolutional layer in sequence, the third parallel path is configured with a 1×1 convolutional layer and two consecutive 3×3 convolutional layers in sequence, and the fourth parallel path is configured with a 3×3 pooling layer and a 1×1 convolutional layer in sequence. The 1×1 convolutional layer does not change the spatial size of the feature map; its main function is to increase or decrease the dimensionality of the feature map's channels, while simultaneously achieving feature fusion and recombination between different channels. Two consecutive 3×3 convolutional layers refer to two 3×3 convolutional layers connected sequentially in the same parallel path. By superimposing two small-sized convolutional operations, the range of feature capture can be expanded while reducing the number of convolutional parameters, thus improving the efficiency of feature extraction. The 3×3 pooling layer is a downsampling operation layer using a 3×3 pooling kernel. It aggregates and statistically analyzes feature values ​​in local regions of the feature map, compressing the spatial size while preserving core features and enhancing feature translation invariance. Nonlinear activation functions are functions that introduce nonlinear transformations, which can perform nonlinear mapping on the fused feature matrix after feature fusion, breaking the linear limitations of convolution operations.

[0099] Understandably, the initial feature map is first input into four parallel paths of the improved Inception structure: the first, second, third, and fourth parallel paths. Each parallel path is configured with different combinations of convolutional and pooling layers to extract features of different scales and types. In the first parallel path, a 1×1 convolutional layer is configured. This 1×1 convolutional layer performs dimensionality reduction and feature reorganization on the initial feature map to obtain the first-dimensional feature. The role of the 1×1 convolutional layer is to reduce the number of channels in the feature map, while effectively reorganizing and filtering features, improving computational efficiency and enhancing the expressive power of the features. In the second parallel path, a 1×1 convolutional layer and a 3×3 convolutional layer are configured sequentially. First, the 1×1 convolutional layer compresses the channels of the initial feature map, reducing the number of channels and lowering the computational cost. Then, the 3×3 convolutional layer captures local detail-related features to obtain the second-dimensional feature. The 3×3 convolutional layer can effectively extract features from local regions, enhancing the model's ability to perceive local details. In the third parallel path, a 1×1 convolutional layer and two consecutive 3×3 convolutional layers are configured sequentially. First, the 1×1 convolutional layer filters the initial feature map, removing redundant information and retaining key features. Then, two consecutive 3×3 convolutional layers capture globally correlated features, resulting in the third-dimensional feature. This design further enhances the ability to extract global information, and the stacking of multiple convolutional layers allows for the capture of more complex feature patterns. In the fourth parallel path, a 3×3 pooling layer and a 1×1 convolutional layer are configured sequentially. First, the 3×3 pooling layer downsamples and compresses the initial feature map, reducing its spatial size and computational cost. Then, the 1×1 convolutional layer optimizes the channels, adjusting the number of channels in the feature map to obtain the fourth-dimensional feature. The pooling layer reduces the spatial resolution of the feature map while retaining important feature information, enhancing the model's robustness to features. Finally, batch normalization is performed on the first, second, third, and fourth-dimensional features respectively. Batch normalization standardizes features, giving them similar distribution characteristics, thus improving model stability and convergence speed. Through batch normalization, standardized first-dimensional features, second-dimensional features, third-dimensional features, and fourth-dimensional features are obtained. Finally, channel concatenation is used to fuse the four standardized features, resulting in a fused feature map. Channel concatenation integrates feature information of different scales and types into a single feature map, forming a comprehensive feature representation. The fused feature map simultaneously contains both local detail features and global correlation features, providing a richer information foundation for subsequent feature processing and classification tasks.

[0100] Step S203: The fused feature map input to the primary capsule layer is converted into a set of vector capsules.

[0101] Specifically, firstly, the number of capsules in the primary capsule layer and the capsule dimension corresponding to each capsule are obtained; then, a convolutional layer of preset size is used to perform feature mapping on the multi-scale features of the fused feature map to obtain a multi-channel feature map; next, the multi-channel feature map is dimensionally split based on the capsule dimension of the primary capsule layer to obtain feature blocks corresponding to the number of capsules; then, each feature block is processed by vector transformation to obtain an initial vector corresponding to the number of capsules; then, each initial vector is normalized and compressed using a non-linear activation function to obtain normalized vector capsules; finally, all normalized vector capsules are integrated to obtain a vector capsule set.

[0102] It should be noted that the number of capsules is determined based on the spatial dimension calculation result of the convolution operation and the number of capsule types. The capsule dimension is a fixed value. The spatial dimension calculation result of the convolution operation refers to the product of the height and width of the output feature map after the input feature map is processed by the convolutional layer. The number of capsule types refers to the number of capsule types existing in parallel at each spatial location in the primary capsule layer. Vector transformation processing is the operation of converting two-dimensional feature blocks into one-dimensional vectors. This operation realizes the vectorized representation of features, enabling features to adapt to the vector operation logic of the capsule layer. Vector capsules are vector units obtained by normalizing and compressing the initial vectors. They are the basic components of the primary capsule layer and can accurately represent the attributes and spatial relationships of multi-scale features in vector form.

[0103] It's important to understand that the first step involves obtaining the number of capsules in the primary capsule layer and the capsule dimension for each capsule. The capsule dimension is a fixed value, typically pre-set based on model design and task requirements. This step aims to determine the structure of the primary capsule layer for subsequent feature processing and classification. Next, convolutional layers of a pre-defined size are used to perform feature mapping on the multi-scale features of the fused feature map, resulting in a multi-channel feature map. These pre-defined convolutional layers further refine the feature map, extracting features from different directions and scales, thus obtaining richer feature representations. This step aims to transform the multi-scale feature map into a form suitable for subsequent processing, preparing it as input for the primary capsule layer. Then, based on the capsule dimensions of the primary capsule layer, the multi-channel feature map is dimensionally split, resulting in feature blocks corresponding to the number of capsules. This step decomposes the multi-channel feature map into multiple feature blocks, each corresponding to the input of a capsule. Dimensional splitting distributes information from the feature map into different capsules, allowing each capsule to focus on processing a specific subset of features. Finally, each feature block undergoes vector transformation to obtain an initial vector corresponding to the number of capsules. Vector transformation converts the feature block from matrix form to vector form, enabling subsequent capsule networks to process it. This step preserves the key information in the feature blocks and organizes them into a vector format suitable for capsule network processing. Then, each initial vector is normalized and compressed using a non-linear activation function to obtain normalized vector capsules. The purpose of the non-linear activation function (such as the squash function) is to compress the vector length to a reasonable range while maintaining the vector direction. The purpose of normalization compression is to ensure that the length of the vector capsule can effectively represent the confidence of its corresponding class, while avoiding excessively large vector values ​​from affecting subsequent calculations. Finally, all normalized vector capsules are integrated to obtain the vector capsule set. This step integrates all normalized vector capsules together to form the output of the primary capsule layer. The vector capsule set contains the outputs of all capsules, with each vector capsule representing a specific aspect or class of the input features. The purpose of this step is to provide a structured feature representation for subsequent dynamic routing and classification tasks.

[0104] Step S204: The vector capsule set is processed through a dynamic routing mechanism to obtain feature space information.

[0105] Specifically, firstly, the preset number of target capsules and vector dimensions of the digital capsule layer are obtained, and the association weight matrix between the target capsules and each vector capsule in the vector capsule set is initialized. Then, based on the association weight matrix, the initial association weight between each vector capsule and each target capsule in the vector capsule set is calculated. Next, a softmax normalization operation is performed on all the initial association weights corresponding to the same target capsule to obtain the coupling coefficient. Then, the corresponding vector capsules in the vector capsule set are weighted and summed according to the coupling coefficient to obtain the original input vector of each target capsule. Subsequently, the original input vector is nonlinearly transformed by a preset compression function to obtain the intermediate output vector of the target capsule. Then, the vector similarity between the intermediate output vector and the corresponding vector capsule is calculated, and the coupling coefficient is iteratively updated based on the similarity to obtain the change in the coupling coefficient until a preset number of iterations is reached or the change in the coupling coefficient is less than a preset threshold, to obtain the optimized target capsule output vector. Finally, all optimized target capsule output vectors are integrated to obtain feature space information containing the spatial relationship between features.

[0106] It should be noted that the number of target capsules refers to the total number of capsules in the digital capsule layer used to represent different drone categories. This number corresponds one-to-one with the number of drone categories to be identified and is the basic configuration for classification. Dimension information refers to the vector dimension corresponding to each target capsule in the digital capsule layer. This dimension determines the target capsule's capacity to carry category features. The association weight matrix is ​​a pre-initialized parameter matrix used to quantify the degree of association between vector capsules and target capsules. Its dimension matches the dimensions of the vector capsules and target capsules. The initial association weight is a numerical value calculated based on the weight matrix, representing the degree of association between a single vector capsule and a single target capsule, reflecting the contribution of lower-level vector capsule features to the higher-level target capsule category. The coupling coefficient is a weight value obtained after softmax normalization, accurately quantifying the contribution ratio of each vector capsule to its corresponding target capsule. The change in coupling coefficient is the difference between the coupling coefficients in two adjacent iterations. When this value is less than a preset threshold, it indicates that the coupling coefficient has stabilized, and the iteration can be terminated to ensure efficiency.

[0107] Understandably, the process begins by obtaining the preset number of target capsules and vector dimensions for the digital capsule layer. These parameters are pre-set based on model design and classification task requirements. Simultaneously, the association weight matrix between the target capsules and each vector capsule in the vector capsule set is initialized. This step provides initial parameters for the subsequent dynamic routing mechanism. Next, based on the initialized association weight matrix, the initial association weight between each vector capsule and each target capsule in the vector capsule set is calculated. The initial association weights reflect the initial contribution of each vector capsule to the target capsule, providing a foundation for subsequent feature fusion. Then, a softmax normalization operation is performed on all initial association weights corresponding to the same target capsule to obtain the coupling coefficient. Softmax normalization ensures that the sum of the coupling coefficients is 1, allowing for a reasonable distribution of the contribution of each vector capsule to the target capsule, avoiding some weights being too large or too small and affecting subsequent calculations. Based on the coupling coefficient, the corresponding vector capsules in the vector capsule set are weighted and summed to obtain the original input vector for each target capsule. This step integrates feature information from different sources according to the coupling coefficient, forming the initial input to the target capsule. The original input vector is nonlinearly transformed using a pre-defined compression function (such as the squash function) to obtain the intermediate output vector of the target capsule. The compression function compresses the vector length to a reasonable range while maintaining the vector direction, ensuring that the length of the output vector effectively represents the confidence level of its corresponding category. The vector similarity between the intermediate output vector and the corresponding vector capsule is calculated, and the coupling coefficient is iteratively updated based on the similarity to obtain the change in the coupling coefficient. The purpose of iterative updating is to allow the model to dynamically adjust the coupling coefficient according to the actual correlation between features, optimizing the feature fusion process. When the preset number of iterations is reached or the change in the coupling coefficient is less than a preset threshold, the iteration stops, and the optimized target capsule output vector is obtained. Finally, all optimized target capsule output vectors are integrated to obtain feature space information containing the spatial correlation between features. This step integrates the output vectors of all target capsules to form a comprehensive feature representation, providing high-quality feature input for subsequent classification tasks. In this way, the model can effectively capture the hierarchical relationship and spatial structure between features, improving the accuracy and robustness of UAV radio frequency signal recognition.

[0108] Step S205: Input the feature space information into the digital capsule layer for classification processing, and output the UAV recognition result.

[0109] Specifically, firstly, preset parameters of the digital capsule layer are obtained; then, feature space information is assigned according to the vector dimension to obtain category-related features corresponding to each digital capsule; next, each category-related feature is linearly transformed through the feature mapping matrix of the digital capsule layer to obtain the category feature vector corresponding to each digital capsule; then, the magnitude of each category feature vector is calculated to obtain the predicted probability value of the corresponding drone category, where the magnitude value is positively correlated with the predicted probability value; then, the predicted probability values ​​of all drone categories are integrated to generate a category prediction probability set; finally, the category with the largest predicted probability value is selected from the category prediction probability set to obtain the drone recognition result.

[0110] It should be noted that the preset parameters include the number of digital capsules and the vector dimension of each digital capsule. The number of digital capsules corresponds to the number of drone categories set in the recognition task. Feature allocation is the operation of decomposing the feature space information according to the vector dimension of the digital capsules and distributing it to the corresponding digital capsules, ensuring that each digital capsule only receives feature information related to its own representation category, avoiding category feature confusion. Category association features are the feature data corresponding to a single digital capsule obtained after feature allocation. These features inherit the core association information related to the corresponding drone category in the feature space information and are the basis for generating category feature vectors. The feature mapping matrix is ​​a pre-trained parameter matrix in the digital capsule layer, used to perform linear transformation on the category association features, enhancing the discriminative power of the category features and making the transformed features more accurately match the representation requirements of the corresponding drone category. The linear transformation is a linear operation on the category association features based on the feature mapping matrix. By adjusting the dimension and numerical distribution of the features through matrix multiplication, it highlights category-specific features and weakens the interference of irrelevant features. The category feature vector is a high-dimensional vector obtained after the linear transformation of the category association features. It can accurately represent the core features of the corresponding drone category, and its vector attributes fully preserve the spatial relationship of the features. The feature mapping matrix in the digital capsule layer is used to perform linear transformation on the associated features of each category to obtain the corresponding prediction vector. Under the action of dynamic routing and squash nonlinear compression, the category feature vector corresponding to each digital capsule is formed.

[0111] The modulus of the feature vector is the length of the category feature vector, calculated as the square root of the sum of the squares of each dimension. This value is positively correlated with the predicted probability of the corresponding drone category and is the core quantitative basis for category determination. The predicted probability value is obtained based on the feature vector modulus and represents the likelihood that the input signal belongs to the corresponding drone category. The larger the modulus, the higher the predicted probability value, directly reflecting the degree of category matching. The category predicted probability set is a collection formed by integrating the predicted probability values ​​corresponding to all digital capsules. It contains the predicted probabilities of the input signal corresponding to all drone categories to be identified, fully presenting the model's classification results.

[0112] Understandably, the first step is to obtain the preset parameters for the digital capsule layer. These parameters include the number of digital capsules and the vector dimension of each digital capsule. The number of digital capsules corresponds to the number of drone categories set in the recognition task, ensuring that each digital capsule corresponds to a specific drone category. The vector dimension is preset according to the model design and task requirements, used to represent the size of the feature vector. Next, the feature space information is assigned according to the preset vector dimension to obtain the category-related features corresponding to each digital capsule. The purpose of this step is to distribute the feature space information to different digital capsules, so that each digital capsule can focus on processing features related to a specific drone category. Then, a linear transformation is performed on each category-related feature through the feature mapping matrix of the digital capsule layer to obtain the category feature vector corresponding to each digital capsule. The role of the feature mapping matrix is ​​to transform the input feature vector into a form more suitable for subsequent processing, preparing for the calculation of the prediction probability. Next, the magnitude of each category feature vector is calculated and converted into the prediction probability value of the corresponding drone category. The magnitude value is positively correlated with the prediction probability value, that is, the larger the magnitude, the higher the prediction probability value. The purpose of this step is to convert the magnitude of the feature vector into a probability value, representing the confidence level that the input signal belongs to a specific drone category. Then, the predicted probability values ​​of all drone categories are integrated to generate a category prediction probability set. This step summarizes the predicted probability values ​​of all categories to form a complete prediction result, providing a basis for the final classification decision. Finally, the category with the highest predicted probability value is selected from the category prediction probability set to obtain the drone identification result. The purpose of this step is to select the most likely drone category based on the predicted probability value as the final identification result. In this way, the model can effectively identify the drone category to which the input signal belongs, completing the classification task.

[0113] This embodiment performs initial feature extraction by inputting the radio frequency time-frequency map into a 7×7 convolutional layer to obtain an initial feature map. Next, the improved Inception feature extraction module further processes the data to obtain a fused feature map. Then, the fused feature map is input into a primary capsule layer to convert it into a set of vector capsules, and a dynamic routing mechanism is used to obtain feature space information. Finally, the feature space information is input into a digital capsule layer for classification, outputting the UAV recognition result. By combining the improved Inception structure and capsule network, multi-scale features are effectively extracted, and the dynamic routing mechanism enhances the feature fusion capability, significantly improving the accuracy and robustness of UAV radio frequency signal recognition while reducing model complexity and increasing computational efficiency.

[0114] Based on the first embodiment of this application, this application also provides a UAV radio frequency signal identification device based on an improved capsule network model. Please refer to... Figure 6 The device includes:

[0115] Acquisition module 10 is used to acquire raw radio frequency signals from the UAV.

[0116] Processing module 20 is used to perform short-time Fourier transform and logarithmic transform on the raw radio frequency signal of the UAV to obtain the radio frequency time-frequency diagram.

[0117] The result module 30 is used to input the radio frequency time-frequency map into the improved capsule network model for feature extraction and classification to obtain the UAV recognition result. The improved capsule network model includes a 7×7 convolutional layer, an improved Inception feature extraction module, a primary capsule layer and a digital capsule layer. The improved Inception feature extraction module includes three improved Inception structures.

[0118] The UAV radio frequency signal identification device based on an improved capsule network model provided in this application, employing the UAV radio frequency signal identification method based on an improved capsule network model in the above embodiments, can solve the technical problem of how to improve the accuracy and stability of UAV radio frequency signal identification. Compared with the prior art, the beneficial effects of the UAV radio frequency signal identification device based on an improved capsule network model provided in this application are the same as the beneficial effects of the UAV radio frequency signal identification method based on an improved capsule network model provided in the above embodiments, and other technical features in the UAV radio frequency signal identification device based on an improved capsule network model are the same as the features disclosed in the methods of the above embodiments, and will not be repeated here.

[0119] This application provides a drone radio frequency signal identification device based on an improved capsule network model. The drone radio frequency signal identification device based on the improved capsule network model includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the drone radio frequency signal identification method based on the improved capsule network model in the above embodiment 1.

[0120] The following is for reference. Figure 7This document illustrates a schematic diagram of a UAV radio frequency signal identification device based on an improved capsule network model, suitable for implementing embodiments of this application. The UAV radio frequency signal identification device based on the improved capsule network model in this application can include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), and vehicle-mounted terminals (e.g., vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. Figure 7 The drone radio frequency signal identification device based on the improved capsule network model shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0121] like Figure 7 As shown, the UAV RFID device based on the improved capsule network model may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage device 1003 into a random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the UAV RFID device based on the improved capsule network model. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the UAV RFID device based on an improved capsule network model to communicate wirelessly or wiredly with other devices to exchange data. Although various UAV RFID devices based on improved capsule network models are shown in the figures, it should be understood that implementation or possession of all shown is not required. More or fewer of these devices may be implemented alternatively.

[0122] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable storage medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0123] The UAV radio frequency signal identification device based on the improved capsule network model provided in this application, employing the UAV radio frequency signal identification method based on the improved capsule network model in the above embodiments, can solve the technical problem of how to improve the accuracy and stability of UAV radio frequency signal identification. Compared with the prior art, the beneficial effects of the UAV radio frequency signal identification device based on the improved capsule network model provided in this application are the same as the beneficial effects of the UAV radio frequency signal identification method based on the improved capsule network model provided in the above embodiments, and other technical features in this UAV radio frequency signal identification device based on the improved capsule network model are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0124] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0125] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0126] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the UAV radio frequency signal identification method based on the improved capsule network model in the above embodiments.

[0127] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible storage medium containing or storing a program that can be executed by instructions, used by a device, or used in conjunction with it. The program code contained on the computer-readable storage medium may be transmitted using any suitable storage medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0128] The aforementioned computer-readable storage medium may be included in a drone radio frequency signal identification device based on an improved capsule network model; or it may exist independently and not assembled into a drone radio frequency signal identification device based on an improved capsule network model.

[0129] The aforementioned computer-readable storage medium carries one or more programs that, when executed by a UAV RFID device based on an improved capsule network model, enable the UAV RFID device based on the improved capsule network model to write computer program code for performing the operations of this application in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—and conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0130] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using dedicated hardware-based implementations that perform the specified functions or operations, or can be implemented using a combination of dedicated hardware and computer instructions.

[0131] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0132] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described UAV radio frequency signal identification method based on an improved capsule network model. This addresses the technical problem of improving the accuracy and stability of UAV radio frequency signal identification. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the UAV radio frequency signal identification method based on the improved capsule network model provided in the above embodiments, and will not be elaborated upon here.

[0133] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the UAV radio frequency signal identification method based on the improved capsule network model described above.

[0134] The computer program product provided in this application can solve the technical problem of how to improve the accuracy and stability of UAV radio frequency signal identification. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the UAV radio frequency signal identification method based on the improved capsule network model provided in the above embodiments, and will not be repeated here.

[0135] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for identifying radio frequency signals of unmanned aerial vehicles (UAVs) based on an improved capsule network model, characterized in that, The method includes: Collect raw radio frequency signals of drones at different distances. Raw radio frequency signals of drones refer to the electromagnetic signals transmitted by drones through communication links during flight without any processing. They include the drone's communication commands and status data. They are the core data basis for identifying drone types and will exhibit unique frequencies and amplitudes depending on the drone model and flight status. Short-time Fourier transform and logarithmic transform are performed on the raw radio frequency signal of the UAV to obtain the radio frequency time-frequency diagram; The radio frequency time-frequency map is input into the improved capsule network model for feature extraction and classification to obtain the UAV recognition result. The improved capsule network model includes a 7×7 convolutional layer, an improved Inception feature extraction module, a primary capsule layer and a digital capsule layer. The improved Inception feature extraction module includes three improved Inception structures. The step of inputting the radio frequency time-frequency map into the improved capsule network model for feature extraction and classification to obtain the UAV identification result includes: The radio frequency time-frequency map is input into a 7×7 convolutional layer of the improved capsule network model for preliminary feature extraction to obtain an initial feature map; The initial feature map is processed by the improved Inception structure of the improved Inception feature extraction module to obtain the fused feature map; The fused feature map is input into the primary capsule layer and converted into a set of vector capsules; The vector capsule set is processed by a dynamic routing mechanism to obtain feature space information; The feature space information is input into the digital capsule layer for classification processing, and the UAV recognition result is output.

2. The method as described in claim 1, characterized in that, The step of processing the initial feature map through the improved Inception structure of the improved Inception feature extraction module to obtain the fused feature map includes: The initial feature map is input into the first parallel path, the second parallel path, the third parallel path, and the fourth parallel path of the improved Inception structure. The first parallel path is configured with a 1×1 convolutional layer, the second parallel path is configured with a 1×1 convolutional layer and a 3×3 convolutional layer in sequence, the third parallel path is configured with a 1×1 convolutional layer and two consecutive 3×3 convolutional layers in sequence, and the fourth parallel path is configured with a 3×3 pooling layer and a 1×1 convolutional layer in sequence. The initial feature map is reduced in dimensionality and reorganized by a 1×1 convolutional layer in the first parallel path to obtain the first-dimensional feature. The initial feature map is compressed by channel compression through a 1×1 convolutional layer in the second parallel path, and then local detail-related features are captured by a 3×3 convolutional layer to obtain the second-dimensional features. The initial feature map is filtered by a 1×1 convolutional layer in the third parallel path, and then global correlation features are captured by two consecutive 3×3 convolutional layers to obtain the third dimension feature. The initial feature map is downsampled and compressed using a 3×3 pooling layer in the fourth parallel path, and then channel optimization is performed using a 1×1 convolutional layer to obtain the fourth dimension feature. The first dimension feature, the second dimension feature, the third dimension feature and the fourth dimension feature are respectively subjected to batch normalization processing to obtain the standardized first dimension feature, the standardized second dimension feature, the standardized third dimension feature and the standardized fourth dimension feature; The standardized first-dimensional features, standardized second-dimensional features, standardized third-dimensional features, and standardized fourth-dimensional features are fused using a channel splicing operation to obtain a fused feature map.

3. The method as described in claim 1, characterized in that, The step of converting the fused feature map into a set of vector capsules by inputting it into the primary capsule layer includes: Obtain the number of capsules in the primary capsule layer and the capsule dimension corresponding to each capsule. The number of capsules is determined based on the spatial size calculation result of the convolution operation and the number of capsule types. The capsule dimension is a fixed value. The spatial size calculation result of the convolution operation refers to the product of the height and width of the output feature map after the input feature map is processed by the convolution layer. The number of capsule types is the number of capsule types that exist in parallel at each spatial location in the primary capsule layer. The multi-scale features of the fused feature map are mapped using a convolutional layer of a preset size to obtain a multi-channel feature map. Based on the capsule dimension of the primary capsule layer, the multi-channel feature map is dimensionally split to obtain feature blocks corresponding to the number of capsules; Each feature block is subjected to vector transformation to obtain an initial vector corresponding to the number of capsules; Each initial vector is normalized and compressed using a nonlinear activation function to obtain a normalized vector capsule; Integrate all the normalized vector capsules to obtain a vector capsule set.

4. The method as described in claim 1, characterized in that, The step of processing the vector capsule set through a dynamic routing mechanism to obtain feature space information includes: Obtain the preset number of target capsules and vector dimensions of the digital capsule layer, and initialize the association weight matrix between the target capsules and each vector capsule in the vector capsule set; The initial association weight between each vector capsule and each target capsule in the vector capsule set is calculated based on the association weight matrix. Perform softmax normalization on all initial association weights corresponding to the same target capsule to obtain the coupling coefficient; The corresponding vector capsules in the vector capsule set are weighted and summed according to the coupling coefficient to obtain the original input vector of each target capsule; The original input vector is nonlinearly transformed by a preset compression function to obtain the intermediate output vector of the target capsule; Calculate the vector similarity between the intermediate output vector and the corresponding vector capsule, and iteratively update the coupling coefficient based on the similarity to obtain the change in the coupling coefficient, until a preset number of iterations is reached or the change in the coupling coefficient is less than a preset threshold, to obtain the optimized target capsule output vector; By integrating all the optimized target capsule output vectors, feature space information containing spatial relationships between features is obtained.

5. The method as described in claim 1, characterized in that, The step of inputting the feature space information into the digital capsule layer for classification processing and outputting the UAV recognition result includes: Obtain preset parameters for the digital capsule layer, wherein the preset parameters include the number of digital capsules and the vector dimension of each digital capsule, and the number of digital capsules corresponds to the number of drone categories set in the identification task; The feature space information is assigned features according to the vector dimension to obtain category-related features corresponding to each digital capsule; The category-related features of each digital capsule are linearly transformed by the feature mapping matrix of the digital capsule layer to obtain the category feature vector corresponding to each digital capsule. Calculate the magnitude of the feature vector for each category and obtain the predicted probability value for the corresponding UAV category, wherein the magnitude value is positively correlated with the predicted probability value; Integrate the predicted probability values ​​of all drone categories to generate a category predicted probability set; The category with the highest predicted probability value is selected from the predicted probability set to obtain the drone identification result.

6. The method as described in claim 1, characterized in that, The step of performing short-time Fourier transform and logarithmic transform on the raw radio frequency signal of the UAV to obtain the radio frequency time-frequency diagram includes: The raw radio frequency signal of the drone is sliced ​​into multiple continuous signal segments. A Hamming window is applied to each of the signal segments to obtain windowed signal segments. Perform a local Fourier transform on each of the signal segments to obtain the corresponding local Fourier transform result; The windowing and local Fourier transform operations are repeated on the unprocessed signal regions in all the signal segments according to the preset step size sliding Hamming window to obtain the full local Fourier transform results; All the full local Fourier transform results are concatenated in chronological order to form a two-dimensional time-frequency matrix and obtain the joint distribution characteristics of the signal; A logarithmic transformation is performed on the joint distribution characteristics to map the values ​​of the two-dimensional time-frequency matrix to a preset dynamic range, thereby obtaining a radio frequency time-frequency diagram.

7. A UAV radio frequency signal identification device based on an improved capsule network model, characterized in that, The device includes: The acquisition module is used to collect raw radio frequency signals of drones at different distances. Raw radio frequency signals of drones refer to the electromagnetic signals transmitted by drones through the communication link during flight without any processing. They include the drone's communication commands and status data, and are the core data basis for identifying drone types. They will exhibit unique frequencies and amplitudes depending on the drone model and flight status. The processing module is used to perform short-time Fourier transform and logarithmic transform on the raw radio frequency signal of the UAV to obtain the radio frequency time-frequency diagram; The result module is used to input the radio frequency time-frequency map into the improved capsule network model for feature extraction and classification to obtain the UAV recognition result. The improved capsule network model includes a 7×7 convolutional layer, an improved Inception feature extraction module, a primary capsule layer, and a digital capsule layer. The improved Inception feature extraction module includes three improved Inception structures. The module is also used to input the radio frequency time-frequency map into the 7×7 convolutional layer of the improved capsule network model for preliminary feature extraction to obtain an initial feature map; process the initial feature map through the improved Inception structure of the improved Inception feature extraction module to obtain a fused feature map; input the fused feature map into the primary capsule layer to convert it into a set of vector capsules; process the set of vector capsules through a dynamic routing mechanism to obtain feature space information; input the feature space information into the digital capsule layer for classification processing and output the UAV recognition result.

8. A drone radio frequency signal identification device based on an improved capsule network model, characterized in that, The device includes: a memory, a processor, and a drone radio frequency signal identification program based on an improved capsule network model stored in the memory and running on the processor, the drone radio frequency signal identification program based on the improved capsule network model being configured to implement the steps of the drone radio frequency signal identification method based on an improved capsule network model as described in any one of claims 1-6.

9. A storage medium, characterized in that, The storage medium stores a UAV radio frequency signal identification program based on an improved capsule network model. When the UAV radio frequency signal identification program based on the improved capsule network model is executed by the processor, it implements the steps of the UAV radio frequency signal identification method based on the improved capsule network model as described in any one of claims 1-6.

Images