Unmanned aerial vehicle radio frequency fingerprinting method, apparatus, medium, and device

Abstract

The application provides a UAV radio frequency fingerprint identification method and device, medium and equipment, relates to the technical field of UAV identity recognition, and comprises the following steps: collecting a radio frequency signal of a to-be-identified UAV; acquiring time domain features, modulation domain features and code domain features of the radio frequency signal; calculating domain contribution degrees corresponding to the time domain features, the modulation domain features and the code domain features; performing dynamic fusion weight distribution according to the domain contribution degrees corresponding to the time domain features, the modulation domain features and the code domain features and the signal complexity of the radio frequency signal to obtain radio frequency signal fusion features; and inputting the radio frequency signal fusion features into a local attention residual network to obtain an identification result of the to-be-identified UAV. The application realizes the dual improvement of UAV identification accuracy and robustness in a complex scene, and effectively meets the technical requirements of low-altitude control in terms of wide coverage, high accuracy and strong anti-interference.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicle identity recognition, in particular to an unmanned aerial vehicle radio frequency fingerprint identification method and device, a medium and an electronic equipment. BACKGROUND

[0002] With the deep penetration of Internet of Things technology in the field of low-altitude traffic control, unmanned aerial vehicles as typical Internet of Things terminals have been widely used in civil aerial photography, logistics transportation, power inspection and other scenarios. Unmanned aerial vehicle technology deeply integrated into the city operation system is driving the development of low-altitude economy. At the same time, it also intensifies the safety risks caused by unmanned aerial vehicle black flight and misflight.

[0003] To cope with the above-mentioned low-altitude safety risks, current researchers have developed unmanned aerial vehicle detection and identification solutions covering radar monitoring, visual recognition, acoustic detection and other technical paths. However, all kinds of traditional technologies have significant limitations in complex actual scenarios, and it is difficult to meet the comprehensive control requirements of wide coverage, high accuracy, strong anti-interference and low delay. In view of the limitations of traditional technologies, radio frequency (RF) signal analysis methods with the core advantages of non-line-of-sight, all-weather and low interference have gradually become a research hotspot in the field of unmanned aerial vehicle control. Unmanned aerial vehicles need to maintain communication with ground remote controllers and satellite navigation systems (such as GPS / Beidou) throughout the flight cycle. The RF signals emitted by them contain rich device identity information, including subtle distortions caused by hardware manufacturing differences (such as crystal oscillator drift, power amplifier nonlinearity), communication protocol characteristics (such as signal frame structure, modulation method), working parameters (such as transmission power, channel bandwidth), etc. These information constitutes the radio frequency fingerprint that can uniquely identify the unmanned aerial vehicle. Compared with traditional technologies, the radio frequency detection system passively listens to the RF signals in the airspace without actively transmitting electromagnetic waves. It can capture signals in advance before the unmanned aerial vehicle enters the visual / acoustic detection range, and the warning time is 5-10 times longer than visual recognition. It is not affected by light, weather, and terrain obstructions, and can still work stably in complex scenarios such as night, rain, and urban dense areas. The cost of a single SDR (Software Defined Radio) receiving device is only a few thousand yuan, and it is flexible and convenient to deploy. Early radio frequency signal classification relies on Fourier transform, Welch power spectrum estimation, wavelet transform and other artificial feature extraction methods, combined with traditional machine learning algorithms such as support vector machine (SVM) and random forest. However, artificial feature design relies on professional knowledge and is difficult to cover all signal distortion characteristics, and the feature dimension is high, the computational overhead is large, and in the case of multiple unmanned aerial vehicle signal superposition and low signal-to-noise ratio, the recognition accuracy is easily affected by noise and multipath effect interference.

[0004] With the development of deep learning technology, Convolutional Neural Networks (CNNs) and their improved architecture, namely Residual Networks, have provided a new solution for radio frequency (RF) signal recognition. Residual Networks, through their cross-layer skip connection design, directly transfer shallow features to deep networks, effectively alleviating the gradient vanishing and performance degradation problems in deep network training. They can deeply mine subtle distortion features in RF signals and automatically learn multi-level features of signals through end-to-end training without the need for manual feature design. Under the same test conditions, recognition accuracy can be improved, and robustness to noise and multipath effects is significantly enhanced. However, existing ResNet-based RF recognition schemes still have problems such as insufficient local feature capture and weak generalization ability with small samples. Summary of the Invention

[0005] To address the aforementioned technical problems, this application provides a method, apparatus, medium, and electronic device for radio frequency fingerprinting of unmanned aerial vehicles (UAVs), which at least partially solves the problems existing in the prior art.

[0006] In a first aspect of this application, a method for radio frequency fingerprinting of unmanned aerial vehicles is provided, the method comprising the following steps:

[0007] Collect radio frequency signals of the drone to be identified;

[0008] Obtain the time-domain, modulation-domain, and code-domain characteristics of the radio frequency signal;

[0009] Calculate the domain contribution of time-domain features, modulation-domain features, and code-domain features; where the domain contribution is determined based on the ratio of inter-class distance to intra-class distance of the corresponding feature domain; intra-class distance represents the average distance of features of several UAV samples of the same class within the corresponding feature domain; inter-class distance represents the average distance of features of several UAV samples of different classes within the corresponding feature domain;

[0010] Dynamic fusion weights are assigned based on the domain contribution values ​​corresponding to time-domain features, modulation-domain features, and code-domain features, as well as the signal complexity of the radio frequency signal, to obtain the fusion features of the radio frequency signal; wherein, the signal complexity of the radio frequency signal is determined based on the energy fluctuation ratio and frequency fluctuation ratio of the radio frequency signal.

[0011] The fused features of radio frequency signals are input into a local attention residual network to obtain the identification result of the drone to be identified. The local attention residual network is obtained by embedding a local attention module into the original residual network after the second convolution. The local attention module divides the radio frequency signal features into several local regions according to the preset number of drone types through adaptive local mean pooling.

[0012] In a second aspect of this application, a radio frequency fingerprint recognition device for unmanned aerial vehicles is provided, the device comprising:

[0013] The acquisition unit is used to acquire the radio frequency signals of the drone to be identified;

[0014] an acquisition unit configured to acquire time domain features, modulation domain features and code domain features of the radio frequency signal;

[0015] a calculation unit configured to calculate domain contribution degrees corresponding to the time domain features, the modulation domain features and the code domain features, wherein the domain contribution degrees are determined according to a ratio of inter-class distance and intra-class distance in a corresponding feature domain, the intra-class distance representing an average distance of a plurality of same-class UAV sample features in the corresponding feature domain, and the inter-class distance representing an average distance of a plurality of non-same-class UAV sample features in the corresponding feature domain;

[0016] a fusion unit configured to perform dynamic fusion weight distribution according to the domain contribution degrees corresponding to the time domain features, the modulation domain features and the code domain features, and a signal complexity of the radio frequency signal, to obtain radio frequency signal fusion features, wherein the signal complexity of the radio frequency signal is determined according to an energy fluctuation ratio and a frequency fluctuation ratio of the radio frequency signal;

[0017] a recognition unit configured to input the radio frequency signal fusion features into a local attention residual network to obtain a recognition result of the UAV to be recognized, wherein the local attention residual network is obtained by embedding a local attention module into an original residual network after a second convolution, and the local attention module divides the radio frequency signal features into a plurality of local regions according to a preset number of UAV categories through adaptive local mean pooling.

[0018] In a third aspect of the present application, a non-transitory computer readable storage medium is provided, and the storage medium stores at least one instruction or at least one program, and the at least one instruction or the at least one program is loaded and executed by a processor to implement the foregoing UAV radio frequency fingerprint identification method.

[0019] In a fourth aspect of the present application, an electronic device is provided, which includes a processor and the foregoing non-transitory computer readable storage medium.

[0020] The present application has at least the following beneficial effects:

[0021] The unmanned aerial vehicle radio frequency fingerprint identification method provided by the application firstly provides reliable data sources for subsequent identification based on the collection of unmanned aerial vehicle radio frequency signals; secondly, by synchronously acquiring time domain, modulation domain and code domain features, the limitation that artificial feature extraction only depends on a single dimension and is difficult to cover full signal distortion information is broken through, multi-dimensional radio frequency fingerprint information such as hardware manufacturing differences (such as crystal oscillator drift, power amplifier nonlinearity), communication protocol characteristics (such as frame structure, modulation mode) is comprehensively included in the analysis, laying a feature foundation for accurate identification; thirdly, the domain contribution degree is calculated based on the ratio of intra-class distance (average distance of features of samples of the same class) and inter-class distance (average distance of features of samples of different classes), which can objectively quantify the discrimination ability of each feature domain and avoid invalid feature interference, and meanwhile, dynamic fusion weight distribution is performed in combination with signal complexity determined by energy fluctuation ratio (reflecting the degree of energy time-varying) and frequency fluctuation ratio (reflecting the degree of frequency domain stability), so that the contribution proportion of each feature domain can be adaptively adjusted according to the actual state of the signal (such as multi-unmanned aerial vehicle superposition, low signal-to-noise ratio scene), the problem of insufficient robustness of fixed weight fusion in a complex scene is solved, and the fused features can focus on high discrimination information; finally, by embedding a local attention module after the second convolution of the original residual network, the local area is divided according to the preset number of unmanned aerial vehicle types by using adaptive local mean pooling, the defects of insufficient local feature capture of the existing ResNet architecture are targetedly made up, the mining ability of signal subtle distortion features is strengthened, the small sample generalization performance is improved, and finally the unmanned aerial vehicle recognition accuracy and robustness in a complex scene are both improved, effectively meeting the low-altitude control technology requirements of wide coverage, high accuracy and strong anti-interference. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0023] Figure 1 The flowchart of the unmanned aerial vehicle radio frequency fingerprint identification method provided by the embodiment of the application is shown in the figure.

[0024] Figure 2 The structural block diagram of the unmanned aerial vehicle radio frequency fingerprint identification device provided by the embodiment of the application is shown in the figure. DETAILED DESCRIPTION

[0025] With reference to the drawings and embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work belong to the scope of protection of the present application.

[0026] It should be noted that the terms "first", "second" and the like in the description and claims of the present application and the above drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or a chronological sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or server including a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0027] It should be noted that various aspects of the embodiments described below are within the scope of the appended claims. It should be apparent that the aspects described herein can be embodied in a wide variety of forms and that any specific structure and / or function described herein is merely illustrative. Based on the present application, a person of ordinary skill in the art should understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, an apparatus can be implemented and / or a method can be practiced using any number of the aspects set forth herein. In addition, this apparatus and / or method can be implemented and / or practiced using other structure and / or functionality in addition to or other than one or more of the aspects set forth herein.

[0028] Reference is made to Figure 1 As shown in the drawings, the embodiments of the present application provide a method for identifying a UAV by radio frequency fingerprinting, the method comprising the following steps:

[0029] S100, collecting a radio frequency signal of a UAV to be identified.

[0030] A radio frequency (RF) signal is a high-frequency electromagnetic wave signal radiated into space through an antenna for wireless communication. It is the carrier of information transmitted by wireless devices (mobile phones, WiFi, Bluetooth, satellites, radar, etc.). The RF signal emitted by a UAV contains rich device identity information, including subtle distortions caused by hardware manufacturing differences (such as crystal oscillator drift, power amplifier nonlinearity), communication protocol characteristics (such as signal frame structure, modulation method), operating parameters (such as transmit power, channel bandwidth), etc. These information constitutes a radio frequency fingerprint that can uniquely identify a UAV.

[0031] S200, acquire the time domain feature, the modulation domain feature and the code domain feature of the radio frequency signal.

[0032] Specifically, step S200 further comprises:

[0033] S210, digitally quadrature demodulate the radio frequency signal to obtain I / Q two-way digital zero intermediate frequency signals; and then normalize to obtain normalized I / Q two-way digital zero intermediate frequency signals.

[0034] Here, digital quadrature demodulation refers to the process of decomposing the collected radio frequency digital signal into in-phase component (I component) and quadrature component (Q component), through which the carrier frequency of the signal can be removed to obtain a zero intermediate frequency signal; I / Q two-way digital zero intermediate frequency signals refer to digital signals representing the in-phase and quadrature directions respectively after digital quadrature demodulation, without carrier frequency; normalization refers to the process of adjusting the amplitudes of I / Q two-way digital zero intermediate frequency signals to a unified range, eliminating signal amplitude fluctuations caused by gain differences of different devices.

[0035] The originally collected radio frequency signal contains a carrier frequency, which is not conducive to subsequent feature extraction. By removing the carrier frequency through digital quadrature demodulation, a zero intermediate frequency signal is obtained, which can simplify the feature extraction process; the receiving gain of different SDR devices may differ, resulting in different amplitudes of the radio frequency signals collected from the same unmanned aerial vehicle. Normalization can eliminate the effects of such device differences and ensure the consistency of the signals, providing a stable data foundation for subsequent multi-domain feature extraction.

[0036] S220, according to the normalized I / Q two-way digital zero intermediate frequency signals, extract the time domain feature, the modulation domain feature and the code domain feature.

[0037] Among them, the modulation domain feature is extracted according to the following steps:

[0038] Step a1, according to the communication frame structure, the normalized I / Q two-way digital zero intermediate frequency signals are segmented and extracted to obtain data codes.

[0039] Here, the communication frame structure refers to the organization form of the signal in time when the unmanned aerial vehicle communicates with external devices, usually including preamble, synchronization code, data code, check code and other parts; the data code refers to the symbol sequence in the communication frame structure that carries the actual data of the unmanned aerial vehicle control instruction, state information, etc.

[0040] First, the communication frame structure parameters (such as frame length, number of symbols in each part) corresponding to the commonly used communication protocols (such as WiFi, Bluetooth, and special data transmission protocol) of the unmanned aerial vehicle are obtained, then the frame synchronization algorithm (such as the synchronization algorithm based on preamble correlation) is used to find the starting position and ending position of the communication frame in the normalized I / Q two-way digital zero intermediate frequency signal, the signal is divided into preamble, synchronization code, data code and other parts according to the frame structure, and finally the data code part is extracted to obtain the data code.

[0041] It should be noted that the unmanned aerial vehicle radio frequency signal is transmitted in units of communication frames, the data code is the part containing the core information in the communication frame, and the modulation characteristics of the data code are closely related to the unmanned aerial vehicle hardware (such as modulator). The manufacturing difference of the hardware will cause slight deviation in the modulation characteristics of the data code. Extracting the data code can focus on the key modulation information that reflects the identity of the unmanned aerial vehicle.

[0042] And exclude the preamble, synchronization code and other parts of the communication frame which have high standardization degree and are difficult to reflect the device difference, and focus on the data code which can reflect the key part of the hardware difference.

[0043] Step a2, obtaining a symbol-level constellation diagram according to the data code; wherein the symbol-level constellation diagram contains a plurality of constellation points, and each constellation point represents the I / Q value of a symbol.

[0044] Here, the symbol-level constellation diagram refers to the graph formed by taking the I component value of each symbol in the data code as the horizontal coordinate and the Q component value as the vertical coordinate in the rectangular coordinate system; the constellation point refers to each point in the symbol-level constellation diagram, each point uniquely corresponds to the I / Q value of a symbol, reflecting the modulation state of the symbol.

[0045] The data code obtained in step a1 is subjected to symbol synchronization to determine the starting and ending positions of each symbol, and then the I component value and the Q component value of each symbol are extracted. Taking the I component value as the horizontal coordinate and the Q component value as the vertical coordinate, points are plotted one by one in the rectangular coordinate system to form a symbol-level constellation diagram. Each point in the constellation diagram is a constellation point, representing the I / Q value of the corresponding symbol.

[0046] Step a3, filtering the constellation points according to the local density of each constellation point and the preset density threshold to obtain a plurality of key constellation points; wherein the local density of the key constellation point is greater than the preset density threshold.

[0047] Here, the local density refers to the number of other constellation points within a preset radius range centered on a certain constellation point, reflecting the density of constellation points around the constellation point; the preset density threshold refers to a density value determined according to statistical analysis of a large number of unmanned aerial vehicle sample symbol-level constellation diagrams, used to filter out representative constellation points; the key constellation point refers to a constellation point with a local density greater than the preset density threshold, which is a constellation point with a higher frequency of occurrence in the data code and can stably reflect the modulation characteristics.

[0048] First, a preset radius (determined according to the distribution range of the symbol-level constellation diagram, ensuring that the adjacent constellation points around a certain constellation point can be covered) and a preset density threshold (by statistically analyzing the symbol-level constellation diagrams of multiple models of unmanned aerial vehicles, calculating the average value of the local density of each constellation point, and taking the average value as the preset density threshold) are set; then for each constellation point in the symbol-level constellation diagram, the number of other constellation points within a preset radius range centered on the constellation point is calculated, i.e. the local density of the constellation point is obtained; finally, the local density of each constellation point is compared with the preset density threshold, and the constellation points with a local density greater than the preset density threshold are retained to obtain a number of key constellation points.

[0049] The calculation method of the local density can be any density calculation method known to those skilled in the art.

[0050] It should be noted that in the symbol-level constellation diagram, some constellation points may occasionally appear due to noise, interference and other factors, with a low local density, and these constellation points cannot stably reflect the modulation characteristics of the unmanned aerial vehicle; while the constellation points with a high local density are the constellation points corresponding to frequently occurring symbols in the data code, which can stably reflect the modulation characteristics of the unmanned aerial vehicle, and filtering out the low local density constellation points can remove invalid information caused by noise and interference.

[0051] Step a4, according to the minimum distance between each key constellation point and other key constellation points and the local density of each key constellation point, determine the clustering score of each key constellation point, and sort them in descending order of clustering score, select a number of key constellation points as clustering centers; wherein the clustering score of the key constellation point is proportional to the minimum distance between the key constellation point and other key constellation points; the clustering score of the key constellation point is proportional to the local density of the key constellation point.

[0052] Here, for each key constellation point, the Euclidean distance from it to all other key constellation points is calculated, and the minimum distance value is selected as the minimum distance of the key constellation point. The clustering score = local density x minimum distance, and the clustering score of each key constellation point is calculated. All key constellation points are sorted in descending order of clustering score, and according to the constellation point distribution of the symbol-level constellation and experience (such as the number of clusters is usually close to the number of standard constellation points of the modulation mode), a number of key constellation points with high ranking are selected as cluster centers.

[0053] It should be noted that the key constellation point with high local density indicates that the modulation state it represents appears frequently in the data code and has strong representativeness; the key constellation point with large minimum distance indicates that it is significantly different from other key constellation points and can effectively distinguish different modulation states; the key constellation point with high clustering score obtained by comprehensively considering the two has high representativeness and strong distinguishability, and is suitable to be used as a cluster center to provide reliable core points for subsequent clustering analysis.

[0054] Step a5, according to the nearest neighbor assignment algorithm, each key constellation point is assigned to the corresponding cluster center to obtain a plurality of cluster clusters.

[0055] Specifically, the nearest neighbor assignment algorithm refers to an algorithm that assigns each key constellation point to be assigned to the class in which the cluster center with the shortest distance is located.

[0056] For each key constellation point obtained in step a4, the Euclidean distance from it to all cluster centers is calculated, and the cluster center with the smallest distance is found. The key constellation point is assigned to the class corresponding to the cluster center, and the process is repeated until all key constellation points are assigned, and finally a plurality of sets with cluster centers as cores, i.e. a plurality of cluster clusters are formed.

[0057] The key constellation points in the same cluster cluster are distributed around the same cluster center, which indicates that they represent similar modulation states and are the embodiment of the same modulation characteristics; through the nearest neighbor assignment algorithm, key constellation points with similar modulation characteristics can be classified into a class, which is convenient for subsequent extraction of features reflecting the modulation characteristics of the class.

[0058] Step a6, feature extraction is performed according to the constellation points in the plurality of cluster clusters to obtain modulation domain features.

[0059] Here, as an example: the modulation domain features can include: the I component mean, the Q component mean, the I component standard deviation, the Q component standard deviation, the deviation value of the center coordinates of the cluster cluster and the standard modulation constellation point coordinates, etc. of all constellation points in the cluster cluster; all these features of the cluster clusters are combined to form a feature vector, and the feature vector is the modulation domain feature.

[0060] It should be noted that the hardware manufacturing differences of different UAVs (such as crystal oscillator drift, power amplifier nonlinearity) will cause the mean, standard deviation of the I / Q components of the clustering cluster, and the deviation value from the standard constellation point to be different. These differences are the embodiment of the unique radio frequency fingerprint of the UAV in the modulation domain. Extracting these features can obtain modulation domain features that can distinguish different UAVs.

[0061] Code domain features are extracted according to the following steps:

[0062] Step b1, frame synchronization is performed on the normalized I / Q two-way digital zero intermediate frequency signal to obtain a complete communication frame.

[0063] Here, frame synchronization refers to finding the starting position and ending position of the communication frame in the continuous radio frequency signal, and realizing the process of dividing the signal by frame; a complete communication frame refers to a communication frame that contains all components such as the preamble, synchronization code, data code, and check code, and is the basic unit of UAV radio frequency signal transmission.

[0064] Specifically, a frame synchronization algorithm based on the preamble is used. First, the known sequence of the preamble in the UAV communication protocol (such as a fixed 0, 1 alternating sequence) is obtained. Then, the normalized I / Q two-way digital zero intermediate frequency signal is processed by a sliding window, and the correlation between the signal in the sliding window and the known preamble sequence is calculated. When the correlation value reaches a preset threshold, it is determined that the signal in the sliding window is the preamble, and the starting position of the preamble is the starting position of the communication frame. Then, according to the communication frame length specified in the communication protocol, the signal of the corresponding length is cut off from the starting position, and a complete communication frame is obtained.

[0065] Step b2, extract the preamble and synchronization code in the complete communication frame.

[0066] Here, the preamble refers to a section of symbols with fixed sequence characteristics at the beginning of the communication frame, which is used to realize frame synchronization, carrier synchronization, etc.; the synchronization code refers to a specific symbol sequence located after the preamble and before the data code in the communication frame, which is used to further realize symbol synchronization and ensure that the receiving end can accurately identify the boundary of each symbol. Some synchronization codes also contain auxiliary data such as frame type and channel information.

[0067] According to the preamble length (such as 16 symbols, 32 symbols) specified in the UAV communication protocol, the symbol sequence of the corresponding length is cut off from the starting position of the complete communication frame, and the preamble is obtained. Then, according to the position relationship (such as the synchronization code immediately following the preamble) and the synchronization code length of the synchronization code in the protocol, the symbol sequence of the corresponding length is cut off from the end of the preamble, and the synchronization code is obtained. As an example: if the protocol specifies that the preamble length is 32 symbols and the synchronization code length is 16 symbols, the preamble is extracted from the 1st-32nd symbols of the complete communication frame, and the synchronization code is extracted from the 33rd-48th symbols.

[0068] The preamble is mainly used for frame synchronization, but its sequence characteristics are affected by the subtle differences in the hardware of the UAV (such as the signal generator), and small sequence deviations may occur. The synchronization code is the key basis for symbol synchronization, and its symbol characteristics are closely related to the precision of the clock circuit of the UAV. Differences in the clock circuits of different UAVs can cause subtle shifts in the symbol boundaries of the synchronization code. Extracting the preamble and the synchronization code can capture these hardware difference-related code domain features.

[0069] Step b3, according to the UAV communication protocol, the preamble and the synchronization code, generate the standard preamble sequence and the standard synchronization code sequence.

[0070] Here, the standard preamble sequence refers to a theoretically unbiased preamble sequence generated according to the preamble format and sequence rules specified by the communication protocol followed by the UAV; the standard synchronization code sequence refers to a theoretically unbiased synchronization code sequence generated according to the synchronization code format and sequence rules specified by the corresponding communication protocol.

[0071] First, determine the communication protocol used by the UAV to be identified, then consult the technical documents of the protocol to obtain the standard sequence format of the preamble and the standard sequence format of the synchronization code, and finally generate a standard preamble sequence consistent with the length of the extracted preamble in step b2, and a standard synchronization code sequence consistent with the length of the extracted synchronization code. As an example, if the WiFi protocol specifies that the preamble is a 128-bit pseudo-random sequence, then generate the 128-bit standard preamble sequence; if the synchronization code is a 16-bit fixed sequence 0101010101010101, then generate the standard synchronization code sequence.

[0072] The standard preamble sequence and the standard synchronization code sequence are ideal unbiased sequences specified by the communication protocol, while the actual extracted preamble and synchronization code from the UAV signal may deviate from the standard sequence due to hardware manufacturing differences (such as signal generator precision, clock jitter). The deviation between the actual sequence and the standard sequence is a unique code domain fingerprint of the UAV, and generating the standard sequence can provide a reference for quantifying this deviation.

[0073] Step b4, according to the standard preamble sequence and the standard synchronization code sequence, perform feature extraction to obtain the code domain features.

[0074] Among them, the code domain features can include calculating deviation features, calculating correlation features, and calculating timing features, etc.

[0075] Here, the calculation of the deviation feature is to compare the actual preamble sequence extracted in step b2 with the standard preamble sequence symbol by symbol, calculate the number of inconsistent symbols (i.e. the number of errors) and the error rate (the number of errors / the total number of preamble symbols); in the same way, the number of errors and the error rate of the actual synchronization code sequence and the standard synchronization code sequence are calculated; the calculation of the correlation feature is to calculate the correlation coefficient (reflecting the degree of similarity between the two, the value range [-1, 1]) of the actual preamble sequence and the standard preamble sequence by using the cross-correlation algorithm; the correlation coefficient of the actual synchronization code sequence and the standard synchronization code sequence is also calculated; the calculation of the timing feature is to determine the starting time of each symbol in the actual preamble and synchronization code by using the symbol synchronization algorithm, calculate the deviation value of the duration of each symbol from the standard symbol duration (determined according to the baud rate of the communication protocol), and take the mean and standard deviation of all deviation values.

[0076] The hardware differences of different drones (such as the symbol generation accuracy of the signal generator and the jitter degree of the clock circuit) will cause the error rate, correlation coefficient and symbol timing deviation of the actual preamble / synchronization code and the standard sequence to be different - for example, the symbol duration deviation will be larger and the error rate may be higher for drones with large clock jitter; these differences are the unique identification of drones in the code domain, and extracting these features can obtain code domain features that can effectively distinguish different drones.

[0077] The extracted code domain features are directly related to the hardware differences of drones, have strong distinguishing ability, can be complementary to the time domain and modulation domain features, further enrich the dimensions of the radio frequency fingerprints of drones, and improve the accuracy of subsequent identification.

[0078] S300, calculate the domain contribution degree of the time domain feature, the modulation domain feature and the code domain feature; wherein the domain contribution degree is determined according to the ratio of the inter-class distance and the intra-class distance in the corresponding feature domain; the intra-class distance represents the average distance between the feature vectors of a plurality of samples of the same type (same class) of drones in the corresponding feature domain; the inter-class distance represents the average distance between the feature vectors of a plurality of samples of different types (non-same class) of drones in the corresponding feature domain.

[0079] Specifically, the domain contribution degree refers to a quantitative index of the contribution of the features in a certain feature domain (time domain, modulation domain, code domain) to the identification of different classes of drones in the process of drone identification; the larger the domain contribution degree, the stronger the distinguishing ability of the features in the feature domain; the intra-class distance refers to the average distance between the feature vectors of a plurality of samples of the same type (same class) of drones in a certain feature domain, reflecting the similarity of the features of the same class of samples; the inter-class distance refers to the average distance between the feature vectors of a plurality of samples of different types (non-same class) of drones in a certain feature domain, reflecting the difference between the features of different classes of samples.

[0080] First, a plurality of samples of different models of unmanned aerial vehicles are collected, and a plurality of samples are collected for each model. Time domain feature vectors, modulation domain feature vectors and code domain feature vectors are extracted from each sample to form three sample feature libraries (time domain sample library, modulation domain sample library and code domain sample library).

[0081] Taking the time domain feature domain as an example, the Euclidean distance between any two vectors of a plurality of time domain feature vectors of each model is calculated to obtain a plurality of distance values of the model, and the average of the distance values is taken as the time domain intra-class distance of the model. The average of the time domain intra-class distances of all models is taken to obtain the intra-class distance of the time domain feature domain. In the same way, the intra-class distances of the modulation domain feature domain and the code domain feature domain are calculated.

[0082] When calculating the inter-class distance, still taking the time domain feature domain as an example, two different models are randomly selected from the plurality of models, the Euclidean distance between each time domain feature vector of the first model and each time domain feature vector of the second model is calculated to obtain a plurality of distance values, and the average is taken as the time domain inter-class distance of the two models. The time domain inter-class distance of each combination is calculated by traversing all model combinations, and the average of all combination distances is taken to obtain the inter-class distance of the time domain feature domain. In the same way, the inter-class distances of the modulation domain feature domain and the code domain feature domain are calculated.

[0083] The contribution degree of the time domain feature domain = inter-class distance / intra-class distance; because the larger the inter-class distance, the more significant the feature difference between different classes of unmanned aerial vehicles, and the smaller the intra-class distance, the more similar the features of unmanned aerial vehicles in the same class, the larger the ratio of inter-class distance / intra-class distance, the higher the contribution degree. The ratio of the two can intuitively quantify the distinguishing ability of the feature domain for different classes of unmanned aerial vehicles, provide a basis for subsequent dynamic fusion weight allocation, and avoid allocating weights to feature domains with weak distinguishing ability, which may lead to a decrease in the quality of the fused features.

[0084] S400, dynamically allocating weights according to the domain contribution degrees corresponding to the time domain features, the modulation domain features and the code domain features, and the signal complexity of the radio frequency signal to obtain a radio frequency signal fusion feature; wherein the signal complexity of the radio frequency signal is determined according to the energy fluctuation ratio and the frequency fluctuation ratio of the radio frequency signal.

[0085] Specifically, step S400 further comprises:

[0086] S410, determining a gating threshold for each feature domain according to the signal complexity of the radio frequency signal; wherein the gating threshold of the time domain feature is inversely proportional to the signal complexity; the gating threshold of the modulation domain feature is proportional to the signal complexity; and the gating threshold of the code domain feature is proportional to the signal complexity.

[0087] The stability of each feature domain is different in different scenarios. For example, in a high complexity scenario (such as multi-signal superposition), the time domain feature is easily disturbed by noise, and the stability decreases, while the modulation domain and code domain features are bound to the hardware depth, and the stability is relatively high; in a low complexity scenario, the stability of the time domain feature is high, and its distinguishing ability can be fully utilized. Dynamic weight allocation is performed in combination with the domain contribution (inherent distinguishing ability of the feature) and the signal complexity (current stability of the feature), so that the fused feature can always focus on the feature with strong distinguishing ability and current stability, and the interference of the feature with poor stability to the fusion result is avoided.

[0088] The gating threshold refers to a quantitative index for controlling the threshold for participating in fusion of the feature domain, and the value range is (0, 1). The lower the threshold value, the easier the feature domain participates in fusion (the higher the weight ratio), and the higher the threshold value, the higher the threshold for the feature domain to participate in fusion (the lower the weight ratio).

[0089] In one embodiment, in a high complexity scenario, the time domain feature is easily disturbed by noise and multi-signal, and the stability decreases, so the gating threshold is reduced to reduce the weight participating in fusion; while the modulation domain and code domain features are bound to the hardware depth (such as modulation deviation and code timing deviation), and have strong anti-interference ability and high stability, so the gating threshold is increased to increase the weight participating in fusion; in a low complexity scenario, the stability of the time domain feature is high, and the gating threshold needs to be increased to fully play a role.

[0090] The specific numerical setting can be determined according to the experience of those skilled in the art.

[0091] S420, obtaining a dynamic weight of each feature domain according to the gating threshold of each feature domain and the domain contribution of each feature domain; wherein the dynamic weight is proportional to the domain contribution; and the dynamic weight is inversely proportional to the gating threshold.

[0092] Specifically, the dynamic weight ; wherein d is a feature domain, d=a is a time domain; d=b is a modulation domain; d=e is a code domain; is a regulation coefficient, to avoid meaningless minimum value of the denominator when the signal complexity D=0. d is the domain contribution; T d is the gating threshold.

[0093] The above formula binds the distinguishing ability of the feature itself to the adaptability of the current scenario through an exponential regulation factor, and finally obtains the actual effective contribution value of the feature in the current signal, which retains the distinguishing advantage and adapts to the scenario demand. It is ensured that the feature with strong distinguishing ability and adaptability to the current scenario obtains a higher effective weight, and otherwise the weight is reduced to avoid invalid features interfering with the fusion effect.

[0094] The higher the domain contribution degree is, the stronger the inherent distinguishing ability of the feature is. The lower the gating threshold is, the higher the current stability of the feature is. The combination of the two ensures that the fused feature focuses on the core effective information. The normalization processing ensures that the sum of the weights is 1, satisfying the mathematical logic of the fusion calculation.

[0095] The embodiment combines the inherent distinguishing ability of the feature with the current scene stability, realizes accurate allocation of the dynamic weight, avoids the limitation of the traditional fixed weight, and ensures that the fused feature has strong distinguishing ability and high reliability in different complexity scenes.

[0096] The signal complexity is determined according to the following steps:

[0097] S401, the normalized I / Q two-way digital zero intermediate frequency signal is pre-processed frame by frame, and the average energy of each preset frame is calculated.

[0098] Specifically, the preset frame refers to the pre-set frame length (such as 1024 sampling points, 2048 sampling points) according to the signal processing efficiency and feature extraction accuracy requirement, which is used to divide the continuous signal into several short periods for energy calculation in the period; the average energy of each preset frame refers to the average value of the energy of all sampling points in a preset frame, which reflects the energy level of the signal in the period.

[0099] Firstly, the preset frame length (such as 1024 sampling points) and the frame shift (i.e. step, such as 512 sampling points, to avoid information loss between frames) are set, and the normalized I / Q two-way digital zero intermediate frequency signal is divided into frames by using the sliding window method and is rounded up to obtain several preset frames. Secondly, for each sampling point in each preset frame, the energy is calculated, wherein the energy is the sum of the square of the I component value of the sampling point and the square of the Q component value of the sampling point. Finally, for each preset frame, the average value of the energy of all sampling points in the frame is calculated to obtain the average energy sequence of all preset frames.

[0100] Here, the energy fluctuation of the signal is an important embodiment of its time characteristic. In high complexity scenarios (such as multiple signal superposition, noise interference), the energy of the signal will fluctuate sharply. By calculating the average energy by frame, the continuous energy change can be converted into a discrete energy sequence, which is convenient for subsequent quantification of the energy fluctuation degree.

[0101] S402, according to the average energy of each preset frame, the energy fluctuation ratio is calculated; wherein the energy fluctuation ratio represents the time-varying amplitude of the energy of the radio frequency signal.

[0102] Here, the energy fluctuation ratio refers to the ratio of the difference between the maximum average energy and the minimum average energy in the preset frame average energy sequence to the average value of the average energy sequence, which is used to quantify the change amplitude of the signal energy in time. The larger the ratio is, the more intense the energy time variation is.

[0103] The energy fluctuation ratio is in the form of extreme difference / mean value, which eliminates the influence of the overall energy level of the signal and only focuses on the relative fluctuation amplitude of the energy. As an example: two signals with different overall energy, if their extreme difference / mean value ratio is the same, it means that the energy fluctuation intensity of the two signals is consistent; this index can objectively reflect the time-varying characteristics of the signal energy, and is an important dimension for measuring the complexity of the signal.

[0104] S403, the instantaneous frequency of the normalized I / Q two-path digital zero intermediate frequency signal is obtained.

[0105] Here, the instantaneous frequency refers to the instantaneous frequency value of the radio frequency signal at each sampling point, which reflects the instantaneous change of the signal frequency with time, and is different from the carrier frequency (fixed value) of the signal. The instantaneous frequency will fluctuate due to factors such as hardware distortion and noise interference. The present embodiment uses an instantaneous frequency estimation algorithm based on Hilbert transform to obtain the instantaneous frequency.

[0106] The instantaneous frequency of the radio frequency signal of the unmanned aerial vehicle is affected by the hardware (such as the crystal oscillator) and external interference. Crystal oscillator drift will cause slow fluctuations in the instantaneous frequency, and noise interference will cause rapid fluctuations in the instantaneous frequency; obtaining the instantaneous frequency can capture these frequency change characteristics and provide a basis for quantifying the frequency fluctuation degree.

[0107] S404, the frequency fluctuation ratio is calculated according to the instantaneous frequency; wherein the frequency fluctuation ratio represents the frequency domain fluctuation amplitude of the radio frequency signal.

[0108] Here, the frequency fluctuation ratio refers to the ratio of the difference between the maximum instantaneous frequency and the minimum instantaneous frequency in the instantaneous frequency sequence to the mean value of the instantaneous frequency sequence, which is used to quantify the change amplitude of the signal frequency over time. The larger the ratio, the more intense the frequency domain fluctuation.

[0109] The frequency fluctuation ratio is also in the form of extreme difference / mean value to eliminate the influence of the carrier frequency of the signal itself, and only focuses on the relative fluctuation amplitude of the frequency. As an example: two kinds of unmanned aerial vehicle signals with carrier frequencies of 2.4GHz and 5.8GHz, if their frequency fluctuation ratios are both 0.05, it means that the frequency domain fluctuation intensity of the two signals is the same. This index can objectively reflect the time-varying characteristics of the signal frequency, and together with the energy fluctuation ratio, it constitutes the basic dimension of the signal complexity.

[0110] S405, the time-frequency coupling fluctuation coefficient is obtained according to the energy sequence and the frequency sequence of the normalized I / Q two-path digital zero intermediate frequency signal; wherein the time-frequency coupling fluctuation coefficient represents the coupling correlation between energy and frequency.

[0111] Here, the energy and frequency of the UAV radio frequency signal are not independent fluctuations, and hardware defects (such as power amplifier nonlinearity) will cause energy and frequency coupling correlation (for example, when the output power of the power amplifier increases, the output frequency will be slightly offset due to the temperature rise of the device), and external interference (such as multi-signal superposition) will also exacerbate this coupling. The time-frequency coupling fluctuation coefficient can capture this unique coupling characteristic, which is different from the traditional single energy or frequency fluctuation evaluation, and can more comprehensively reflect the signal complexity.

[0112] Specifically, the average value of all instantaneous frequencies in each frame is calculated according to the preset frame length and frame shift of step S401.

[0113] S406, the energy fluctuation ratio, the frequency fluctuation ratio and the time-frequency coupling fluctuation coefficient are weighted and summed to obtain the signal complexity.

[0114] Here, the three indicators are weighted and summed using preset weights. Among them, the preset weight is determined according to a large number of sample experiments, and according to the calculation result, the complexity level is divided according to the preset threshold. The weight of the time-frequency coupling fluctuation coefficient is the largest.

[0115] The energy fluctuation ratio and the frequency fluctuation ratio reflect the degree of fluctuation of the signal in a single dimension, and the time-frequency coupling fluctuation coefficient reflects the correlation strength between dimensions, which together constitute a complete evaluation system for signal complexity from the two aspects of single dimension fluctuation and cross-dimension correlation. The weight of the time-frequency coupling fluctuation coefficient is the highest because the time-frequency coupling characteristic can better reflect the depth of the signal affected by hardware or interference.

[0116] S500, input the radio frequency signal fusion feature into the local attention residual network to obtain the identification result of the to-be-identified UAV; wherein the local attention residual network is obtained by embedding a local attention module into the original residual network after the second convolution; the local attention module divides the radio frequency signal feature into several local regions according to a preset number of UAV categories through adaptive local mean pooling.

[0117] The radio frequency signal fusion feature refers to the feature map (dimension HxWxC, H is the height, W is the width, and C is the number of channels) obtained by weighting and combining the time domain, modulation domain and code domain features according to the step dynamic weight. Although the original residual network can alleviate the gradient disappearance problem, it lacks attention to local key features (such as clustering bias in modulation domain features and time sequence bias in code domain features), resulting in weak small sample generalization ability. Embedding a local attention module after the second convolution is because the feature map still retains a lot of detailed information at this stage, and by dividing the local area according to the number of categories, the network can focus on the local features directly related to "category differentiation", strengthen the expression of key features, and suppress irrelevant feature interference.

[0118] Specifically, step S500 further comprises:

[0119] S510, divide the spatial dimension of the radio frequency signal fusion feature into several local areas according to a preset number of unmanned aerial vehicle categories.

[0120] Among them, let the spatial dimension of the radio frequency signal fusion feature map be HxW, and the preset number of unmanned aerial vehicle categories be M. Then, H is divided into M equal segments along the height direction (each segment has a height of H / M, and if H cannot be divided by M, edge zero padding or trimming is used to adjust it to be divisible), and W remains unchanged along the width direction, to obtain M local areas, each with a dimension of (H / M)xWxC.

[0121] S520, calculate the mean value of each channel in each local area according to global adaptive mean pooling.

[0122] For the M local areas, each with a dimension of (H / M)xWxC, for the cth channel (c=1 to C) of each area, the mean value of all (H / M)xW pixel values in the channel is calculated to obtain the mean value of the cth channel of the area. All C channels are traversed to form a 1x1xC mean vector for each area, and finally M 1x1xC mean vectors are obtained.

[0123] The channel mean value of each local area can represent the feature strength of the channel in the area. For example, a certain channel corresponds to a clustering bias in the modulation domain feature, and a high channel mean value indicates that the area contains a strong modulation bias feature. Calculating the channel mean value can extract the core feature information of each area, reduce the computational power caused by the spatial dimension, and at the same time retain the feature distinguishability of the channel dimension.

[0124] S530, obtain the mean value of the global maximum pooling layer.

[0125] For the M 1x1xC mean vectors obtained in step S520, for each channel c (c=1 to C), the maximum value is selected from the channel mean values of the M areas to form a 1x1xC global maximum mean vector [G_c] (c=1 to C). Each value in the vector is the mean value of the global maximum pooling layer for the corresponding channel.

[0126] The mean value of the global maximum pooling layer can represent the strongest feature level of each channel in the entire fusion feature map. For example, a high global maximum mean value of a certain channel indicates that the feature corresponding to the channel (such as the bit error rate bias in the code domain) is prominent in a certain local area and is a key feature for distinguishing unmanned aerial vehicle categories. Obtaining this mean value can serve as a "baseline reference" for subsequent channel weight distribution, ensuring that the weight is tilted towards strong feature channels.

[0127] S540, obtain the weight of each channel according to the fully connected layer.

[0128] The 1 × 1 × C global maximum mean vector obtained in step S530 is flattened into a 1 × C feature vector as the input of the fully connected layer, the fully connected layer includes two layers of neurons, the number of neurons in the first layer is set to C / 4 (for example, when C = 64, the number of neurons in the first layer is 16, and the calculation amount is reduced by dimension reduction), and the activation function uses ReLU; the number of neurons in the second layer is set to C (consistent with the number of channels), and the activation function uses Sigmoid (the output value is compressed to the range of [0, 1]); through the forward propagation calculation of the fully connected layer, the input feature vector of 1 × C is mapped to the output vector of 1 × C, and each element in the output vector is the weight [W c ] (c = 1 to C) of the corresponding channel, for example, the third element W 3 is the weight of the third channel.

[0129] The fully connected layer can learn the association between the channel features and the recognition result through training. As an example: if the fifth channel corresponds to the clustering deviation feature in the modulation domain, and this feature is crucial for distinguishing between category 1 and category 2, the fully connected layer will assign a higher weight to the fifth channel after training; the Sigmoid activation function ensures that the weight is within the range of [0, 1], which facilitates subsequent dot multiplication operations with channel features.

[0130] S550, multiply the weight of each channel with the corresponding input feature to obtain the enhanced radio signal fusion feature.

[0131] The 1 × C channel weight vector obtained in step S540 is expanded into an H × W × C weight matrix consistent with the dimension of the fusion feature map (the weight value of each channel is repeated H × W times to form the weight matrix of this channel); for each channel c (c = 1 to C) of the original fusion feature map, multiply the H × W pixel values of the channel with the corresponding H × W weight values element by element to obtain the weighted channel feature; combine all C weighted channel features in the original channel order to form the enhanced radio signal fusion feature map (the dimension remains H × W × C unchanged).

[0132] S560, according to the enhanced radio signal fusion feature, obtain the recognition result of the unmanned aerial vehicle to be identified.

[0133] The enhanced radio signal fusion feature map is input into the subsequent layers (the third convolutional layer to the fully connected output layer) of the local attention residual network; the subsequent convolutional layers continue to extract deeper features, the global average pooling layer compresses the feature map of H × W × C into a feature vector of 1 × C, and finally the Softmax activation function of the output layer (the number of neurons = the preset number of unmanned aerial vehicle categories M) outputs the probability value of each category (the sum of the probability values of all categories is 1); select the category with the maximum probability value as the recognition result of the unmanned aerial vehicle to be identified, as an example: if the probability value of category 1 is 0.92 and is the maximum value among all categories, then the unmanned aerial vehicle to be identified is determined to be category 1.

[0134] The preset number of UAV categories is the target number that the network needs to distinguish. Dividing the feature map into a corresponding number of local regions can allow each region to correspond to the potential key features of a category of UAV. As an example: the first region focuses on the modulation deviation features of the first category of UAV, and the second region focuses on the symbol timing deviation features of the second category of UAV, to realize the association mapping of the region and the category, and guide the network to learn the key features of each category of UAV. The enhanced fusion feature map has highlighted the key features, and the subsequent network layer can more efficiently mine the mapping relationship between the features and the categories; the Softmax activation function converts the network output into a probability distribution, which is convenient for intuitively judging the confidence of the to-be-identified signal belonging to a category of UAV, and the category with the highest probability is the most likely identification result.

[0135] In an exemplary embodiment of the present application, first, the initial time-frequency features are extracted, and the short-time Fourier transform (STFT) converts the normalized time-domain signal into a two-dimensional time-frequency spectrum. The STFT is a multi-window, and in an embodiment, the number of windows is 3, each window is a single-channel time-frequency map with pixels of 224x224x1, forming a multi-window time-frequency feature group. Specifically, 3 STFTs with different window sizes are used to process the normalized I / Q signals in parallel, generating 3 types of time-frequency maps, which capture features of different dimensions: small window STFT (window size = 256 samples, overlap rate = 50%): high time resolution, low frequency resolution, good at capturing transient features, such as instantaneous frequency jumps caused by changes in UAV attitude; medium window STFT (window size = 512 samples, overlap rate = 50%): balanced time / frequency resolution, capturing medium-term trend features, such as stable frequency bands during hovering; large window STFT (window size = 1024 samples, overlap rate = 50%): low time resolution, high frequency resolution, capturing steady-state features, such as fixed frequency distortions caused by power amplifier nonlinearity.

[0136] The 3 single-channel time-frequency maps are fused, specifically as follows:

[0137] First, feature concatenation: concatenate the 3 single time-frequency channel time-frequency maps along the channel dimension to form a multi-time-frequency channel time-frequency map of 224x224x3 (channel 1 = small window, channel 2 = medium window, channel 3 = large window), and second, cross-channel attention weight distribution: use the SE (Squeeze-and-Excitation) attention mechanism to dynamically distribute the weights of the 3 time-frequency channels. Through global average pooling, the 224x224x3 features are compressed into time-frequency channel statistics of 1x1x3, and then through two fully connected layers (dimensions 3→1→3) and a Sigmoid activation function, the weight of each time-frequency channel is generated.

[0138] After that, the full connection layer and the Sigmoid activation are performed to generate the weight of each time-frequency channel and to strengthen the high-weight time-frequency channel:

[0139] Then, feature fusion is performed: the generated time-frequency channel weight is multiplied with the 224x224x3 multi-time-frequency channel time-frequency graph point by point to strengthen the features of the high-weight time-frequency channel and to suppress the noise of the low-weight time-frequency channel.

[0140] Finally, time-frequency channel compression is performed: the fused multi-channel features are compressed into 224x224x32 through a 1x1 convolution kernel to reduce the subsequent calculation amount and to realize the information interaction between channels.

[0141] Finally, the fused time-frequency feature graph in the form of 224x224x32 is output, which is the final time-frequency feature.

[0142] It should be noted that the above time-frequency channel refers to the time-frequency channel obtained through multi-window during time-frequency feature extraction, which is different from the channel in the above residual network.

[0143] Please refer to Figure 2 The embodiment of the application provides a kind of unmanned aerial vehicle radio frequency fingerprint identification device, the device includes:

[0144] Acquisition unit 110 is used to collect the radio frequency signal of the unmanned aerial vehicle to be identified.

[0145] Acquisition unit 120 is used to acquire the time domain feature, modulation domain feature and code domain feature of the radio frequency signal.

[0146] The calculation unit 130 is used to calculate the domain contribution degree corresponding to the time domain feature, the modulation domain feature and the code domain feature;Wherein, the domain contribution degree is determined according to the ratio of the inter-class distance and the intra-class distance of the corresponding feature domain;The intra-class distance represents the average distance of the features of several same-class unmanned aerial vehicle samples in the corresponding feature domain;The inter-class distance represents the average distance of the features of several non-same-class unmanned aerial vehicle samples in the corresponding feature domain.

[0147] Fusion unit 140 is used for dynamic fusion weight distribution according to the domain contribution degree corresponding to the time domain feature, the modulation domain feature and the code domain feature and the signal complexity of the radio frequency signal to obtain the radio frequency signal fusion feature;Wherein, the signal complexity of the radio frequency signal is determined according to the energy fluctuation ratio and the frequency fluctuation ratio of the radio frequency signal.

[0148] Identification unit 150 is used to input the radio frequency signal fusion feature into the local attention residual network to obtain the identification result of the unmanned aerial vehicle to be identified;Wherein, the local attention residual network is obtained by embedding a local attention module into the original residual network after the second convolution;The local attention module divides the radio frequency signal feature into several local areas according to the preset number of unmanned aerial vehicle categories through adaptive local mean pooling.

[0149] Embodiments of the present application also provide a computer program product, which comprises program codes for causing an electronic device to perform the steps of the methods according to various exemplary embodiments of the present application described above in the specification when the program product is run on the electronic device.

[0150] In addition, although various steps of the methods in the present application are described in a particular order in the accompanying drawings, this does not require or imply that the steps must be performed in this particular order, or that all of the illustrated steps must be performed to achieve the desired results. Additionally or alternatively, some steps can be omitted, a plurality of steps can be combined into one step, one step can be divided into a plurality of steps, and the like.

[0151] From the above description of the embodiments, those skilled in the art will readily appreciate that the example embodiments described herein can be implemented by software and / or by hardware combined with software that is necessary for the embodiments of the present application. Accordingly, the technical solutions according to the embodiments of the present application can be embodied in the form of a software product. The software product can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, or the like) or on a network, and includes a number of instructions for causing a computing device (which can be a personal computer, a server, a mobile terminal, or a network device, etc.) to perform the methods according to the embodiments of the present application.

[0152] In the example embodiments of the present application, an electronic device capable of implementing the above-described methods is also provided.

[0153] Those skilled in the art can understand that various aspects of the present application can be implemented as a system, a method, or a program product. Therefore, various aspects of the present application can be embodied in the form of a complete hardware implementation, a complete software implementation (including firmware, microcode, etc.), or a combination of hardware and software aspects, which can be collectively referred to as "circuitry", "module", or "system" herein.

[0154] The electronic device according to this embodiment of the present application. The electronic device is merely an example, and should not impose any limitation on the functions and use range of the embodiments of the present application.

[0155] The electronic device is in the form of a general computing device. The components of the electronic device can include, but are not limited to, the at least one processor described above, the at least one storage described above, and a bus connecting different system components (including the storage and the processor).

[0156] The storage stores program codes that can be executed by the processor, so that the processor performs the steps according to various exemplary embodiments of the present application described in the above "example method" section of the specification.

[0157] The storage can include a readable medium in the form of volatile storage such as random access memory (RAM) and / or cache memory, and can further include a non-volatile storage such as read only memory (ROM).

[0158] The storage can also include a program / utility, having a set of program modules that include, but are not limited to: an operating system, one or more application programs, other program modules, and program data, each of which or a combination thereof, can include an implementation of a networking environment.

[0159] The bus can represent one or more of several types of bus structures, including a storage bus or bus controller, a peripheral bus, a graphics bus, a processor or local bus using any of a variety of bus architectures.

[0160] The electronic device can also communicate with one or more external devices such as a keyboard or a pointing device, through an I / O interface. Additionally, the electronic device can communicate with one or more devices that enable a user to interact with the electronic device through an input device or devices 110. In the illustrated embodiment, the input device or devices 110 include a keyboard and a pointing device such as a mouse. In some embodiments, the input device or devices 110 include a touch screen, a microphone, an input port, etc. The electronic device can also include an output device, such as a display, a speaker, etc. The electronic device can further include one or more communication devices, such as a transmitter, a receiver, a transceiver, a modem, a LAN card, a wireless communication device, etc. The communication devices enable the electronic device to communicate with one or more other devices over a one or more communication coupling or bus structures. In these and other embodiments, therefore, the electronic device includes communication devices that enable the electronic device to communicate with one or more other computing devices, such as over a network.

[0161] Those skilled in the art will readily understand that the example embodiments described herein can be implemented by software and / or by software in combination with the necessary hardware. Thus, the technical solutions according to the embodiments of the present application can be embodied in the form of a software product. The software product can be stored in a non-volatile storage medium (which can be a CD-ROM, a U disk, a mobile hard disk, etc.) or a network, and includes a number of instructions to enable a computing device (which can be a personal computer, a server, a terminal device, or a network device, etc.) to perform the methods according to the embodiments of the present application.

[0162] In the example embodiments of the present application, a computer readable storage medium is also provided, on which a program product capable of implementing the method described above is stored. In some possible implementation manners, various aspects of the present application can also be implemented in the form of a program product, which includes program codes for causing the terminal device to perform the steps described in the "example method" section above according to various example embodiments of the present application when the program product is run on the terminal device.

[0163] The program product can adopt any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium may, for example, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or apparatus, or any combination thereof. More specific examples (non-exhaustive list) of the readable storage medium include an electrical connection having one or more wires, a portable disc, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0164] The computer readable signal medium can include a data signal propagated in a baseband or as part of a carrier wave, in which readable program codes are borne. Such a propagated data signal can take on many forms, including but not limited to an electromagnetic signal, an optical signal, or any suitable combination thereof. The readable signal medium can also be any readable medium that is not a readable storage medium and that can send, propagate or transmit the program for use by or in connection with an instruction execution system, apparatus or device.

[0165] The program codes contained on the readable medium can be transmitted by any suitable medium, including but not limited to wireless, wired, optical cable, RF, etc., or any suitable combination thereof.

[0166] The program code may be executed by one or more programmable processing devices to perform the operations described herein. The processing devices may include processors, microprocessors, microcontrollers, programmable logic devices, field programmable gate arrays (FPGAs), or other processing devices. The processing devices may be configured to execute program code stored in memory. The memory may include one or more types of memory such as volatile memory (e.g., random access memory (RAM), dynamic random access memory (DRAM), or static random access memory (SRAM)), non-volatile memory (e.g., read only memory (ROM), programmable read only memory (PROM), erasable programmable read only memory (EPROM), electrically erasable programmable read only memory (EEPROM), flash memory, solid state memory, or other non-volatile memory), or other types of memory.

[0167] Furthermore, the above-described diagrams are only schematic and are non-limiting. It is expressly contended that the processes set forth in the diagrams can be performed in an order other than that depicted. Additionally, it is expressly contemplated that the processes set forth in the diagrams can be performed synchronously or asynchronously, e.g., in a plurality of modules.

[0168] It should be noted that, although reference is made in the above detailed description to several modules or units of the device for action execution, such a division is not mandatory. Indeed, according to an implementation of the present application, the features and functionalities of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functionalities of one module or unit described above can be further divided into several modules or units embodied by.

[0169] The above merely illustrates the specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any changes or replacements within the technical scope disclosed by the present application should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for radio frequency fingerprint recognition of unmanned aerial vehicles, characterized in that, The method includes: Collect radio frequency signals of the drone to be identified; Obtain the time-domain, modulation-domain, and code-domain characteristics of the radio frequency signal; Calculate the domain contribution of time-domain features, modulation-domain features, and code-domain features; where the domain contribution is determined based on the ratio of inter-class distance to intra-class distance of the corresponding feature domain; intra-class distance represents the average distance of features of several UAV samples of the same class within the corresponding feature domain; inter-class distance represents the average distance of features of several UAV samples of different classes within the corresponding feature domain; Dynamic fusion weights are assigned based on the domain contribution values ​​corresponding to time-domain features, modulation-domain features, and code-domain features, as well as the signal complexity of the radio frequency signal, to obtain the fusion features of the radio frequency signal; wherein, the signal complexity of the radio frequency signal is determined based on the energy fluctuation ratio and frequency fluctuation ratio of the radio frequency signal. The fused features of radio frequency signals are input into a local attention residual network to obtain the identification result of the drone to be identified. The local attention residual network is obtained by embedding a local attention module into the original residual network after the second convolution. The local attention module divides the radio frequency signal features into several local regions according to the preset number of drone types through adaptive local mean pooling.

2. The UAV radio frequency fingerprinting method according to claim 1, characterized in that, The acquisition of the time-domain features, modulation-domain features, and code-domain features of the radio frequency signal includes: The radio frequency signal is digitally quadrature demodulated to obtain two digital zero intermediate frequency signals, I and Q; then normalization is performed to obtain normalized two digital zero intermediate frequency signals, I and Q. Based on the normalized I / Q two-channel digital zero-IF signals, time-domain features, modulation-domain features, and code-domain features are extracted.

3. The UAV radio frequency fingerprinting method according to claim 2, characterized in that, The modulation domain features are extracted according to the following steps: The normalized I / Q two-channel digital zero-IF signals are segmented and extracted according to the communication frame structure to obtain the data code; Based on the data code, a symbol-level constellation diagram is obtained; the symbol-level constellation diagram contains several constellation points, and each constellation point represents the I / Q value of a symbol. The constellation points are filtered based on their local density and a preset density threshold to obtain several key constellation points; where the local density of the key constellation points is greater than the preset density threshold. Based on the minimum distance between each key constellation point and other key constellation points and the local density of each key constellation point, a cluster score for each key constellation point is determined, and the key constellation points are sorted in descending order of their cluster scores. Several key constellation points are then selected as cluster centers. The cluster score of a key constellation point is proportional to the minimum distance between that key constellation point and other key constellation points, and the cluster score of a key constellation point is proportional to its local density. Based on the nearest neighbor assignment algorithm, each key constellation point is assigned to the corresponding cluster center to obtain several clusters; Feature extraction is performed on constellation points in several clusters to obtain modulation domain features.

4. The UAV radio frequency fingerprinting method according to claim 1, characterized in that, The code field features are extracted according to the following steps: Frame synchronization is performed on the normalized I / Q two-channel digital zero-IF signals to segment and obtain a complete communication frame; Extract the preamble and synchronization code from the complete communication frame; Based on the UAV communication protocol, preamble, and synchronization code, generate standard preamble and standard synchronization code sequences; Feature extraction is performed based on the standard preamble sequence and the standard synchronization code sequence to obtain code domain features.

5. The UAV radio frequency fingerprinting method according to claim 2, characterized in that, The dynamic fusion weight allocation based on the domain contribution degrees corresponding to time-domain features, modulation-domain features, and code-domain features, as well as the signal complexity of the radio frequency signal, to obtain the fused radio frequency signal features includes: The gate threshold for each feature domain is determined based on the signal complexity of the radio frequency signal; among them, the gate threshold for the time domain feature is inversely proportional to the signal complexity; the gate threshold for the modulation domain feature is directly proportional to the signal complexity; and the gate threshold for the code domain feature is directly proportional to the signal complexity. The dynamic weight of each feature domain is obtained based on the gating threshold and the domain contribution of each feature domain; the dynamic weight is directly proportional to the domain contribution and inversely proportional to the gating threshold.

6. The UAV radio frequency fingerprinting method according to claim 5, characterized in that, The signal complexity of an radio frequency (RF) signal is determined by the following steps: The normalized I / Q two-channel digital zero intermediate frequency signals are divided into preset frames, and the average energy of each preset frame is calculated. The energy fluctuation ratio is calculated based on the average energy of each preset frame; where the energy fluctuation ratio represents the time-varying amplitude of the radio frequency signal's energy. Obtain the instantaneous frequency of the normalized I / Q two-channel digital zero-IF signals; The frequency fluctuation ratio is calculated based on the instantaneous frequency; where the frequency fluctuation ratio represents the frequency domain fluctuation amplitude of the radio frequency signal. The time-frequency coupling fluctuation coefficient is obtained from the energy sequence and frequency sequence of the normalized I / Q two-channel digital zero-IF signals; whereby the time-frequency coupling fluctuation coefficient represents the correlation between energy and frequency coupling. The signal complexity is obtained by weighted summation of the energy fluctuation ratio, frequency fluctuation ratio, and time-frequency coupling fluctuation coefficient.

7. The UAV radio frequency fingerprinting method according to claim 1, characterized in that, The step of inputting the fused radio frequency signal features into a local attention residual network to obtain the identification result of the drone to be identified includes: The spatial dimension of the fused radio frequency signal features is divided into several local regions according to the preset number of UAV types; The mean value of each channel within each local region is calculated using global adaptive mean pooling. Obtain the mean of the global max pooling layer; The weight of each channel is obtained from the fully connected layer; The weight of each channel is multiplied by the corresponding input feature to obtain the enhanced RF signal fusion feature. Based on the enhanced radio frequency signal fusion characteristics, the identification result of the drone to be identified is obtained.

8. A radio frequency fingerprint recognition device for unmanned aerial vehicles, characterized in that, The device includes: The acquisition unit is used to acquire the radio frequency signals of the drone to be identified; The acquisition unit is used to acquire the time-domain characteristics, modulation-domain characteristics, and code-domain characteristics of the radio frequency signal; The calculation unit is used to calculate the domain contribution of time-domain features, modulation-domain features, and code-domain features. The domain contribution is determined based on the ratio of inter-class distance to intra-class distance of the corresponding feature domain. The intra-class distance represents the average distance of features of several UAV samples of the same class in the corresponding feature domain. The inter-class distance represents the average distance of features of several UAV samples of different classes in the corresponding feature domain. The fusion unit is used to dynamically allocate fusion weights based on the domain contribution rates corresponding to the time domain features, modulation domain features, and code domain features, as well as the signal complexity of the radio frequency signal, in order to obtain the fusion features of the radio frequency signal; wherein, the signal complexity of the radio frequency signal is determined based on the energy fluctuation ratio and frequency fluctuation ratio of the radio frequency signal. The recognition unit is used to input the fused features of radio frequency signals into the local attention residual network to obtain the recognition result of the drone to be identified. The local attention residual network is obtained by embedding a local attention module into the original residual network after the second convolution. The local attention module divides the radio frequency signal features into several local regions according to the preset number of drone types through adaptive local mean pooling.

9. A non-transitory computer-readable storage medium, characterized in that, The storage medium stores at least one instruction or at least one program segment, which is loaded and executed by a processor to implement the method as described in any one of claims 1-7.

10. An electronic device, characterized in that, Includes a processor and the non-transitory computer-readable storage medium as described in claim 9.

Images