Unmanned aerial vehicle identification method and device, electronic equipment and storage medium

Abstract

The invention relates to the field of artificial intelligence, and provides an unmanned aerial vehicle identification method and device, electronic equipment and a storage medium, and the method comprises the steps: obtaining an electromagnetic signal of a to-be-identified unmanned aerial vehicle; based on a feature extraction model, extracting unmanned aerial vehicle features corresponding to the electromagnetic signals; the feature extraction model comprises a feature extraction module, and the feature extraction module is used for performing state space modeling by applying time sequence features and spatial features of the electromagnetic signals, and constructing unmanned aerial vehicle features by applying long-range features obtained by the state space modeling; matching the unmanned aerial vehicle features with registration features of each unmanned aerial vehicle in an unmanned aerial vehicle library, and determining an identification result of the to-be-identified unmanned aerial vehicle based on a matching result; the registration features are obtained based on a feature extraction model. According to the method, the device, the electronic equipment and the storage medium provided by the invention, full fusion of the information of the electromagnetic signals in the time domain and the frequency domain is ensured, so that efficient and accurate unmanned aerial vehicle identification in an urban low-altitude complex spectrum environment is realized.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, electronic device, and storage medium for identifying unmanned aerial vehicles (UAVs). Background Technology

[0002] With the rapid deployment of unmanned aerial vehicles (UAVs), the detection and identification of UAVs is becoming increasingly important for maintaining airspace, infrastructure, and public privacy.

[0003] Currently, methods for detecting and identifying drones mainly include radar-based methods, visual image-based methods, sound-based methods, and remote identification (Remote ID) methods. However, in the complex low-altitude spectrum environment of cities, radar-based methods are susceptible to multipath effects caused by dense buildings and ground clutter interference; visual image-based methods are limited by changes in lighting, inclement weather, and visual obstruction; sound-based methods are easily drowned out by the rich background noise of cities; and remote identification methods, as a cooperative surveillance technology, cannot deal with non-cooperative drones that refuse to broadcast or tamper with signals.

[0004] In this context, how to effectively detect and identify drones in complex urban environments remains a problem that urgently needs to be solved in this field. Summary of the Invention

[0005] This invention provides a drone identification method, device, electronic device, and storage medium to address the shortcomings of related technologies in complex urban environments, such as the susceptibility of drone identification technology to interference and its inability to cope with non-cooperative drones.

[0006] This invention provides a method for identifying unmanned aerial vehicles (UAVs), comprising: Acquire the electromagnetic signals of the drone to be identified; Based on the feature extraction model, the UAV features corresponding to the electromagnetic signal are extracted; the feature extraction model includes a feature extraction module, which is used to perform state space modeling by applying the temporal and spatial features of the electromagnetic signal, and to construct the UAV features by applying the long-range features obtained by the state space modeling. The drone features are matched with the registration features of each drone in the drone database, and the identification result of the drone to be identified is determined based on the matching result; the registration features are obtained based on the feature extraction model.

[0007] According to a drone identification method provided by the present invention, the step of extracting drone features corresponding to the electromagnetic signal based on a feature extraction model includes: Based on the temporal extraction unit in the feature extraction module, the temporal features of the electromagnetic signal are extracted; Based on the frequency domain extraction unit in the feature extraction module, the spatial features of the frequency domain representation of the electromagnetic signal are extracted; Based on the state space unit in the feature extraction module, state space modeling is performed on the temporal features and the spatial features to obtain the long-range features; Based on the feature output unit in the feature extraction module, high-level features are constructed by combining the frequency domain representation and the long-range features. Based on the output module of the feature extraction model, the high-level features are used to construct the UAV features.

[0008] According to a UAV identification method provided by the present invention, the step of constructing high-level features based on the feature output unit in the feature extraction module, combined with the frequency domain representation and the long-range features, includes: Based on the feature output unit in the feature extraction module, the residual features of the frequency domain representation are extracted, and the high-level features are constructed by combining the long-range features, the residual features, and the frequency domain representation.

[0009] According to the UAV identification method provided by the present invention, the feature extraction model includes multiple cascaded feature extraction modules, wherein the high-level features output by the previous feature extraction module are the electromagnetic signals input by the next feature extraction module.

[0010] According to a drone identification method provided by the present invention, the step of constructing the drone features based on the high-level features using the output module of the feature extraction model includes: Based on the output module in the feature extraction model, the high-level features output by each of the multiple cascaded feature extraction modules are fused to obtain the UAV features.

[0011] According to the UAV identification method provided by the present invention, the training steps of the feature extraction model include: Acquire the electromagnetic signals of the sample drone; Based on the initial model, the sample UAV features of the sample electromagnetic signals are extracted; Based on the differences in features between sample drones belonging to the same sample drone, and the differences in features between sample drones belonging to different sample drones, the initial model is iterated to obtain the feature extraction model.

[0012] According to a drone identification method provided by the present invention, the sample electromagnetic signal includes the original electromagnetic signal and the electromagnetic signal after data enhancement of the original electromagnetic signal.

[0013] The drone identification method provided by the present invention further includes: Acquire the electromagnetic signals of the drone to be registered; Based on the feature extraction model, the drone features corresponding to the electromagnetic signals of the drone to be registered are extracted; Based on the drone characteristics corresponding to the electromagnetic signals of the drone to be registered, the registration characteristics of the drone to be registered are determined. The registration features and drone identifier of the drone to be registered are stored in the drone database.

[0014] The present invention also provides a drone identification device, comprising: Acquisition unit, used to acquire electromagnetic signals of the drone to be identified; The inference unit is used to extract UAV features corresponding to the electromagnetic signal based on the feature extraction model; the feature extraction model includes a feature extraction module, which is used to perform state space modeling by applying the temporal and spatial features of the electromagnetic signal, and to construct the UAV features by applying the long-range features obtained by the state space modeling. The identification unit is used to match the drone features with the registration features of each drone in the drone database, and determine the identification result of the drone to be identified based on the matching result; the registration features are obtained based on the feature extraction model.

[0015] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement any of the above-described drone identification methods.

[0016] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the unmanned aerial vehicle (UAV) identification method as described above.

[0017] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements any of the above-described drone identification methods.

[0018] The UAV identification method, apparatus, electronic device, and storage medium provided by this invention, due to the all-weather, stealthy, and long-range acquisition attributes of electromagnetic signal capture, make UAV identification based on electromagnetic signals more adaptable to the complex low-altitude spectrum environment of cities, and do not require UAV cooperation. Furthermore, by applying the temporal and spatial features of electromagnetic signals to perform state-space modeling through a feature extraction module, efficient inference of long sequences of electromagnetic signals can be achieved while ensuring the full fusion of information in the time and frequency domains, thereby rapidly acquiring high-quality UAV features. Based on the UAV features obtained in this way, efficient and accurate UAV identification can be achieved in complex low-altitude spectrum environments of cities. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this invention or related technologies, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is one of the flowcharts illustrating the drone identification method provided by the present invention.

[0021] Figure 2 This is a schematic diagram of the feature extraction module provided by the present invention.

[0022] Figure 3 This is a schematic diagram of the feature extraction model provided by the present invention.

[0023] Figure 4 This is a schematic diagram of the training process of the feature extraction model provided by the present invention.

[0024] Figure 5 This is a schematic diagram of the incremental learning process for unmanned aerial vehicles provided by the present invention.

[0025] Figure 6 This is the second flowchart illustrating the drone identification method provided by the present invention.

[0026] Figure 7 This is a schematic diagram of the structure of the drone identification device provided by the present invention.

[0027] Figure 8 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0029] All actions involving the acquisition of signal information or data in this invention are carried out in compliance with the relevant data protection laws and policies of the country where the device is located, and with the authorization granted by the owner of the device.

[0030] In recent years, drones have been rapidly adopted in various fields, including civilian and commercial applications, due to their low cost, high maneuverability, and ease of operation. However, with the rapid growth in the number of drones, the illegal use of drones has become increasingly serious. Although traditional regulatory methods such as electronic fences and real-name authentication have played a role in maintaining safety, these methods are insufficient to address dynamic and covert non-cooperative flight behaviors, and cannot effectively cover drones that are not registered, have been technically tampered with, or are flying across borders. Therefore, drone detection and identification technologies have become an urgent technological direction for breakthroughs.

[0031] Currently, detection and identification technologies for drones are mainly divided into four categories: radar-based methods, visual image-based methods, sound-based methods, and remote identification-based methods.

[0032] Among them, radar-based methods use echoes to infer information such as the shape and attitude of UAVs, thereby identifying different types of UAVs. They are suitable for long-range detection, are unaffected by lighting conditions, have strong anti-jamming capabilities, and can provide rich UAV parameter information. However, radar-based equipment deployment requires a wide line of sight, and in the complex low-altitude spectrum environment of cities, it is highly susceptible to multipath effects caused by dense buildings and ground clutter interference, making it difficult to effectively separate and track low, slow, and small targets.

[0033] Visual image-based methods rely on visual sensors to acquire drone images and then identify drones by analyzing features such as color and contour. However, in complex low-altitude urban environments, the detection performance of visual image-based methods is easily affected by weather, lighting changes, and visual occlusion. Furthermore, real-time processing of high-resolution images presents computational challenges, resulting in significant difficulties in achieving accurate and real-time detection of small targets.

[0034] Sound-based methods perform non-line-of-sight detection by extracting noise signals emitted by the drone's engine, rotor, and other mechanical structures. However, these methods have limited detection range and are easily drowned out by the rich background noise in urban environments, resulting in a very short effective range and significantly reduced practicality.

[0035] Remote identification-based methods are cooperative surveillance technologies that rely on drones actively broadcasting their identity and location information in accordance with regulations. This allows for direct association of drones with their registered identities, achieving high-precision, low-cost real-time identification. However, the effectiveness of this method depends entirely on the drone's compliance. In other words, this method can only handle cooperative drones that actively cooperate; it is completely ineffective against non-cooperative drones that refuse to broadcast or tamper with their signals.

[0036] Therefore, how to overcome the complexity of the urban low-altitude environment and achieve efficient and accurate detection and identification of various drones, including non-cooperative drones, remains an urgent problem to be solved in this field.

[0037] To address the above problems, embodiments of the present invention provide a method for identifying unmanned aerial vehicles (UAVs). Figure 1 This is one of the flowcharts illustrating the drone identification method provided by the present invention, such as... Figure 1 As shown, the method includes: Step 110: Obtain the electromagnetic signal of the drone to be identified.

[0038] Here, the drone to be identified is the drone that needs to be detected and identified. The electromagnetic signal of the drone to be identified is the electromagnetic signal of a specific communication frequency band in the monitored airspace captured by scanning. The monitored airspace is the airspace where the drone is monitored, which is usually the airspace where the drone may be active; the specific communication frequency band is the frequency range used by the drone's communication links such as remote control and image transmission, such as the 5.8 GHz or 2.4 GHz sub-band in the ISM (Industrial, Scientific, and Medical) band.

[0039] The electromagnetic signals obtained are electromagnetic radiation data, which can specifically be electromagnetic radiation data of communication between the drone to be identified and the remote controller, such as electromagnetic IQ (In-phase and Quadrature) data sequences.

[0040] Step 120: Based on the feature extraction model, extract the UAV features corresponding to the electromagnetic signal; the feature extraction model includes a feature extraction module, which is used to perform state space modeling by applying the temporal and spatial features of the electromagnetic signal, and to construct the UAV features by applying the long-range features obtained from the state space modeling.

[0041] Specifically, the electromagnetic signals of the drones to be identified typically have large data bandwidth and high sampling rate, resulting in a massive amount of data. For example, the electromagnetic signal bandwidth of a drone during image transmission is around 10MHz or even higher, and the sampling rate of its signal receiver is as high as 100MHz. A 10ms electromagnetic signal has a sequence length of up to 1 million IQ sampling points.

[0042] Currently, traditional CNNs (Convolutional Neural Networks) struggle to establish global contextual relationships due to their limited receptive field. While the Transformer architecture can effectively model long-distance dependencies, its complexity and memory consumption gradually become bottlenecks as sequence length increases. When performing UAV recognition based on electromagnetic signals, using an architecture like the Transformer for inference is insufficient to meet the real-time processing requirements of electromagnetic signals.

[0043] To address the aforementioned issues, this invention provides an embodiment that constructs a feature extraction model to extract features from the electromagnetic signals of a UAV. This model may include one or more feature extraction modules. The model extracts features from the input electromagnetic signals and then constructs UAV features based on the extracted features. Here, the UAV features are high-level representation vectors that characterize the UAV information reflected by the electromagnetic signals.

[0044] Furthermore, the process of extracting UAV features corresponding to electromagnetic signals based on the feature extraction model can be represented as inputting the electromagnetic signal into the feature extraction model, extracting UAV features from the electromagnetic signal by the feature extraction model, and then obtaining the UAV features output by the feature extraction model.

[0045] In this process, the feature extraction module in the feature extraction model is responsible for extracting and constructing UAV features from electromagnetic signals. Specifically, the feature extraction module can extract temporal and spatial features from electromagnetic signals, and apply the extracted temporal and spatial features to perform state-space modeling to obtain long-range features of the electromagnetic signals. Then, the long-range features are applied to construct UAV features.

[0046] Among them, the temporal feature is the representation vector of the electromagnetic signal in the time domain. The feature extraction module can directly extract features from the electromagnetic signal to obtain the temporal feature. For example, the feature extraction module can perform linear projection on the electromagnetic signal to project it into a high-dimensional space, and the features obtained from the linear projection are denoted as the temporal feature.

[0047] Spatial features are the representation vectors of electromagnetic signals in the frequency domain. The feature extraction module can first perform a time-frequency transformation on the electromagnetic signal to obtain its frequency domain representation, and then extract features from this frequency domain representation to obtain the spatial features. For example, the feature extraction module can first perform a Short-Time Fourier Transform (STFT) on the electromagnetic signal to obtain its frequency domain representation, and then perform linear projection, Layer Normalization (LN), and 2D Convolution (2DConv) on the frequency domain representation to obtain the spatial features.

[0048] Building upon this foundation, state-space modeling is performed using temporal and spatial features. Specifically, these features can be considered as input sequences for state-space modeling. Through discretized state-space equations, long-range input sequences can be processed while maintaining efficient reasoning, thereby obtaining the long-range features of the electromagnetic signal. These long-range features can be understood as feature representations that compress the global long-range context dependencies of the electromagnetic signal. For example, the temporal and spatial features can be added together and used as the input sequence for state-space modeling.

[0049] Furthermore, state-space modeling can be implemented using the State-Space Model (SSM) within the Mamba architecture. Specifically, a state-space model can be set up in the feature extraction module, using temporal and spatial features as input sequences. During this process, the state-space model effectively establishes the connection between the current input and historical information in the input sequence, thereby understanding and processing the extremely long contextual information of the input sequence, and ultimately outputting the long-range features of the electromagnetic signal. It is understandable that the application of the state-space model can replace the self-attention mechanism in the traditional Transformer architecture, exhibiting linear computational complexity and superior performance in long sequence modeling tasks. Applying state-space modeling to the feature extraction module enables efficient and high-quality processing of long sequence features that integrate the temporal and spatial features of the electromagnetic signal, thereby significantly improving the overall inference efficiency of the feature extraction model and optimizing the feature extraction quality.

[0050] After obtaining the long-range features, the feature extraction module can determine the UAV features corresponding to the electromagnetic signals based on these features. For example, the long-range features can be directly used as UAV features, or the long-range features after layer normalization can be fused with the frequency domain representation after activation by the activation function as UAV features, or the long-range features after layer normalization can be fused with the frequency domain representation as UAV features. This embodiment of the invention does not limit the specific approach.

[0051] In this embodiment of the invention, the feature extraction module implements a dual-channel, lightweight feature extraction structure for electromagnetic signals. Specifically, the feature extraction module can extract features from the electromagnetic signal in the time domain and its frequency domain representation to obtain temporal features in the time domain and spatial features in the frequency domain. Subsequently, the temporal features and spatial features are fully and effectively fused through state-space modeling to fully extract the information of the electromagnetic signal in the time and frequency domains, thereby obtaining the UAV features corresponding to the electromagnetic signal.

[0052] Compared to the traditional Mamba architecture, which fails to effectively adapt to electromagnetic signals, the feature extraction module in this embodiment of the invention achieves efficient long sequence processing by applying state space modeling, while fully and effectively fusing the temporal and spatial features of electromagnetic signals, greatly improving the quality of UAV feature extraction.

[0053] Step 130: Match the drone features with the registration features of each drone in the drone database, and determine the identification result of the drone to be identified based on the matching result; the registration features are obtained based on the feature extraction model.

[0054] Specifically, a drone database can be pre-set. This drone database is a feature database of registered drones, storing the registration features of registered drones. It is understood that for any registered drone, its registration features are obtained by inputting the drone's electromagnetic signals into a feature extraction model. For example, the drone's registration features can be drone features obtained by inputting the drone's electromagnetic signals into the feature extraction model; or, for another example, the drone's registration features can be obtained by inputting multiple electromagnetic signals of the drone into the feature extraction model respectively, obtaining multiple drone features, and then calculating the average or center point of the multiple drone features. This embodiment of the invention does not specifically limit this approach.

[0055] After obtaining the drone features corresponding to the electromagnetic signals of the drone to be identified, these features can be matched against the registration features of each registered drone in the drone database to obtain a matching result. This matching can be achieved by calculating the feature similarity between the drone feature and each registered feature in the database, and then determining whether the drone feature matches a registered feature based on the magnitude of the feature similarity. The resulting matching result may be a match with one registered feature or a mismatch with all registered features.

[0056] After obtaining the matching results, the identification result of the drone to be identified can be determined. Here, if the matching result matches one registered feature, it can be determined that the drone to be identified is the drone corresponding to the matching registered feature, that is, the identified drone is a registered drone. If the matching result does not match any registered features, it can be determined that the drone to be identified is an unregistered drone.

[0057] In the method provided in this invention, due to the all-weather, covert, and long-range acquisition attributes of electromagnetic signal capture, UAV identification based on electromagnetic signals is more adaptable to the complex low-altitude spectrum environment in urban areas and does not require UAV cooperation. Furthermore, by applying the temporal and spatial features of electromagnetic signals to perform state-space modeling through a feature extraction module, efficient inference of long sequences of electromagnetic signals can be achieved while ensuring the full fusion of information in the time and frequency domains, thereby rapidly acquiring high-quality UAV features. Based on the UAV features obtained in this way, efficient and accurate UAV identification can be achieved in complex low-altitude spectrum environments in urban areas.

[0058] Based on the above embodiments, step 120, which involves extracting the UAV features corresponding to the electromagnetic signal based on the feature extraction model, includes: Based on the temporal extraction unit in the feature extraction module, the temporal features of the electromagnetic signal are extracted; Based on the frequency domain extraction unit in the feature extraction module, the spatial features of the frequency domain representation of the electromagnetic signal are extracted; Based on the state space unit in the feature extraction module, state space modeling is performed on the temporal features and the spatial features to obtain the long-range features; Based on the feature output unit in the feature extraction module, high-level features are constructed by combining the frequency domain representation and the long-range features. Based on the output module of the feature extraction model, the high-level features are used to construct the UAV features.

[0059] Specifically, the feature extraction model may include a feature extraction module and an output module. If there is only one feature extraction module in the feature extraction model, the output of the feature extraction module is connected to the input of the output module. If there are multiple feature extraction modules in the feature extraction model, the output of each feature extraction module is connected to the input of the output module, or the output of the last feature extraction module is connected to the input of the output module. This embodiment of the invention does not impose any specific limitations on this.

[0060] Furthermore, the feature extraction module may include a temporal extraction unit, a frequency domain extraction unit, a state space unit, and a feature output unit. The outputs of the temporal extraction unit and the frequency domain extraction unit are connected to the input of the state space unit, respectively, and the output of the state space unit is connected to the input of the feature output unit.

[0061] In the process of extracting UAV features based on the feature extraction model, for any feature extraction module in the feature extraction model, each unit in the feature extraction module performs the following operations: The timing extraction unit is specifically designed to extract timing features. Electromagnetic signals can be input into the timing extraction unit, which then extracts timing features from the electromagnetic signals. The timing extraction unit can then output the timing features of the electromagnetic signals. For example, the timing extraction unit could be a linear projection unit.

[0062] A frequency domain extraction unit is a unit used to extract spatial features. The frequency domain representation of an electromagnetic signal can be input into the frequency domain extraction unit, which then extracts spatial features from this representation. Thus, the frequency domain extraction unit can output the spatial features of the electromagnetic signal. For example, a frequency domain extraction unit may include a linear projection unit, a layer normalization unit, and a two-dimensional convolution unit.

[0063] A state-space element is a unit used to implement state-space modeling. The temporal and spatial characteristics of an electromagnetic signal can be input into the state-space element, which then performs state-space modeling based on these characteristics. Thus, the state-space element can output the long-range characteristics of the electromagnetic signal. For example, a state-space element can be a State-Space Model (SSM).

[0064] The feature output unit is used to output high-level features extracted from electromagnetic signals. The frequency domain representation and long-range features of the electromagnetic signal can be input into the feature output unit, which then combines these two data points for further feature extraction, resulting in the high-level features output by the feature output unit. Here, high-level features can be understood as high-level representation vectors of the electromagnetic signal output by the feature extraction model. For example, the feature output unit may include layer normalization units, dot product units, linear projection units, etc.

[0065] After the feature extraction module outputs high-level features, these features can be input into the output module of the feature extraction model. The output module then constructs the UAV features corresponding to the electromagnetic signals based on these high-level features. For example, the output module can directly output the high-level features from a single feature extraction module as the UAV feature output, or it can fuse the high-level features output from multiple feature extraction modules as the UAV feature output. This embodiment of the invention does not impose specific limitations on these methods.

[0066] In the method provided in the embodiments of the present invention, a dual-channel feature extraction scheme is implemented in the feature extraction module by combining a time-series extraction unit, a frequency domain extraction unit, and a state space unit, thereby realizing high-level feature extraction of UAV electromagnetic signals.

[0067] Based on any of the above embodiments, in step 120, constructing high-level features based on the feature output unit in the feature extraction module, combined with the frequency domain representation and the long-range features, includes: Based on the feature output unit in the feature extraction module, the residual features of the frequency domain representation are extracted, and the high-level features are constructed by combining the long-range features, the residual features, and the frequency domain representation.

[0068] Specifically, in the feature output unit, residual features can be extracted from the frequency domain representation of the electromagnetic signal. These residual features can be features extracted from different parts of the frequency domain representation, similar to spatial features. For example, the frequency domain representation can be divided into two parts: one part is input into the frequency domain extraction unit to extract spatial features, and the other part is input into the feature output unit to extract residual features.

[0069] In the feature output unit, long-range features, residual features, and frequency domain representations can be combined to construct high-level features. For example, the residual features can be multiplied by the long-range features after layer normalization, and the resulting feature can be linearly projected and added to the frequency domain representation to obtain the high-level features.

[0070] Based on any of the above embodiments Figure 2 This is a schematic diagram of the feature extraction module provided by the present invention, as shown below. Figure 2 As shown, the feature extraction module employs a dual-channel design for electromagnetic signals and their frequency domain representation.

[0071] Among them, for electromagnetic signals Electromagnetic signals can be directly transmitted. The input is fed into the time series extraction unit to obtain the time series features. .exist Figure 2 In this context, the timing extraction unit uses electromagnetic signals. For the input linear projection unit, the temporal characteristics The extraction can be specifically expressed as the following formula: In the formula, Indicates electromagnetic signal The output obtained from the input linear projection unit. and All of these are learnable weight parameter matrices.

[0072] Frequency domain representation of electromagnetic signals The frequency domain can be directly represented. The input is fed into the frequency domain extraction unit to obtain spatial features. .exist Figure 2 In this context, the frequency domain extraction unit may include elements represented in the frequency domain. The input consists of linear projection units, layer normalization units, and two-dimensional convolution units connected in series. Furthermore, it is assumed that the frequency domain is represented... The features obtained after inputting into the linear projection unit are After that, It is divided into two parts, namely and , here will Input to the feature output unit, The input is then fed into subsequent layer normalization units and two-dimensional convolutional units to obtain spatial features. .

[0073] Spatial features The extraction can be specifically expressed as the following formula: In the formula, and These represent layer normalization and two-dimensional convolution, respectively. The kernel size of the two-dimensional convolution can be... .

[0074] The extraction can be specifically expressed as the following formula: In the formula, This represents the frequency domain representation of electromagnetic signals. The output obtained from the input linear projection unit. and All of these are learnable weight parameter matrices.

[0075] Subsequently, the temporal features output by the temporal extraction unit were analyzed. And the spatial features output by the frequency domain extraction unit. Both of these can be input into the state-space unit. The state-space unit is... Figure 2 The state-space model in [the context]. It can be... The input sequence is used as the input sequence for the state-space model to perform state-space modeling, thereby obtaining long-range features. , It can be composed of multiple consecutive moments. Composition. Here, " "" indicates element-wise addition.

[0076] in, , That is, the state-space model.

[0077] Specifically, It can be represented as , It can be composed of multiple consecutive moments. Composition. A state-space model can be modeled using a system of differential equations: In the formula, For the input sequence in Input data at any time, for The state vector at time t, for Output data at any given time. , , , The learnable weight parameters in the state-space model are the state transition matrix, input projection matrix, output projection matrix, and feedforward matrix, respectively.

[0078] in, The state equation describes the update process of the state vector representing memory. express The rate of change. The output equation describes the process of generating the output based on the input and the state vector representing the memory.

[0079] Next, the long-range characteristics output by the state-space unit are analyzed. and frequency domain representation and features after linear projection unit These three elements can be input together into the feature output unit. The feature output unit may include... Figure 2 The branches consisting of layer normalization units and activation function units, as well as the layer normalization units and linear projection units connected after the state space module.

[0080] In the feature output unit, for the features The features can be transformed sequentially through layer normalization units and activation function units, thereby obtaining... , ,in, This represents the SiLU activation function. Representation layer normalization.

[0081] In the feature output unit, for long-range features First, perform layer normalization, then... Multiplication, i.e., correspondence Figure 2 "in The '' sign indicates element-wise multiplication. Subsequently, a linear projection is performed on the result of the element-wise multiplication, and then the result of the linear projection is compared with the frequency domain representation. Add them together to obtain high-level features. .

[0082] High-level characteristics The acquisition of can be expressed as the following formula: In the formula, and These represent layer normalization and linear projection, respectively.

[0083] Based on any of the above embodiments, the feature extraction model includes multiple cascaded feature extraction modules, where the high-level features output by the previous feature extraction module are the electromagnetic signals input by the next feature extraction module.

[0084] Specifically, in a feature extraction model, multiple feature extraction modules can be set up, and these modules can be cascaded, meaning they are connected in series, with the output of the previous module serving as the input of the next.

[0085] For example, in a feature extraction model with two cascaded feature extraction modules, the first module takes the electromagnetic signal of the drone to be identified as input, extracts features from this signal, and outputs its high-level features. The second module takes the high-level features output by the first module as input, specifically treating these features as the electromagnetic signal of the drone to be identified, extracting features from this signal, and outputting its high-level features. It can be understood that the high-level features output by each of the two feature extraction modules are actually features at different levels for the same drone's electromagnetic signal.

[0086] In the method provided in this embodiment of the invention, by setting multiple cascaded feature extraction modules in the feature extraction model, high-level features at multiple levels can be obtained, thereby enhancing the feature representation extraction capability of UAV features.

[0087] Based on any of the above embodiments, in step 120, the step of constructing the UAV features by applying the high-level features based on the output module of the feature extraction model includes: Based on the output module in the feature extraction model, the high-level features output by each of the multiple cascaded feature extraction modules are fused to obtain the UAV features.

[0088] Specifically, in traditional feature extraction model design, since the features at the highest level possess rich semantic information for known types, applying these features to classify known types can achieve high accuracy, and they are typically used as the output of feature extraction. However, for data of unknown types that the feature extraction model has not yet learned, using only the features at the highest level will reduce the model's ability to represent unknown types and will also miss high-contrast features between known and unknown types.

[0089] To address the above situation and enhance the feature extraction capability of UAV features based on the output of the feature extraction model for electromagnetic signals of unknown UAVs, for cases where multiple cascaded feature extraction modules exist in the model, each feature extraction module can input its own high-level features into the output module of the feature extraction model. Thus, the output module of the feature extraction model can fuse the high-level features output by multiple cascaded feature extraction modules and output the fused features as the UAV features. It can be understood that the resulting UAV features fuse multiple levels of high-level features, thereby ensuring the ability to extract electromagnetic signals for known types of UAVs while also enhancing the ability to extract electromagnetic signals for unknown types of UAVs, thus providing conditions for the detection and identification of unauthorized UAVs.

[0090] For example, Figure 3 This is a schematic diagram of the feature extraction model provided by the present invention, as shown below. Figure 3 As shown, the feature extraction model includes two cascaded feature extraction modules, namely feature extraction module 1 and feature extraction module 2. The feature extraction model also includes an output module, namely... Figure 3 The Chinese label is " The part marked "". For example Figure 3 As can be seen, a residual structure is introduced in the feature extraction model. That is, the high-level feature 1 output by feature extraction module 1 is not only the input of feature extraction module 2, but also input together with the high-level feature 2 output by feature extraction module 2 into the output module. Through the jump connection of the above residual structure, the high-level feature 1 and high-level feature 2 are added pixel by pixel in the output module, thereby outputting the UAV feature containing high-level features of multiple levels.

[0091] In the method provided in this embodiment of the invention, by fusing the high-level features output by multiple cascaded feature extraction modules, drone features with strong characterization and extraction capabilities for electromagnetic signals of both known and unknown drones are obtained, thus providing conditions for achieving reliable and accurate drone identification.

[0092] Based on any of the above embodiments, the training steps of the feature extraction model include: Acquire the electromagnetic signals of the sample drone; Based on the initial model, the sample UAV features of the sample electromagnetic signals are extracted; Based on the differences in features between sample drones belonging to the same sample drone, and the differences in features between sample drones belonging to different sample drones, the initial model is iterated to obtain the feature extraction model.

[0093] Here, the sample electromagnetic signal refers to the electromagnetic signal of the sample UAV used to train the feature extraction model. The sample electromagnetic signal can be a directly captured electromagnetic signal or an electromagnetic signal after processing such as adding noise and cropping the directly captured electromagnetic signal. This embodiment of the invention does not make specific limitations on this.

[0094] The initial model is the model used to train the feature extraction model. The initial model has the same model structure as the feature extraction model, and the parameters of the initial model can be the parameters obtained during initialization.

[0095] Sample electromagnetic signals can be input into an initial model, thereby extracting UAV features from the sample electromagnetic signals using the initial model. In this embodiment of the invention, the UAV features from the sample electromagnetic signals extracted via the initial model are denoted as sample UAV features.

[0096] After obtaining the sample UAV features corresponding to the electromagnetic signals of each sample, the loss value required for parameter iteration of the initial model can be calculated based on this. In this embodiment of the invention, the loss value can be divided into two parts: intra-class loss and inter-class loss.

[0097] The intra-class loss refers to the loss caused by the differences between the drone features extracted from different samples of electromagnetic signals belonging to the same drone sample. It can be understood that the smaller the differences between the drone features of drone samples belonging to the same sample, the more concentrated the drone features extracted by the initial model for drone samples of the same sample, and the stronger the robustness of the initial model in extracting drone features; conversely, the greater the differences between the drone features of drone samples belonging to the same sample, the more dispersed the drone features extracted by the initial model for drone samples of the same sample, and the weaker the robustness of the initial model in extracting drone features.

[0098] In some embodiments, the intra-class loss can be expressed as the following formula: in, This is an intra-class loss. This represents the number of categories, which is the number of sample drones; that is, each sample drone can be classified as an independent category. It is a category The set of features of all sample drones in the dataset. It is a category The mean of the features of the sample drones, yes The first in Features of individual sample drones.

[0099] By minimizing intra-class loss This allows features belonging to the same drone to be more tightly clustered around the class center. To enhance the model's stability against intra-class variations, reduce the distribution of drone features within the same category, and improve feature robustness.

[0100] Inter-class loss refers to the loss caused by the differences in features between drone samples belonging to different sample drones. It can be understood that the greater the differences in features between drone samples, the more obvious the distinction between drone features extracted by the initial model for different sample drones; conversely, the smaller the differences in features between drone samples, the smaller the distinction between drone features extracted by the initial model for different sample drones.

[0101] In some embodiments, the inter-class loss can be expressed as the following formula: in, This is the inter-class loss. The total number of features of the sample drones For the first Characteristics of individual drone samples To and Characteristics of sample drones belonging to the same sample drone, It can be a similarity function, such as a cosine similarity function; This refers to the temperature parameter. To and Characteristics of sample drones belonging to different sample drones.

[0102] By minimizing the inter-class loss, the initial model can be encouraged to output more similar drone features for drones of the same type, and more distinct drone features for drones of different types. This effectively widens the gap between drone features belonging to different types of drones, while improving the robustness of drone features belonging to the same type of drone.

[0103] After obtaining the intra-class loss and between-class loss respectively, the loss value can be determined by combining these two values. The loss value is then applied to iterate the parameters of the initial model, and the resulting initial model after parameter iteration is used as the feature extraction model. Here, the loss value can be obtained by superimposing the intra-class loss and the between-class loss, or by weighted summing the intra-class loss and the between-class loss. This embodiment of the invention does not specifically limit this.

[0104] For example, the loss value can be calculated using the following formula: In the formula, The loss value used for parameter iteration. and The weighting coefficients are used to balance the intra-class loss and inter-class loss.

[0105] In the method provided in this embodiment of the invention, the feature extraction model is trained by combining the differences between the features of sample drones belonging to the same sample drone and the differences between the features of sample drones belonging to different sample drones, which can effectively enhance the drone feature extraction capability of the feature extraction model.

[0106] Based on any of the above embodiments, the sample electromagnetic signal includes the original electromagnetic signal and the electromagnetic signal after data enhancement of the original electromagnetic signal.

[0107] Specifically, when training a feature extraction model, for the same sample UAV, the sample electromagnetic signal includes the original electromagnetic signal and the electromagnetic signal after data augmentation of the original electromagnetic signal.

[0108] The original electromagnetic signal refers to the electromagnetic signal directly captured. Data augmentation can include operations such as adding noise and cropping. Augmenting the original electromagnetic signal can simulate the interference of the complex electromagnetic environment in a city, resulting in an augmented electromagnetic signal that better reflects the characteristics of electromagnetic signals in a real urban low-altitude complex spectrum environment.

[0109] Therefore, during the training of the feature extraction model, both the original electromagnetic signal and the data-augmented electromagnetic signal can be considered as sample electromagnetic signals for training. Furthermore, when iterating the parameters of the initial model, the features of sample drones belonging to the same sample drone can be the features from the original electromagnetic signal and the features from the data-augmented electromagnetic signal. Iterating the parameters of the initial model based on the differences between these two can make the sample drone features before and after data augmentation as similar as possible, thereby enhancing the feature extraction model's robustness and generalization ability in complex urban electromagnetic environments.

[0110] Figure 4 This is a schematic diagram of the training process of the feature extraction model provided by the present invention, as shown below. Figure 4 As shown, for sample drones 1 to n, each sample drone has a corresponding original electromagnetic signal, and an enhanced electromagnetic signal can be obtained by data enhancement of the original electromagnetic signal.

[0111] For the raw and enhanced electromagnetic signals of each sample UAV, its features can be extracted separately using an initial model. Here, the initial model includes feature extraction module 1, feature extraction module 2, and a module labeled "...". The output module is used to extract the initial model shared weights for the original and enhanced electromagnetic signals of different sample drones. After obtaining the sample drone features of the original and enhanced electromagnetic signals for each sample drone, intra-class loss can be calculated for the sample drone features of the original and enhanced electromagnetic signals of the same sample drone; inter-class loss can be calculated for the sample drone features of the original electromagnetic signals of different sample drones.

[0112] Based on this, the initial model can be iterated by combining intra-class loss and inter-class loss to obtain the feature extraction model.

[0113] Based on any of the above embodiments, the drone identification method further includes: Acquire the electromagnetic signals of the drone to be registered; Based on the feature extraction model, the drone features corresponding to the electromagnetic signals of the drone to be registered are extracted; Based on the drone characteristics corresponding to the electromagnetic signals of the drone to be registered, the registration characteristics of the drone to be registered are determined. The registration features and drone identifier of the drone to be registered are stored in the drone database.

[0114] Specifically, the drone to be registered refers to the newly added, incremental drones that are yet to be registered. The electromagnetic signals of the drone to be registered can be input into the feature extraction model, which then extracts the corresponding drone features from the electromagnetic signals of the drone to be registered, thereby obtaining the drone features of the drone to be registered output by the feature extraction model.

[0115] Based on this, registration features can be determined according to the drone characteristics of the drone to be registered. For example, if the drone to be registered has only one electromagnetic signal, its drone characteristics can be directly used as registration features; or if the drone to be registered has multiple electromagnetic signals, the drone characteristics corresponding to each of the multiple electromagnetic signals can be statistically analyzed, and the statistical results can be used as registration features. The statistical results here can be the average value calculated for multiple drone characteristics, or the center point calculated for multiple drone characteristics. This embodiment of the invention does not specifically limit this.

[0116] After obtaining the registration features, the registration features and the drone identifier of the drone to be registered can be stored in the drone database, thus adding a newly registered drone to the drone database.

[0117] The method provided in this embodiment of the invention realizes incremental learning of drones for a drone library. In this process, there is no need to retrain the feature extraction model or update its parameters; registration features can be obtained simply by applying the feature extraction model for forward inference. This improves the efficiency of incremental learning, reduces the computational resources and time required for incremental learning, and enables rapid drone registration.

[0118] Based on any of the above embodiments Figure 5 This is a schematic diagram of the incremental learning process for unmanned aerial vehicles provided by this invention. For example... Figure 5 As shown, assume the drone database stores the registration features of drones A, B, and C. Currently, there is a drone D to be registered, and m electromagnetic signals of drone D have been obtained, specifically electromagnetic signals 1 to m. Electromagnetic signals 1 to m can be input into a feature extraction model, which will then extract the drone features corresponding to each of the electromagnetic signals 1 to m. Therefore, the feature extraction model can output the drone features corresponding to each of the electromagnetic signals 1 to m, i.e., drone features 1 to drone features m.

[0119] Based on this, statistics can be performed on drone features 1 to m, and the statistical results can be used as the registration features of drone D, i.e., registration feature D. Then, registration feature D can be stored in the drone database, at which point the drone database stores the registration features of drones A, B, C, and D.

[0120] Based on any of the above embodiments Figure 6 This is the second flowchart illustrating the drone identification method provided by the present invention, as shown below. Figure 6 As shown, the electromagnetic signal of the drone to be identified can be input into the feature extraction model, thereby obtaining the drone feature corresponding to the electromagnetic signal output by the feature extraction model.

[0121] Subsequently, the similarity between the drone features and the registration features of each drone stored in the drone database is calculated to obtain the similarity between the drone features and each registration feature. This similarity can be represented by cosine similarity; that is, the larger the cosine similarity, the more similar the drone features and registration features are.

[0122] Therefore, the maximum value of the cosine similarity between the drone's features and each registered feature can be determined. This maximum value is then compared to a pre-set threshold. If the maximum value is greater than or equal to the threshold, the drone to be identified belongs to the registered feature corresponding to the maximum value; if the maximum value is less than the threshold, the drone to be identified is not a registered drone and is considered an unknown, unauthorized drone.

[0123] The drone identification device provided by the present invention is described below. The drone identification device described below and the drone identification method described above can be referred to in correspondence.

[0124] Figure 7 This is a schematic diagram of the structure of the drone identification device provided by the present invention, as shown below. Figure 7 As shown, the device includes: Acquisition unit 710 is used to acquire the electromagnetic signals of the drone to be identified; The inference unit 720 is used to extract UAV features corresponding to the electromagnetic signal based on a feature extraction model. The feature extraction model includes a feature extraction module, which is used to perform state space modeling by applying the temporal and spatial features of the electromagnetic signal, and to construct the UAV features by applying the long-range features obtained from the state space modeling. The identification unit 730 is used to match the drone features with the registration features of each drone in the drone database, and determine the identification result of the drone to be identified based on the matching result; the registration features are obtained based on the feature extraction model.

[0125] In the apparatus provided in this invention, due to the all-weather, covert, and long-range acquisition attributes of electromagnetic signal capture, UAV identification based on electromagnetic signals is more adaptable to the complex low-altitude spectrum environment in urban areas and does not require UAV cooperation. Furthermore, by applying the temporal and spatial features of electromagnetic signals to perform state-space modeling through a feature extraction module, efficient inference of long sequences of electromagnetic signals can be achieved while ensuring the full fusion of information in the time and frequency domains, thereby rapidly acquiring high-quality UAV features. Based on the UAV features obtained in this way, efficient and accurate UAV identification can be achieved in complex low-altitude spectrum environments in urban areas.

[0126] Based on any of the above embodiments, the inference unit is specifically used for: Based on the temporal extraction unit in the feature extraction module, the temporal features of the electromagnetic signal are extracted; Based on the frequency domain extraction unit in the feature extraction module, the spatial features of the frequency domain representation of the electromagnetic signal are extracted; Based on the state space unit in the feature extraction module, state space modeling is performed on the temporal features and the spatial features to obtain the long-range features; Based on the feature output unit in the feature extraction module, high-level features are constructed by combining the frequency domain representation and the long-range features. Based on the output module of the feature extraction model, the high-level features are used to construct the UAV features.

[0127] Based on any of the above embodiments, the inference unit is specifically used for: Based on the feature output unit in the feature extraction module, the residual features of the frequency domain representation are extracted, and the high-level features are constructed by combining the long-range features, the residual features, and the frequency domain representation.

[0128] Based on any of the above embodiments, the feature extraction model includes multiple cascaded feature extraction modules, where the high-level features output by the previous feature extraction module are the electromagnetic signals input by the next feature extraction module.

[0129] Based on any of the above embodiments, the inference unit is specifically used for: Based on the output module in the feature extraction model, the high-level features output by each of the multiple cascaded feature extraction modules are fused to obtain the UAV features.

[0130] Based on any of the above embodiments, the device further includes a training unit, used for: Acquire the electromagnetic signals of the sample drone; Based on the initial model, the sample UAV features of the sample electromagnetic signals are extracted; Based on the differences in features between sample drones belonging to the same sample drone, and the differences in features between sample drones belonging to different sample drones, the initial model is iterated to obtain the feature extraction model.

[0131] Based on any of the above embodiments, the sample electromagnetic signal includes the original electromagnetic signal and the electromagnetic signal after data enhancement of the original electromagnetic signal.

[0132] Based on any of the above embodiments, the device further includes a registration unit, used for: Acquire the electromagnetic signals of the drone to be registered; Based on the feature extraction model, the drone features corresponding to the electromagnetic signals of the drone to be registered are extracted; Based on the drone characteristics corresponding to the electromagnetic signals of the drone to be registered, the registration characteristics of the drone to be registered are determined. The registration features and drone identifier of the drone to be registered are stored in the drone database.

[0133] Figure 8 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 8 As shown, the electronic device may include: a processor 810, a communications interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communications interface 820, and the memory 830 communicate with each other via the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute a UAV identification method, which includes: Acquire the electromagnetic signals of the drone to be identified; Based on the feature extraction model, the UAV features corresponding to the electromagnetic signal are extracted; the feature extraction model includes a feature extraction module, which is used to perform state space modeling by applying the temporal and spatial features of the electromagnetic signal, and to construct the UAV features by applying the long-range features obtained by the state space modeling. The drone features are matched with the registration features of each drone in the drone database, and the identification result of the drone to be identified is determined based on the matching result; the registration features are obtained based on the feature extraction model.

[0134] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to related technologies, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0135] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program that can be stored on a non-transitory computer-readable storage medium, wherein when the computer program is executed by a processor, the computer is able to execute the unmanned aerial vehicle (UAV) identification method provided by the above methods, the method comprising: Acquire the electromagnetic signals of the drone to be identified; Based on the feature extraction model, the UAV features corresponding to the electromagnetic signal are extracted; the feature extraction model includes a feature extraction module, which is used to perform state space modeling by applying the temporal and spatial features of the electromagnetic signal, and to construct the UAV features by applying the long-range features obtained by the state space modeling. The drone features are matched with the registration features of each drone in the drone database, and the identification result of the drone to be identified is determined based on the matching result; the registration features are obtained based on the feature extraction model.

[0136] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the unmanned aerial vehicle (UAV) identification method provided by the above methods, the method comprising: Acquire the electromagnetic signals of the drone to be identified; Based on the feature extraction model, the UAV features corresponding to the electromagnetic signal are extracted; the feature extraction model includes a feature extraction module, which is used to perform state space modeling by applying the temporal and spatial features of the electromagnetic signal, and to construct the UAV features by applying the long-range features obtained by the state space modeling. The drone features are matched with the registration features of each drone in the drone database, and the identification result of the drone to be identified is determined based on the matching result; the registration features are obtained based on the feature extraction model.

[0137] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0138] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of software products. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0139] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying unmanned aerial vehicles (UAVs), characterized in that, include: Acquire the electromagnetic signals of the drone to be identified; Based on the feature extraction model, the UAV features corresponding to the electromagnetic signals are extracted; The feature extraction model includes a feature extraction module, which is used to perform state space modeling by applying the temporal and spatial features of the electromagnetic signal, and to construct the UAV features by applying the long-range features obtained from the state space modeling. The drone features are matched with the registration features of each drone in the drone database, and the identification result of the drone to be identified is determined based on the matching result; the registration features are obtained based on the feature extraction model.

2. The UAV identification method according to claim 1, characterized in that, The step of extracting UAV features corresponding to the electromagnetic signals based on the feature extraction model includes: Based on the temporal extraction unit in the feature extraction module, the temporal features of the electromagnetic signal are extracted; Based on the frequency domain extraction unit in the feature extraction module, the spatial features of the frequency domain representation of the electromagnetic signal are extracted; Based on the state space unit in the feature extraction module, state space modeling is performed on the temporal features and the spatial features to obtain the long-range features; Based on the feature output unit in the feature extraction module, high-level features are constructed by combining the frequency domain representation and the long-range features. Based on the output module of the feature extraction model, the high-level features are used to construct the UAV features.

3. The UAV identification method according to claim 2, characterized in that, The construction of high-level features based on the feature output unit in the feature extraction module, combined with the frequency domain representation and the long-range features, includes: Based on the feature output unit in the feature extraction module, the residual features of the frequency domain representation are extracted, and the high-level features are constructed by combining the long-range features, the residual features, and the frequency domain representation.

4. The UAV identification method according to claim 2, characterized in that, The feature extraction model includes multiple cascaded feature extraction modules, where the high-level features output by the previous feature extraction module are the electromagnetic signals input to the next feature extraction module.

5. The UAV identification method according to claim 4, characterized in that, The step of constructing the UAV features by applying the high-level features based on the output module of the feature extraction model includes: Based on the output module in the feature extraction model, the high-level features output by each of the multiple cascaded feature extraction modules are fused to obtain the UAV features.

6. The UAV identification method according to any one of claims 1 to 5, characterized in that, The training steps of the feature extraction model include: Acquire the electromagnetic signals of the sample drone; Based on the initial model, the sample UAV features of the sample electromagnetic signals are extracted; Based on the differences in features between sample drones belonging to the same sample drone, and the differences in features between sample drones belonging to different sample drones, the initial model is iterated to obtain the feature extraction model.

7. The UAV identification method according to claim 6, characterized in that, The sample electromagnetic signal includes the original electromagnetic signal and the electromagnetic signal after data enhancement of the original electromagnetic signal.

8. The UAV identification method according to any one of claims 1 to 5, characterized in that, Also includes: Acquire the electromagnetic signals of the drone to be registered; Based on the feature extraction model, the drone features corresponding to the electromagnetic signals of the drone to be registered are extracted; Based on the drone characteristics corresponding to the electromagnetic signals of the drone to be registered, the registration characteristics of the drone to be registered are determined. The registration features and drone identifier of the drone to be registered are stored in the drone database.

9. A drone identification device, characterized in that, include: Acquisition unit, used to acquire electromagnetic signals of the drone to be identified; The inference unit is used to extract the UAV features corresponding to the electromagnetic signals based on the feature extraction model. The feature extraction model includes a feature extraction module, which is used to perform state space modeling by applying the temporal and spatial features of the electromagnetic signal, and to construct the UAV features by applying the long-range features obtained from the state space modeling. The identification unit is used to match the drone features with the registration features of each drone in the drone database, and determine the identification result of the drone to be identified based on the matching result; the registration features are obtained based on the feature extraction model.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the drone identification method as described in any one of claims 1 to 8.

11. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the drone identification method as described in any one of claims 1 to 8.

Images