A method for identifying unknown UAV swarms based on signal semantic space construction
By constructing a signal semantic space, employing short-time Fourier transform and multi-level perception mechanism for feature extraction, and combining an outlier detection strategy, the problem of refining the identification of unknown UAVs was solved, achieving efficient identification and cluster segmentation of unknown UAV signals.
Patent Information
- Application Number
- CN202411113655.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-14
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2044-08-14
AI Technical Summary
Existing drone identification algorithms are unable to effectively identify unknown types of drones, especially given the increasing variety of drones. They are unable to achieve precise identification of unknown drone signals and effective response to swarm intrusions.
By employing a signal semantic space construction method, short-time Fourier transform, multi-level perception mechanism feature extraction, joint optimized semantic space construction, and outlier detection strategy, sparsely expressed signal semantics are obtained, enabling precise identification of unknown UAV signals.
It achieves refined identification of unknown UAV signals, breaks the closed set assumption, enhances the semantic expression of UAV signals, and realizes refined identification of unknown UAV swarm signals in high-dimensional space.
Smart Images

Figure CN119025888B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of signal processing and pattern recognition, and in particular to a method for identifying unknown drone swarms based on signal semantic space construction. Background Technology
[0002] In recent years, with the increasing maturity of drone technology, drones have been widely used in fields such as geographic surveying, power line inspection, and medical rescue. Although the application of drones is transforming people's production and lifestyles, the numerous serious incidents caused by unauthorized and excessive drone flights have also raised an urgent need for stronger drone management.
[0003] Radio frequency (RF)-based drone detection is one of the most effective methods in urban scenarios, offering advantages such as insensitivity to non-line-of-sight environments and long detection range. By deploying RF signal receivers to collect RF signals from the surrounding environment and analyzing drone communication activities, intruding drones can be effectively detected and identified.
[0004] Some radio frequency (RF) signal-based drone identification algorithms adhere to the closed-set assumption, requiring the test signal to be of the same type as the training signal. This leads to the inability to identify unknown drone types. Given the increasing variety of drones, collecting all possible drone signals to expand the training closed set is impractical and costly. Therefore, there is an urgent need to develop RF identification algorithms for open sets. Furthermore, to effectively address the challenge of drone swarm intrusion, unknown drones not only need to be detected but also further refined in their identification, i.e., inferring their number of types and determining the unknown type to which each signal belongs. To our knowledge, refined identification of unknown drone signals is a novel challenge that has not yet been considered in existing work. Developing algorithms for this task can effectively expand drone RF datasets and drive existing supervised learning models for drone RF signals towards a new incremental learning model. Summary of the Invention
[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method for identifying unknown drone swarms based on signal semantic space construction. Through robust characterization of drone radio frequency signals, feature extraction using multi-level sensing mechanisms, and joint optimization of signal semantic space construction, sparsely expressed signal semantics are obtained. Furthermore, an outlier detection strategy is employed to determine the boundary thresholds of the semantic space, thereby effectively discovering unknown drone signals. For refined identification of a large number of unknown samples (swarm signals), basic features are introduced to expand the original signal semantic space to a higher-dimensional space, making the semantics of unknown drones separable in the higher-dimensional space. Cluster analysis ultimately enables refined identification of drone swarm signals.
[0006] The objective of this invention is achieved through the following technical solution: a method for identifying unknown drone swarms based on signal semantic space construction, the method comprising:
[0007] S1. Perform a short-time Fourier transform on the raw signal acquired by the radio frequency receiver to obtain the signal characterization;
[0008] S2. A semantic feature extractor composed of multi-level perceptron convolution operators is used to learn features of the signal representation, and a nonlinear perceptron is used to construct the semantic space of the output feature map.
[0009] S3. Semantic space optimization is performed by combining center loss, clustering loss, and cross-entropy function.
[0010] S4. Based on the Gaussian distribution property of the semantics of the training closed set signal, an outlier detection strategy is adopted to determine the boundary threshold of the semantic space and segment the semantic space, thereby determining the explicit expression for discovering unknown signals.
[0011] S5. Introduce the features learned but not compressed in S2 to expand the semantic space of the original signal to a higher-dimensional space. In this higher-dimensional space, use clustering to evaluate the number of unknown signal types and determine their specific unknown categories.
[0012] Furthermore, S1 specifically includes:
[0013] Short-time Fourier analysis is used to transform the time-domain signal to the time-frequency domain. A fast Fourier transform is performed independently on each segment, enabling the signal to be jointly represented in the two-dimensional domain of time and frequency. This process is expressed as:
[0014]
[0015] Where R represents the sliding window of the short-time Fourier transform, and g(t) is the Hamming window function used to prevent spectral leakage. This represents the analysis frequency points, and W represents the analysis window length. This represents the analysis time point, and N represents the number of analysis windows.
[0016] Furthermore, the semantic feature extractor includes three parallel, multi-level perceptron convolution operators and a corresponding nonlinear perceptron. The convolution operators perform feature learning on the signal representation, including:
[0017]
[0018] In this context, MaxPool() represents the max pooling function, ReLU() represents the nonlinear activation function, and * represents the convolution operation. This represents the output of the i-th feature layer using a dilated convolution operator with dilation coefficient c, specifically Z0 = X, Imax Indicates the number of feature layers. Let represent the learnable parameters of the dilated convolution operator with dilation coefficient c for the i-th feature layer.
[0019] Furthermore, the semantic space construction includes: reducing the feature map to a one-dimensional vector using global max pooling, and compressing and extracting effective features using a nonlinear perceptron, specifically including:
[0020]
[0021] Where GlobalAvgPool() represents the max pooling operation. That is, the mean of all elements in a feature map of size d1×d2 is calculated as the output.
[0022] The nonlinear perceptron compresses and extracts effective features, represented as:
[0023]
[0024] in and b c These are the learnable parameters of the nonlinear perceptron corresponding to the dilated convolution operator with an expansion coefficient of c; ultimately, the constructed signal semantics are the concatenation of the compressed feature results of all nonlinear perceptrons.
[0025] Furthermore, the semantic space optimization specifically includes: using a gradient descent algorithm to jointly optimize three loss functions to update and optimize the learnable coefficients of the multi-level receptive convolution operator and the nonlinear perceptron: θ←θ-▽ θ (L center (θ)+L cluster (θ)+L ce (θ))
[0026] The central loss is expressed as follows:
[0027]
[0028] in, Represents a known class y i The corresponding semantic center of the signal, ||·||2 denotes the calculation of the L2 norm of the vector, where The acquisition is Represented as
[0029]
[0030] Where δ(·) represents the conditional decision function, δ(·) = 1 when the condition is true, otherwise δ(·) = 0;
[0031] The clustering loss function is used to optimize the spatial distance between different semantic classes, denoted as:
[0032]
[0033] in, It is a segmentation threshold given in the signal semantic space.
[0034] The cross-entropy function is expressed as:
[0035]
[0036] Wherein, P(X) i ;θ) represents the input X i Predicted as y i The probability of the category.
[0037] Furthermore, the outlier detection strategy is used to segment the signal semantic space, including:
[0038]
[0039] Where, Θ k Let dist(z,μ) represent the boundary threshold of the k-th class of signal, where d represents the dimension of the signal semantic space. k ) represents
[0040]
[0041] in, Let represent the inverse matrix of the covariance matrix of the k-th semantic vector.
[0042] Furthermore, the calculation of the boundary threshold includes: firstly, the distance set from the semantics of samples of category k in the training closed set to their corresponding semantic centers is represented as:
[0043]
[0044] Where N k This represents the number of samples of class k in the training closed set, based on the semantic center μ. k Reorder the distance sets from largest to smallest to obtain a descending distance set, represented as:
[0045]
[0046] Based on the 3-sigma criterion, traverse... The first non-outlier value is used as the boundary threshold for the k-th sub-semantic space, denoted as:
[0047]
[0048] Where, σk express The standard deviation.
[0049] Furthermore, the number of types of unknown signals in the inference specifically includes: considering the cases where the unknown signals are the same type and multiple types, wherein when the unknown signals are the same type, the determination criterion is: for the set of unknown signal samples... Calculate its boundary threshold If the inequality If true, then u = 1;
[0050] If there are multiple unknown signals, the number of categories obtained by segmenting the unknown semantic set is determined by an evaluation function ψ(·). Specifically, the evaluation function ψ(·) is:
[0051] And there are:
[0052]
[0053]
[0054] Then the inference of the optimal unknown class u * Represented as
[0055]
[0056] The k-means clustering algorithm is used to obtain the category to which the unknown signal belongs.
[0057] On the other hand, the specification also provides an unknown drone swarm identification device based on signal semantic space construction, including a memory and one or more processors. The memory stores executable code, and when the processor executes the executable code, it implements the unknown drone swarm identification method based on signal semantic space construction.
[0058] On the other hand, the specification also provides a computer-readable storage medium on which a program is stored, which, when executed by a processor, implements the aforementioned method for identifying unknown drone swarms based on signal semantic space construction.
[0059] The beneficial effects of this invention are:
[0060] (1) This invention takes into account the open set identification problem of UAV signals, breaks the closed set assumption of most existing work, and enables the radio frequency detection system to cope with the identification challenge of unknown UAV signals.
[0061] (2) This invention designs a dedicated feature extractor for UAV signal representation instead of using a general universal feature extractor, which enhances the expression of UAV signal semantics.
[0062] (3) The present invention designs a high-dimensional extension method of semantic space so that the signals of unknown UAV clusters can be accurately identified in the high-dimensional semantic space without introducing additional computational complexity. Attached Figure Description
[0063] Figure 1 This is an overall flowchart of the present invention.
[0064] Figure 2 This invention is a dedicated feature extractor for UAV signal characterization.
[0065] Figure 3 This is a schematic diagram of the semantic space constructed in this invention.
[0066] Figure 4 This invention presents the semantic space segmentation performance based on outlier detection.
[0067] Figure 5 This demonstrates the effectiveness of the high-dimensional expansion method for the signal semantic space proposed in this invention.
[0068] Figure 6 This is a schematic diagram of an unknown drone swarm identification device based on signal semantic space provided in an embodiment of the present invention. Detailed Implementation
[0069] The following is for reference only. Figure 1 The technical solution of the present invention will be further described in detail below.
[0070] To address the robust characterization of the UAV signal in step one, a short-time Fourier transform is used to convert the acquired time-domain signal to the time-frequency domain, thereby decoupling the target UAV signal. Specifically, the ambient radio frequency signal acquired by the radio frequency receiver can be expressed as:
[0071] x(t)=x video (t)+x control (t)+x interference (t)+n(t) (1)
[0072] Where x video (t) represents the video stream transmitted by the UAV's onboard camera to the ground controller, x control (t) indicates that the UAV's flight control signals are used to carry flight logs and flight commands, x interference(t) represents interference signals in urban scenarios, such as those from Bluetooth and WiFi devices, and n(t) is the inherent noise generated by the RF receiver itself, such as Gaussian white noise. Since the power of the interference signal may be greater than that of the drone signal, causing the characteristics of the drone signal to be obscured, short-time Fourier analysis is used to transform the time-domain signal to the time-frequency domain. Short-time Fourier analysis divides the input time-domain signal into several segments of equal length and performs a fast Fourier transform independently on each segment, thus enabling the signal to be jointly represented in the two-dimensional domain of time and frequency. This process is expressed as:
[0073]
[0074] Where R represents the sliding window of the short-time Fourier transform, and g(t) is the Hamming window function used to prevent spectral leakage. This represents the analysis frequency points, and W represents the analysis window length. This represents the analysis time point, and N represents the number of analysis windows.
[0075] For signal representation learning in step two, a multi-level receptive convolution operator is employed. Due to the significant differences in modulation methods between UAV image transmission signals and flight control signals, their representations in X differ significantly. Specifically, because image transmission signals employ orthogonal frequency division multiplexing (OFDM) with frequency hopping spread spectrum, they exhibit short-term, time-invariant frequency characteristics, while flight control signals have continuously time-varying frequencies. Furthermore, the bandwidth of image transmission signals is typically greater than 10MHz, while the bandwidth of flight control signals is typically 1–5MHz. Since the characteristics of flight control signals are weaker than those of image transmission signals, a single-level feature extractor cannot effectively learn both types of signals simultaneously.
[0076] Appendix Figure 2 The proposed semantic feature extractor for UAVs incorporates three multi-level receptive convolution operators, DCL-c, where c = 1, 3, and 5 represent the dilation coefficients of the dilated convolution. Furthermore, a global max-pooling operation is used to flatten the output of the dilated convolution operator into a one-dimensional vector.
[0077] The signal representation is learned using a multi-level receptive convolution operator as follows:
[0078]
[0079] In this context, MaxPool() represents the max pooling function, ReLU() represents the nonlinear activation function, and * represents the convolution operation. This represents the output of the i-th feature layer using the c-dilated convolution operator, specifically Z0 = X. max Indicates the number of feature layers. Let c represent the learnable parameters of the c-dilated convolution operator of the i-th feature layer.
[0080] For constructing the signal semantic space, we first use global max pooling to reduce the feature map to a one-dimensional vector, as follows:
[0081]
[0082] Where GlobalAvgPool() represents the max pooling operation. That is, the mean of all elements in a feature map of size d1×d2 is calculated as the output. Further, the nonlinear perceptron compresses and extracts the effective features, represented as:
[0083]
[0084] in and b c These are the learnable parameters of the nonlinear perceptron corresponding to the convolution operator with dilation coefficient c. Finally, the constructed signal semantic representation is:
[0085]
[0086] in This represents concatenating two one-dimensional vectors, i.e. However, the semantic space constructed from this signal semantics is not sparsity, therefore it is necessary to introduce a semantic space optimization strategy. Let E(·) represent the semantic construction process for X, and θ represent all the learnable parameters mentioned above. Given a batch size of M training data... in and Let represent the set of all known classes. The centrality loss is expressed as follows:
[0087]
[0088] in, Represents a known class The corresponding signal semantic center, ||·||2 represents the calculation of the vector's L2 norm. Given a learning rate α, The acquisition is in Represented as
[0089]
[0090] Here, δ(·) represents the conditional decision function, where δ(·) = 1 when the condition is true, and δ(·) = 0 otherwise.
[0091] Furthermore, a clustering loss function is introduced to optimize the spatial distance between different classes, expressed as:
[0092]
[0093] in, This is a segmentation threshold given in the signal semantic space. For optimized signal semantics, it needs to possess the ability for directly supervised classification; therefore, the cross-entropy function is introduced, expressed as:
[0094]
[0095] Wherein, P(X) i ;θ) represents the input X i Predicted as The probability of each category. Ultimately, to obtain a sparsified signal semantic space, a gradient descent algorithm can be used to jointly optimize three loss functions, expressed as follows:
[0096] θ←θ-▽ θ (L center (θ)+L cluster (θ)+L ce (θ)) (11)
[0097] Regarding the detection of the unknown signal in step four, since L was introduced in step three... center Therefore, the signal semantic space is approximately Gaussian distributed. Based on this property, an outlier detection algorithm is proposed to segment the signal semantic space as follows:
[0098]
[0099] Where, Θ k Let be the boundary threshold of the k-th class of signal, and d represent the dimension of the signal's semantic space. dist(z,μ) k ) represents
[0100]
[0101] in, Let be the inverse matrix of the covariance matrix of the k-th semantic vector. Based on this, for any semantic vector z in the test set... t If a known category exists Make If true, then z t It should be considered a known category. Conversely, z t It is considered an unknown category. Therefore, the problem of detecting signals of unknown class is essentially to find a set of boundary thresholds to divide the semantic space, such that each segmented sub-semantic space contains as few as possible signals of known class. Since L center The introduction of this feature results in the closed-set signals in the semantic space exhibiting an approximately Gaussian distribution. An outlier detection strategy is then employed to determine the boundary threshold. Specifically, the implementation is as follows: First, the distance set from the semantics of samples of category k in the closed set to their corresponding semantic centers is represented as:
[0102]
[0103] Where N k This represents the number of samples of class k in the training closed set. Based on the semantic center μ... k Reorder the distance sets from largest to smallest to obtain a descending distance set, represented as:
[0104]
[0105] Based on the 3-sigma criterion, traverse... The first non-outlier value is used as the boundary threshold for the k-th sub-semantic space, denoted as:
[0106]
[0107] Where, σ k express The standard deviation of . The cutoff threshold is now determined. For a test sample, if it belongs to a known class, then . And its known categories are
[0108]
[0109] Regarding step five, due to the original space The signal obtained from training on a closed set is insufficient for representing the test signal in an open set. Therefore, the uncompressed features learned in step two are introduced to expand the semantic space of the original signal to a higher-dimensional space, represented as follows:
[0110]
[0111] The set of detected unknown signal samples is defined as follows: Where NU represents the number of detected unknown samples. The problem of finely dividing unknown signals is transformed into how to... The problem of dividing the unknown semantics in C into u subgroups, i.e. u ={C1,C2,...,C u At this time, C u Each subgroup is categorized as the unknown class with the highest percentage within that group. For u, this can be further divided into two cases, A and B.
[0112] A: Consider the case where u=1. When all unknown class semantics are grouped together, all unknown samples should be classified as the same unknown class. The criterion is: for Calculate its boundary threshold according to equation (16). If the inequality If true, then u = 1.
[0113] B: Consider the case where u ≠ 1. Given any u = 2, 3, ..., NU, the optimal segmentation of the unknown semantic set can be expressed as:
[0114]
[0115] in It is the Cth i The center of each subgroup. Equation (19) above can be solved by the conventional k-means clustering algorithm, the premise of which is to determine the value of u. For this purpose, an evaluation function ψ(·) is introduced, which has the following properties: when the number of subgroups u is close to the true number of unknown subgroups, ψ(C u The function of u reaches an extreme point. Therefore, the chosen ψ(·) is denoted as...
[0116]
[0117] And there are
[0118]
[0119]
[0120] Therefore, the inference of the optimal unknown class u * Represented as
[0121]
[0122] Given u * Under the given conditions, the k-means clustering algorithm is used again to solve the optimization problem equation (19), and the category to which the unknown signal belongs can be obtained.
[0123] The effects of the present invention will be further described below with reference to specific experiments.
[0124] The experiment used the publicly available dataset DroneRFa, which contains 24 drone signal types. Eighteen of these were selected as a closed set (known) for both training and testing, with a training-to-test data ratio of 7:3. The remaining six were selected as an open set (unknown) for testing only. For signal representation, the parameters were set as follows: W = 2048, N = 1024. For signal representation learning, I was set... max =3. For signal semantic space construction, the parameters are set as follows: d = 128.
[0125] Appendix Figure 2The proposed semantic feature extractor for UAVs incorporates three multi-level receptive convolution operators, DCL-c, where c = 1, 3, and 5 represent the dilation coefficients of the dilated convolution. Furthermore, a global max-pooling operation is used to flatten the output of the dilated convolution operator into a one-dimensional vector.
[0126] Appendix Figure 3 The visualization represents the constructed signal semantic space, where red indicates all known signal classes, which exhibit good sparse representation due to optimization by the three loss functions. The remaining colors represent open-set unknown cluster semantics. Due to the designed high-dimensional signal semantic space expansion method, these unknown cluster semantics can present a sparse distribution, thus enabling effective and precise classification.
[0127] In the appendix Figure 4 The performance of the proposed semantic space segmentation based on outlier detection strategy is validated in this study, with comparisons made to two common outlier detectors: Local Outlier Factor (LOF) and Isolation Forest (IF). Figure 2 It can be seen that, in terms of both known acceptance rate and unknown rejection rate, the proposed method is significantly better than the comparison method.
[0128] Appendix Figure 5 The necessity of the proposed high-dimensional semantic space extension method has been verified. Figure 4 It can be seen that, through semantic space expansion, the detection rate of unknown cluster signals increased from 0.9171 to 0.9517, and the recognition rate increased significantly from 0.7180 to 0.8763.
[0129] Corresponding to the aforementioned embodiment of an unknown drone swarm identification method based on signal semantic space construction, the present invention also provides an embodiment of an unknown drone swarm identification device based on signal semantic space construction.
[0130] See Figure 6 The present invention provides an unknown drone swarm identification device based on signal semantic space construction, comprising a memory and one or more processors. The memory stores executable code, and when the processor executes the executable code, it is used to implement an unknown drone swarm identification method based on signal semantic space construction in the above embodiment.
[0131] The embodiment of the unknown UAV swarm identification device based on signal semantic space construction provided by this invention can be applied to any device with data processing capabilities, such as a computer. The device embodiment can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 5 The diagram shown illustrates a hardware structure of any data processing-capable device, including the unknown UAV swarm identification device based on signal semantic space construction provided by this invention. (Except for...) Figure 6 In addition to the processor, memory, network interface, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.
[0132] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.
[0133] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and 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 the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.
[0134] This invention also provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements a method for identifying an unknown UAV swarm based on a signal semantic space constructed in the above embodiments.
[0135] The computer-readable storage medium can be an internal storage unit of any data processing device as described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device of any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units and external storage devices of any data processing device. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.
[0136] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned method for identifying unknown drone swarms based on signal semantic space construction.
[0137] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this application are indicated by the claims.
[0138] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this application. This application is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.
Claims
1. A method for identifying unknown unmanned aerial vehicle (UAV) swarms based on signal semantic space construction, characterized in that, The method includes: S1. Perform a short-time Fourier transform on the raw signal acquired by the radio frequency receiver to obtain the signal characterization; S2. A semantic feature extractor composed of multi-level perceptron convolution operators is used to learn features of the signal representation, and a nonlinear perceptron is used to construct the semantic space of the output feature map. S3. Semantic space optimization is performed by combining center loss, clustering loss, and cross-entropy function. S4. Based on the Gaussian distribution property of the semantics of the training closed set signal, an outlier detection strategy is adopted to determine the boundary threshold of the semantic space and segment the semantic space, thereby determining the explicit expression for discovering unknown signals. S5. Introduce the learned but uncompressed features from S2 to expand the original signal semantic space to a higher-dimensional space. In this higher-dimensional space, clustering is used to evaluate the number of unknown signal categories and determine their specific unknown categories. The inference of the number of unknown signal categories specifically includes considering the cases where the unknown signals are of the same type and multiple types. When the unknown signals are of the same type, the determination criterion is: for the unknown class signal sample set... Calculate its boundary threshold If the inequality If it is established, then ; If there are multiple unknown signals, the number of categories obtained by segmenting the unknown semantic set is determined by an evaluation function. Specifically: And there are: Then the inference of the optimal unknown class Represented as ; The number of unknown samples detected. The k-means clustering algorithm is used to obtain the category to which the unknown signal belongs.
2. The method for identifying unknown UAV swarms based on signal semantic space construction according to claim 1, characterized in that, S1 specifically includes: Short-time Fourier analysis is used to transform the time-domain signal to the time-frequency domain. A fast Fourier transform is performed independently on each segment, enabling the signal to be jointly represented in the two-dimensional domain of time and frequency. This process is expressed as: in The sliding window represents the short-time Fourier transform. The Hamming window function is used to prevent spectral leakage. Indicates the frequency points to be analyzed and Indicates the length of the analysis window. Indicates the analysis time point and Indicates the number of analysis windows.
3. The method for identifying unknown UAV swarms based on signal semantic space construction according to claim 1, characterized in that, The semantic feature extractor includes three parallel, multi-level perceptron convolution operators and a corresponding nonlinear perceptron. The convolution operator performs feature learning on the signal representation, including: in, Represents the max pooling function. Represents a nonlinear activation function. This represents the convolution operation. Indicates the first Each feature layer uses an expansion coefficient of... The output of the dilated convolution operator, especially , Indicates the number of feature layers. Indicates the first The expansion coefficient of each feature layer is The learnable parameters of the dilated convolution operator.
4. The method for identifying unknown UAV swarms based on signal semantic space construction according to claim 3, characterized in that, The semantic space construction includes: reducing the feature map to a one-dimensional vector using global max pooling, and compressing and extracting effective features using a nonlinear perceptron, specifically including: in Represents max pooling operation That is, the calculated dimensions are The mean of all elements in the feature map is used as the output. The nonlinear perceptron compresses and extracts effective features, represented as: in and The expansion coefficient is The learnable parameters of the nonlinear perceptron corresponding to the dilated convolution operator; finally, the constructed signal semantics is the concatenation of the compressed feature results of all nonlinear perceptrons.
5. The method for identifying unknown UAV swarms based on signal semantic space construction according to claim 1, characterized in that, The semantic space optimization specifically includes: using a gradient descent algorithm to jointly optimize three loss functions to update and optimize the learnable coefficients of the multi-level receptive convolution operator and the nonlinear perceptron. The central loss is expressed as follows: in, The representative batch size is The training data, of which and Represents the set of all known classes. Represents a known class The corresponding signal semantic center, This represents the calculation of the L2 norm of a vector, given a learning rate. Under the conditions, among which The acquisition is , Represented as in, This represents a conditional decision function, which evaluates a function to the condition when it is true. ,otherwise ; with function Expressing The semantic construction process, in order to This represents all learnable parameters; The clustering loss function is used to optimize the spatial distance between different semantic classes, denoted as: in, It is a segmentation threshold given in the signal semantic space. The cross-entropy function is expressed as: in, Indicates input Predicted as The probability of the category.
6. The method for identifying unknown UAV swarms based on signal semantic space construction according to claim 1, characterized in that, The outlier detection strategy is used to segment the signal semantic space, including: in, Represented as the first Boundary threshold of signal class The dimension representing the semantic space of the signal. Represented as in, Indicates the first The inverse of the covariance matrix of semantic vectors.
7. The method for identifying unknown UAV swarms based on signal semantic space construction according to claim 6, characterized in that, The calculation of the boundary threshold includes: first, training the closed set of classes as... The set of distances from a sample's semantics to its corresponding semantic center is represented as: in Indicates the categories in the training closed set as The number of samples, based on the semantic center Reorder the distance sets from largest to smallest to obtain a descending distance set, represented as: Based on the 3-sigma criterion, traverse... Find the value of the first non-outlier as the first The boundary threshold of the sub-semantic space is expressed as in, express standard deviation This represents a conditional decision function.
8. An unknown unmanned aerial vehicle (UAV) swarm identification device based on signal semantic space construction, comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that, When the processor executes the executable code, it implements a method for identifying unknown drone swarms based on signal semantic space as described in any one of claims 1-7.
9. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the method for identifying unknown drone swarms based on signal semantic space as described in any one of claims 1-7.
