Non-cooperative unmanned aerial vehicle open set perception method based on multi-scale representation learning

Abstract

The application aims to provide a non-cooperative unmanned aerial vehicle open set perception method based on multi-scale feature learning, belonging to the technical field of unmanned aerial vehicles; the method comprises four functional modules of adaptive multi-scale time-frequency feature extraction, known class hypersphere representation, unknown class boundary perception generation and open set decision; specifically, the multi-scale time-frequency feature extraction module alleviates the problem of weak fingerprint extraction difficulty; the known hypersphere representation module shapes the compact known class geometry; the boundary perception virtual point generation module explicitly regularizes the high-risk identity boundary under the condition of no non-cooperative unmanned aerial vehicle sample; the open set decision module realizes the calibrated non-cooperative unmanned aerial vehicle rejection in combination with the energy residual and the cosine interval; the application is experimentally verified under multiple openness based on a public data set, and the results show that the proposed RFGuard architecture is superior to the baseline algorithm in both authorized unmanned aerial vehicle identification and non-cooperative unmanned aerial vehicle rejection, and can provide support for the development of anti-unmanned aerial vehicle technology for point defense and low-altitude joint defense.

Description

Technical Field

[0001] This invention belongs to the field of unmanned aerial vehicle (UAV) technology, specifically relating to a non-cooperative UAV open set perception method based on multi-scale representation learning. Background Technology

[0002] In the rapidly developing low-altitude network system, unmanned aerial vehicles (UAVs) have been widely deployed in various tasks, including agricultural plant protection and emergency rescue. However, the widespread use of UAVs has also led to frequent unauthorized incursions into restricted airspace. With their advantages of small size, flexible takeoff and landing, excellent stealth capabilities, and low deployment costs, UAVs can overcome traditional spatial limitations and quickly complete target missions in complex environments, making UAV perception and detection particularly crucial for low-altitude security governance. Traditional UAV identification methods, such as radar detection, visual recognition, and acoustic perception, have inherent limitations, including short detection range, susceptibility to environmental interference, and the inability to achieve physical-layer UAV identification.

[0003] Radio Frequency Fingerprint Identification (RFFI) technology extracts device-specific features left in the transmitted signal by hardware defects in the transmitter of a radiation source. It provides a physical layer authentication scheme for individual UAVs, independent of upper-layer protocols, and is a crucial technique for distinguishing friend from foe and identifying unauthorized flights in anti-UAV systems. The underlying mechanism of RFFI stems from the inherent non-ideal characteristics of the device's analog front-end: hardware defects such as transmitter nonlinear distortion, local oscillator phase noise, carrier frequency offset, I / Q branch mismatch, and filter amplitude and phase distortion couple unique device features into the radiation source's radio frequency signal, forming a hardware fingerprint that cannot be generated at the software level. Compared to logical layer identifiers such as MAC addresses, ID broadcasts, and encryption tokens, RFFI technology, starting from the signal itself, possesses stronger anti-spoofing capabilities and reliability. Early RFFI research generally employed expert feature engineering: manually designed physical quantities were extracted from transient signals, steady-state modulation, constellation diagrams, and cyclostationary features, and then recognition was completed using learners such as support vector machines, random forests, and Bayesian classifiers. This model exhibits high accuracy under low noise, fixed receiver links, and fixed channels, and has been validated in scenarios such as satellite telemetry and control, radar radiation sources, and wireless / IoT devices. However, expert features heavily rely on prior design, and their reliability and accuracy rapidly decline under complex low-altitude electromagnetic environments and time-varying channel conditions.

[0004] In recent years, deep learning (DL), with its ability to extract nonlinear radio frequency fingerprint features, has gradually replaced traditional methods based on expert knowledge, becoming the mainstream approach for the detection, sorting, and individual identification of radio frequency signals from radar radiation sources, satellite ground stations, and UAVs. Specifically, deep learning models device identification as a multi-classification task, directly using I / Q sequences, time-frequency images, or time-frequency spectra such as STFT, FFS-SPWVD, and Wigner-Ville distributions as input. It achieves end-to-end learning of radio frequency fingerprint representations through convolutional neural networks, Transformers, and ensemble networks. In UAV individual identification, commonly used methods include: combining CNNs and ensemble learning to improve the distinguishing performance of individuals of the same model; employing a high-precision individual identification framework based on time-frequency maps and DIR-YOLOv11; using domain adaptation models based on Transformer encoders and domain adversarial networks to improve cross-domain generalization; and exploring UAV identification problems under conditions of low signal-to-noise ratio and multi-dimensional signal features. However, the aforementioned works are limited by the closed-set assumption, all assuming that all identities to be identified have been enumerated during the training phase, and the model's output space consists only of known classes, failing to provide a rejection mechanism for individuals from radiation sources outside the training distribution. However, the identities of unauthorized drones are unknown and unenumerable during the training phase. In complex electromagnetic environments and with authorized drones operating normally, when unauthorized drones appear, drone identification models based on closed-set classifiers or threshold-based decisions struggle to reliably reject drones outside the training distribution, posing a direct safety hazard to public safety. Therefore, open-set drone identification—that is, how to accurately identify authorized drones and reliably reject enemy / unauthorized equipment—has become a critical issue that urgently needs to be addressed in the low-altitude airspace security governance system.

[0005] Research on open-set identification methods for unmanned aerial vehicles (UAVs) faces three main challenges: First, UAV radio frequency fingerprints originate from minor hardware defects such as transmitter nonlinearity, carrier frequency offset, and phase distortion, and are susceptible to interference from channel fading, noise, and complex electromagnetic environments on the battlefield. This manifests as high similarity and low separability in the time-frequency domain, making it difficult to reliably obtain separable features from different individuals using fixed feature extraction patterns. Second, while individual UAVs of the same model exhibit highly similar radio frequency fingerprints, the same individual may show significant intra-class dispersion under different environments, receivers, or geographical scenarios, making it difficult to simultaneously achieve tight intra-class clustering and clear inter-class separation. Third, the individual identities and radio frequency features of hostile / non-cooperative UAVs are completely unavailable during the training phase. General open-set recognition (OSR) methods are primarily geared towards visual tasks and do not explicitly consider the weak separability of radio frequency signals and the time-varying nature of channel interleaving during discrimination, making it difficult to provide a reliable unauthorized denial mechanism for anti-UAV systems. Summary of the Invention

[0006] In view of this, the purpose of this invention is to provide a non-cooperative UAV open set perception method based on multi-scale representation learning. This method consists of four functional modules: adaptive multi-scale time-frequency feature extraction, known class hyperspherical representation, unknown class boundary perception generation, and open set decision. Specifically, the multi-scale time-frequency feature extraction module alleviates the difficulty of weak fingerprint extraction; the known hyperspherical representation module shapes compact known class geometry; the boundary perception virtual point generation module explicitly regularizes high-risk identity boundaries in the absence of non-cooperative UAV samples; and the open set decision module combines energy residuals and cosine intervals to achieve calibrated non-cooperative UAV rejection. This invention is experimentally verified on a public dataset under various openness conditions. The results show that the proposed RFGuard architecture outperforms baseline algorithms in both authorized UAV identification and non-cooperative UAV rejection, providing support for the development of anti-UAV technologies for key area defense and low-altitude joint defense.

[0007] A non-cooperative open-set perception method for unmanned aerial vehicles (UAVs) based on multi-scale representation learning includes:

[0008] An adaptive multi-scale time-frequency feature extraction module is used to group the radio frequency signals of the UAV to obtain grouped time-frequency samples. Multiple convolutional kernels of different scales are used to extract multi-scale feature maps in parallel from the grouped time-frequency samples. The fusion weight of each scale feature map is determined through a learnable attention mechanism. The multi-scale feature maps are then weighted and fused to obtain a fused feature map. The known class hypersphere representation module is used to map the fused feature map to a unit hypersphere to obtain a normalized representation, and a cosine classifier is used to obtain the classification logits of the known class based on the normalized representation; During the training phase, an unknown class boundary awareness generation module is employed to select a reliable sample pool and an anchor point sample pool from the training samples based on the normalized representation and the classification logits. Based on the samples in the reliable sample pool, a class prototype is estimated for each known class. Based on the class prototype and the samples in the anchor point sample pool, boundary virtual points and difficult virtual points are synthesized to obtain a set of retained virtual points. A joint optimization loss function is then used to train the parameters of the adaptive multi-scale time-frequency feature extraction module and the known class hyperspherical representation module. During the inference phase, an adaptive multi-scale time-frequency feature extraction module is used to extract the fused feature map of the sample to be identified, and a known class hyperspherical representation module is used to obtain the normalized representation and classification logits of the sample to be identified. Then, an open set decision module is used to calculate the free energy score based on the classification logits and to calculate the cosine interval between the predicted class and the strongest competing class. The free energy score and the cosine interval are fused to construct a distribution-normalized energy interval score. The distribution-normalized energy interval score is compared with a preset global rejection threshold to identify authorized or non-cooperative drones.

[0009] Preferably, the adaptive multi-scale time-frequency feature extraction module includes multiple convolutional kernels of different scales, including fine-grained convolutional kernels, medium-scale convolutional kernels, and wide-domain convolutional kernels; the attention mechanism maps the global average pooling vectors of feature maps at each scale to normalized fusion weights through fully connected layers and the Softmax function.

[0010] Preferably, the decision score of the cosine classifier is determined by the cosine similarity between the normalized representation and the learnable classifier parameter vectors of each category, and the classification logits are obtained by scaling the cosine similarity with a scaling factor.

[0011] Preferably, the selection criteria for the reliable sample pool and the anchor sample pool both include: the target class confidence is greater than their respective confidence thresholds, and the cosine interval is greater than their respective cosine interval thresholds; wherein, the target class confidence is the cosine similarity between the sample and its class, and the cosine interval is the difference between the cosine similarity between the sample and its class and the highest cosine similarity between the sample and other known classes.

[0012] Preferably, the boundary virtual points are synthesized as follows: for each class prototype, the competing prototype with the highest cosine similarity is selected, and spherical linear interpolation is performed between the class prototype and the competing prototype, and tangent space perturbation is superimposed; the difficult virtual points are synthesized as follows: for the selected difficult samples, starting from the normalized representation of the sample, tangent noise is superimposed in the direction towards the outside of the class prototype, the direction towards the most difficult competing prototype, and tangential noise.

[0013] Preferably, a joint loss function is constructed based on the reserved set of virtual points and the training samples. The joint loss function includes cross-entropy loss for known authorized drone identity classification, known class representation loss, and virtual point unknown region regularization term.

[0014] Preferably, the formula for calculating the distribution-normalized energy interval score is:

[0015] in, Indicates input sample The fraction of free energy; These are energy targets associated with the predicted empowerment individuals. This corresponds to the intra-class dispersion scale. It is a small constant used for numerical stability. Controlling the contribution of the cosine interval term, To measure the angular advantage of a predicted granting individual relative to its strongest competing granting individual.

[0016] Preferred angular dominance The calculation method is as follows: make express Normalized representation and the first The cosine similarity between the normalized classifier weights of the authorized individuals, the cosine margin is defined as:

[0017] in, , They represent The cosine similarity between the normalized representation of the predicted authorized individual and the normalized classifier weights of authorized individual j.

[0018] Preferably, the method for calculating the free energy fraction is as follows: For any sample, the normalized representation The logits relative to the k-th known class are calculated using a cosine classifier, and are represented as follows: And define the free energy fraction:

[0019] in, This represents the number of authorized drone categories.

[0020] The present invention has the following beneficial effects: 1) This invention proposes a multi-scale time-frequency feature extraction module, which captures fine-grained radio frequency fingerprint features of UAVs from grouped time-frequency samples by dynamically allocating kernel weights across multiple receptive field scales, thus solving the feature truncation problem caused by fixed feature extraction modes in complex electromagnetic environments.

[0021] 2) This invention designs a known class hypersphere representation module, which maps the individual signal features of UAVs to a normalized hypersphere and uses cosine classification. It designs a decision mechanism based on angular geometry to suppress feature norm drift and enhance the compactness of known classes and the separability between classes.

[0022] 3) This invention proposes an unknown class boundary perception generation module, which jointly synthesizes boundary virtual points and difficult virtual points on a normalized hypersphere. During the training process, it can perform structured regular constraints on the boundary of high-risk individual identity decision without relying on real non-cooperative UAV samples.

[0023] 4) This invention designs an open set decision module that integrates energy residual and cosine interval, and proposes a lightweight scoring mechanism based on distributed normalized energy gap (Scatter-Normalized Energy-Gap, SNE). Through scoring calibration, it achieves plug-and-play efficient rejection of non-cooperative UAVs. Attached Figure Description

[0024] Figure 1This is the overall architecture of the RFGuard-UAV of the present invention; Figure 2(a) , 2(b) 2(c) are the confusion matrices of the RFGuard algorithm under different open set settings of U=2, U=4 and U=6 in the embodiments of the present invention, respectively; Figure 3(a) , 3(b) 3(c) are respectively the feature space linear embedding, hyperspherical embedding, and hyperspherical embedding with virtual point constraints of the present invention t-SNE visualization. Detailed Implementation

[0025] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0026] This invention proposes a non-cooperative UAV open set perception (RFGuard) method based on multi-scale representation learning. The overall architecture consists of four functional modules: adaptive multi-scale time-frequency feature extraction, known class hyperspherical representation, unknown class boundary perception generation, and open set decision-making. Each module progressively completes UAV identification and outlier interception determination. The first module, the adaptive multi-scale time-frequency feature extraction module, transcends the limitations of single-scale signal analysis by performing layered mining of UAV radio frequency signals at multiple scales and automatically adjusting feature aggregation weights. This effectively solves the problems of weak hardware fingerprints being difficult to extract effectively and effective features being easily buried by noise in real-world scenarios. The second module, the known class hyperspherical representation module, applies unified geometric spatial constraints to the signal features of authorized UAVs in the field, clustering features of similar devices into a compact and regular feature distribution. This increases the feature spacing between different UAVs, significantly improving the accuracy of authorized UAV identification in the field and enabling precise differentiation between friendly and compliant UAVs. Thirdly, there is the unknown boundary perception virtual point generation module. In practical applications, it is often impossible to collect samples of various non-cooperative, unauthorized, and unidentified drones in advance. This module does not rely on real outlier data. By intelligently generating virtual feature samples of the boundary, it actively delineates the security boundary of the feature domain of legitimate devices, performs explicit constraints and regularization on high-risk areas of the feature space, and predefines the exclusion range of outlier signals. Fourthly, there is the open set intelligent decision-making module. This module integrates the feature energy residual and cosine distance interval as dual judgment criteria to perform comprehensive credibility calibration on the input drone signal. It ensures that authorized drones in the field can pass the identification and verification normally, and can accurately identify illegal external signals, efficiently completing the automatic rejection and interception of non-cooperative and unidentified drones. This invention relies on publicly available radio frequency datasets and has completed comprehensive experimental verification in various open test scenarios. The results show that the proposed RFGuard method outperforms existing mainstream identification algorithms in both the accurate identification of authorized drones and the effective rejection of intrusive non-cooperative drones. It can provide feasible ideas and technical support for the application of anti-drone detection and identification technology in key area security and low-altitude joint defense systems.

[0027] 1. Specific Scheme of the Invention 1.1. Modeling the Open Set Recognition Problem of Unmanned Aerial Vehicles Facing the complex low-altitude electromagnetic warfare environment of the battlefield, considering the existence of Individual authorized drones and The task of open-set drone identification involves two core objectives: first, to accurately identify and classify authorized drones; and second, to reliably reject non-cooperative drones. The received signal can be represented as:

[0028] in, Indicates the time slot of the radiation source The signal sent; Indicates the wireless channel response; Indicates Gaussian white noise; Representing the The device-specific radio frequency fingerprint model of an individual UAV, introduced by transmitter hardware defects, mainly includes three categories: nonlinear amplitude response, carrier frequency offset, and phase distortion. Furthermore, each radio frequency impairment dynamically evolves with changes in device condition and channel conditions, and can be represented as follows:

[0029] in, For signal amplitude, The sampling interval is... Indicates the first Each sampling time. In addition, For carrier frequency, For the first The carrier frequency offset of the drone Indicates the modulation term. This indicates nonlinear amplitude distortion, while This indicates phase distortion.

[0030] To further extract UAV signal features, the received I / Q signal samples can be converted into a time-frequency map (TFM) using a short-time Fourier transform (STFT), denoted as:

[0031] in, This represents the received radio frequency signal after sampling. Indicates length is The window, and These represent the time frame number and the frequency number, respectively. Represents time frame With frequency index The value at that location. To simplify the notation, the time-frequency plot sample obtained from the STFT is denoted as... .make This represents the set of authorized drone identity tags available during the training phase, where The number of authorized drones. This invention defines the three data points as follows: ,in and Representing division The first in Each sample and its corresponding label, and These correspond to the training set, validation set, and test set, respectively. For the training set and validation set, we have... , For the test set, there are { },in This indicates an unauthorized individual drone label.

[0032] The open set recognition model for drones can be represented as: ,in Let represent the trainable parameters. The goal of the UAV open set recognition problem is to maximize the desired recognition accuracy, i.e.:

[0033] in, This represents the expected recognition accuracy. A function representing the true label on the test set. Indicator function. exist The value is 1 if the condition is met, and 0 otherwise.

[0034] Equation (2) and Equation (3) reveal three challenges. First, the inherent amplitude of RF fingerprints, resulting from hardware imperfections, is weak and easily distorted by the superposition of environmental noise, channel fading, and temporal fluctuations, posing a significant challenge to robust fingerprint feature mining. Second, the open-set recognition paradigm requires the model to generalize to training samples of unseen categories, but non-cooperative UAVs in the airspace are diverse in type and have unknown feature distributions, lacking usable prior information. Furthermore, there is high inter-class similarity in RF fingerprints among different UAVs, while the same device exhibits significant intra-class feature differences under complex electromagnetic transmission conditions. In summary, constructing a feature representation space and decision boundary that can accurately distinguish authorized UAV categories and has a stable rejection capability for unknown non-cooperative targets presents significant technical challenges.

[0035] 1.2 RFGuard Architecture To address the aforementioned challenges, this invention proposes a non-cooperative UAV open-set perception (RFGuard) architecture based on multi-scale representation learning, such as... Figure 1 As shown. For the task of identifying non-cooperative drone open sets, RFGuard consists of four collaborative modules: 1) Multi-scale time-frequency feature extraction module: Taking grouped time-frequency samples as input, it captures local texture and cross-time-frequency structure through convolutional kernels with different receptive fields to obtain more robust radio frequency fingerprint features; 2) Given the hyperspherical representation module, the authorized UAV samples are mapped to the normalized spherical feature space, and combined with the cosine metric classifier to achieve compact clustering of intra-class features and effective separation of inter-class features; 3) Unknown class boundary perception generation module, synthesizes boundary virtual points and difficult virtual points based on known class prototypes and competing boundaries, and constrains high-risk areas that may appear in non-cooperative UAV individual samples during the training phase; 4) The open set decision module constructs an SNE confidence score during the inference phase using feature energy response and angular interval, and combines this with an adaptive calibration threshold to complete authorized UAV identity authentication and reliable rejection of non-cooperative unknown UAVs. The design motivation and principles of each module are explained in detail below.

[0036] 1.2.1 Multi-scale time-frequency feature extraction module In open-set UAV identification tasks, the Time-Frequency Map (TFM), as the core feature carrier bearing the transmitter's hardware fingerprint, can jointly characterize the unique time-frequency signature caused by device-specific hardware defects, and is a key basis for distinguishing individual UAVs. However, traditional feature extraction methods use fixed-parameter convolutional kernels to process the TFM, resulting in a fixed receptive field size, which is difficult to adapt to the inherent differences in the time-frequency distribution of different individual UAVs. Constrained by uniform convolutional kernel parameters, static convolution is prone to losing subtle discriminative textures and distorting device-specific fingerprint information, leading to insufficient robustness of feature extraction and significantly reducing the stability and accuracy of UAV identification in complex battlefield low-altitude electromagnetic environments. To address this, this invention proposes an adaptive multi-scale time-frequency feature extraction module, which can adaptively adjust the convolutional kernel combination and receptive field configuration according to the TFM distribution patterns of different individual UAVs, effectively avoiding feature loss and truncation caused by fixed extraction paradigms, and fully extracting fine-grained individual difference features of UAV signals.

[0037] First, the original I / Q signals are used to construct time-frequency pattern samples via STFT. To capture local temporal variations and expand the sample, this invention clips along the time axis with a fixed step size. The grouped samples are obtained and denoted as .in, This indicates the sequence number of the group of samples generated from the same signal segment.V The number of samples in each group. For the first A cropped time-frequency graph sample.

[0038] Given a processed time-frequency sample This invention employs three convolutional kernels of different scales in parallel to extract features, resulting in a feature map. , and Among them, fine-grained convolutional kernels focus on local texture details, medium-scale convolutional kernels capture medium-scale time-frequency patterns, and wide-domain convolutional kernels cover a wider range of global structural information. To make the fusion process adaptive to each sample, global average pooling is first performed on the three feature maps to obtain the aggregated vector. This vector represents the overall structural characteristics of the current time-frequency sample. Subsequently, adaptive weights for the three convolutional kernels are generated through a learnable attention map:

[0039] in, For parameters A learnable attention network maps the summary vector to three initial scores, and then... Normalization ensures that the weights are non-negative and sum to 1. The normalized weights reflect the importance of features at different scales to the current sample. Finally, a weighted summation is used to obtain the fused feature map. The resulting high-dimensional feature maps of the groups It effectively preserves the complementary time-frequency characteristics of the same radio frequency signal at different scales, providing an effective basis for adaptive multi-scale feature extraction for subsequent modules.

[0040] 1.2.2 Known Hyperspherical Representation Module In open set identification scenarios, the feature representations of unauthorized drones often deviate from all known authorized class prototypes. However, the logits score output by traditional classifiers is determined by both feature orientation and feature norm. Even if the feature orientation of unauthorized individual samples differs significantly from that of authorized drones, their larger feature norm can still cause the classifier to produce abnormally high confidence outputs for a certain known class, leading to misclassification and incorrect acceptance. Since open set identification (RFFI) cannot obtain real non-cooperative drone samples during the training phase, it is difficult to directly suppress such overconfident predictions through supervision. Therefore, it is necessary to eliminate the interference caused by norm fluctuations at the feature geometry level, ensuring that classification decisions rely solely on the stable individual identity attribute feature of the representation orientation. To this end, this invention proposes a known class hyperspherical representation module, which constrains the feature space to a unit hypersphere and employs a cosine classifier, ensuring that the classification evidence for known classes is dominated only by the angular similarity between feature vectors, thereby eliminating the impact of norm drift on identification decisions.

[0041] This invention first processes each fused feature map Global average pooling is performed on the spatial dimension to obtain the feature vector. Then, the features from... The feature vectors of the grouped samples are averaged to form the sample representation. The polymerization retains the same radio frequency radiation. The shared identity clues of each cropped sample are used to suppress unnecessary variations.

[0042] Next, the sample will be characterized. conduct The normalization mapping is applied to the unit hypersphere, resulting in the normalized representation of the sample:

[0043] go through After normalization, all samples are constrained to a hypersphere with a unit norm, and inter-class discrimination is mainly determined by the angular similarity between feature vectors. This geometric constraint promotes the formation of clusters with known class features that are compact within classes and separate between classes, providing a clear and robust decision boundary for subsequent open set identification.

[0044] Then, this invention introduces a cosine classifier to be applied to the normalized representation. To calculate the known class logits. Let Indicates the first A vector of learnable classifier parameters for known classes, where This is the embedding dimension. After weight normalization, the decision score is represented by the sample. and The cosine similarity between the defined category directions eliminates the influence of feature magnitude on the classification result.

[0045] Then, the known class logits is calculated as follows: = ; Among them, parameters and All belong to the trainable parameter set ; >0 is the logits scaling factor, which can adjust its numerical range while keeping the cosine response angle order unchanged, thereby obtaining logits that are suitable for subsequent classification loss and energy calculation.

[0046] Finally, the cosine classifier outputs a normalized representation of the samples. logits vectors over all known classes .

[0047] 1.2.3 Unknown Class Perception Boundary Generation Module While the known class hyperspherical representation module effectively enhances the compactness of intra-class features and the separability of inter-class features, it does not impose explicit constraints on rejection regions outside the known class representation space. In the scenario of UAV open set individual recognition, since non-cooperative UAV samples are unavailable during the training phase, decision ambiguities are mainly concentrated near the competitive boundaries between known classes. Inspired by the Virtual Outlier Synthesis (VOS) strategy, this invention proposes an unknown class perception boundary generation module for the normalized hyperspherical representation space. Unlike the original VOS, which randomly samples from low-likelihood regions of feature distribution, this module fully utilizes the geometric properties of the hypersphere to focus the virtual point synthesis process on high-risk decision boundary regions between known classes, thereby achieving structured constraints on the potential distribution of unknown classes.

[0048] make Represents the first [class] from a known class. Normalized representation of each training sample, Represents all known classes of sample i, logits. Let i represent the true known class index of sample i; define the target class confidence and cosine interval as follows:

[0049] in, This represents the cosine similarity between sample i and category k. For sample i and its category cosine similarity, This represents the highest cosine similarity between sample i and other known categories; based on( , Construct two types of complementary sample pools: reliable pool. Used for robust class prototype estimation; anchor pool This is used to detect difficult boundaries and synthesize difficult virtual points. During training, each time a batch of samples is input, two sample pools are selected from the current batch:

[0050] in, and These are the confidence threshold and cosine interval threshold required for reliable prototype estimation, respectively. and It is the threshold corresponding to the anchor point selection.

[0051] For each known class ,make The prototype estimate for the current batch of samples is:

[0052] Among them, Norm( ) Follow the definition in equation (3) Normalization. To mitigate the fluctuations in prototype estimates across different mini-batches, this invention adopts the online update concept of VOS, employing an exponential moving average (EMA) for each category to slide and merge the prototype estimate of the current batch with the historical prototypes of previous batches as the prototype for the current batch.

[0053] Based on the above conditions, this invention synthesizes two types of virtual points: boundary virtual points and hard virtual points, and these two types of virtual points have complementary effects. Boundary virtual points characterize the indistinguishable regions between adjacent class prototypes, thereby providing structured coverage of the decision interface for each class. Hard virtual points anchor selected hard training samples to capture local escape directions caused by intra-class scattering and local boundary irregularities. Using only boundary virtual points will miss such sample-specific hard regions, while using only hard virtual points cannot fully cover the global interface between adjacent classes.

[0054] For each prototype The present invention is based on a selected reliable pool Selection of prototypes in medium sample estimation The highest cosine similarity A competitive prototype and exclude It itself. In this embodiment, =1.

[0055] Subsequently, boundary virtual points are generated by spherical interpolation towards these competing prototypes and superimposing tangent space perturbations:

[0056] in, And Ω = arccos( ).in, This indicates the first number selected based on cosine similarity. A competing prototype These are spherical interpolation coefficients. Control the amplitude of the tangential perturbation. Vector It is a tangential perturbation adapted to the local prototype dispersion. This construction enables boundary virtual points to fill diverse high-risk regions near the prototype point boundaries.

[0057] For each anchor point embed Difficult virtual points are synthesized from difficult samples in known classes; difficult samples are selected based on three criteria: (i) higher energy score; (ii) smaller cosine interval; and (iii) larger offset relative to the class prototype point. Let... Indicates the category to which sample i belongs. The prototype estimation, i.e., the class prototype point, Represents the most difficult competing prototype point, among which Is with The class number of the non-target prototype point with the highest cosine similarity. Difficult virtual points are generated as follows:

[0058] in, Control the overall disturbance step size, and Balancing outward and competitive directions. Controlling the tangential noise amplitude, Pointing away from the source prototype, Pointing to the most difficult competing prototype in the local tangent space. This represents tangential noise. This construction explicitly pushes virtual points into a high-risk region that is both outside the source class and close to competing alternative classes.

[0059] For any sample, the normalized representation This invention calculates its known class logits using a cosine classifier, denoted as: And define the free energy fraction, denoted as . ( ):

[0060] Since not all synthesized virtual points have the same training value, this invention employs a dual screening process for candidate virtual points to avoid retaining virtual points that are too close to known class centers, far from effective boundaries, or have anomalous energy distributions. Let... The type is The set of candidate virtual points, where , and Let the boundary virtual point and the hard virtual point be represented respectively; This represents a candidate virtual point in the set. For each candidate virtual point, first calculate its maximum cosine similarity with all known class prototypes, and require that this value falls within a type-specific interval. This ensures that the virtual point lies near the valid class boundary. Then, their free energies are compared. Energy on the unknown side of the target The difference between them is only valid if the difference does not exceed the energy bandwidth. Only when the energy level is consistent with the expected unknown region is the energy level considered to be consistent. Ultimately, candidate points that simultaneously satisfy both cosine consistency and energy consistency constraints are retained in the virtual point set. The constraints are as shown in formula (14):

[0061] in, It is the minimum cosine threshold. It is a type-specific upper cosine threshold determined based on candidate statistics. It is a type-specific energy range estimated based on the residual statistics of the candidate set. Target energy This specifies the desired free energy fraction level for an unknown class that a virtual point should reach. It is derived from the energy targets of the source class and neighboring classes, with an additional boundary interval added to push the virtual point beyond the boundaries of the corresponding known classes. This invention uses... This indicates the anchor point energy before the boundary interval was added.

[0062] To provide the joint training objective of this invention, let This indicates a mini-batch of the currently known class, where For batch size, For the number of authorized drone categories, For the sample The true known class index. Let To obtain the normalized representation of samples in the two sample pools based on equation (6), The auxiliary projection head output for this sample is the predicted probability vector of sample i across all known categories. For the known class logits generated by the cosine classifier, The free energy fraction is defined by equation (13). The boundary virtual points and difficult virtual points retained after candidate screening are denoted as . and The final set of virtual points is The joint training objective is written as follows:

[0063] in, It is the cross-entropy loss for classifying known authorized drone identities; It is a known class representation loss, which includes a supervised comparison term, a corner margin term for the most difficult negative class, and a known energy-stable term; It is a regularization term for the unknown region of virtual points; It is an optional auxiliary rejection head separation item. , , , , and For the corresponding weights. Here we use... This represents the hardest-gap weights during the training phase, to distinguish them from the SNE scores in the open-set decision module. The known class cross-entropy and supervised comparison terms are defined as follows:

[0064] In equation (16), This represents the set of projected samples with valid positive samples. To monitor and compare temperatures, For the sample With sample Compare the weights of the positive samples. This is the normalization factor. Here we use... Indicates the comparison weights to avoid conflicting with the channel noise mentioned earlier. Or cosine classifier parameters Confusion. For the angular interval term, let... The sample defined by equation (7) For the cosine response of class For the sample set participating in the calculation of the difficult interval term, Preset angular interval; This represents the selected set of known samples participating in the calculation of the known energy stability term. Known energy target. It is obtained by updating the high quantile free energy statistics of the current small batch using EMA. The above two items are defined as follows:

[0065] in, For smooth hinge penalty function; For any reserved virtual point This invention uses unknown side energy constraints, relative energy interval constraints, and source-neighbor boundary balance constraints to construct a regularization term for the unknown region, which can be written as:

[0066] in, For the set of virtual points to be retained, For a single virtual point in the set, For virtual points in unknown regions, regularization loss is applied. The regularization cost for a single virtual point. The free energy fraction defined by equation (13), The unknown side target energy is obtained from the energy targets of the source category and adjacent categories. To add anchor point energy before adding additional boundary spacing, This refers to the relative energy interval. and These represent the cosine responses of the virtual point to the source class prototype and the adjacent competing class prototype, respectively. , and The weights of the unknown side energy constraint, the relative energy interval constraint, and the source-neighbor balance constraint are controlled respectively. The constraint is a smooth hinge penalty function. Three constraints respectively encourage virtual points to approach the unknown side energy target, maintain an energy gap relative to the anchor point, and prevent virtual points from collapsing into a single known category. An optional auxiliary rejection head loss is also included. Used in auxiliary denial heads Upper Separated Virtual Point Set With the current mini-batch of known embeddings .

[0067] 1.2.4 Open Set Decision Module In the inference phase of open-set RFID fingerprinting for drones, the decision-making process needs to simultaneously complete two key tasks: first, predict the identity of authorized drones and output the most likely known class label; second, make a rejection decision for non-cooperative drones, determining whether the current sample belongs to an unknown target not seen during the training phase. Traditional discrimination methods based on maximum Softmax confidence cannot effectively address this challenge: once the RFID fingerprint of a non-cooperative drone falls into the fuzzy decision boundary region between known classes, it may still produce a high logit output for a certain authorized class, thus being misclassified as a friendly target. To address this, this invention designs an open-set decision module for drone scenarios, whose core process consists of two steps: open-set rejection score estimation and adaptive threshold calibration. The rejection score comprehensively utilizes the overall distribution profile (energy term) of the known class logit, as well as the angular interval information between the predicted class prototype and the strongest competing class prototype, to construct an SNE confidence score sensitive to unknown targets; finally, by comparing it with a global calibration threshold, it completes the identity authentication of authorized drones or the reliable rejection of non-cooperative drones.

[0068] Given input sample The cosine classifier outputs known class logits, denoted as... ,in It is related to authorized drone individuals The relevant first Given a known class `logit`, the predicted authorized individual sequence number is:

[0069] For input samples The present invention calculates its free energy fraction according to equation (13). A lower value indicates greater compatibility with known logit profiles, while a higher value indicates weaker compatibility.

[0070] The distribution of the original energy scores changes with the individual drones being predicted, and a single energy score cannot explicitly characterize the proximity of a sample to its competing class. Therefore, this invention designs a Scatter-Normalized Energy-Gap (SNE) score, combining energy-based compatibility with the angular geometry of a cosine classifier. Let... express Normalized representation and the first The cosine similarity between the normalized classifier weights of the authorized individuals, the cosine margin is defined as:

[0071] in, , They represent The cosine similarity between the normalized representation and the predicted authorized individual, and the normalized classifier weights of authorized individual j. This measures the angle advantage of a predicted granting individual relative to its strongest competing granting individual. The RFGuard open-set rejection score is calculated based on the SNE scoring strategy, and is as follows:

[0072] in, These are energy targets associated with the predicted empowerment individuals. This corresponds to the intra-class dispersion scale. It is a small constant used for numerical stability. The contribution of the cosine margin term is controlled. The SNE score consists of two complementary constraints: the first is an energy term, which measures the overall response energy of a sample to the predicted authorized class. This term increases significantly when non-cooperative drone samples produce abnormally high classification responses, thus increasing the rejection tendency; the second is an angular margin term, which calculates the cosine similarity difference between the predicted class prototype and the strongest competing class prototype. When the features of the predicted and competing classes are clearly separated on the hypersphere, this term reduces the open set rejection score, thus increasing the acceptance tendency of authorized drone samples. These two factors work together to allow the SNE score to simultaneously suppress the overconfident responses of non-cooperative samples and enhance the classification reliability of authorized samples. Therefore, a larger... (i.e., larger) A score of 0 indicates stronger evidence that the sample should be rejected as unauthorized. In the threshold calibration step, this score is compared to the global rejection threshold. The final open set prediction is:

[0073] in, This indicates the predicted authorized drone individuals. This indicates an unauthorized drone label. This is the global rejection threshold. Therefore, RFGuard only determines a sample as a known legitimate identity if it matches the features of the predicted authorized individual and maintains sufficient differentiation from other competing authorized individuals. This module relies on the output of the feature extractor and cosine classifier, and makes decisions using only known class log confidence, cosine response, and a single threshold comparison, with minimal additional computational overhead during the inference phase.

[0074] 1.2.5 RFGuard Algorithm The complete workflow of RFGuard is shown below. During the training phase, the following steps are performed sequentially: feature extraction of grouped samples, modeling of known class hypersphere representations, generation of boundary-aware virtual points, and optimization of the joint loss function. During the inference phase, the trained feature extractor and cosine classifier are reused. First, the sample SNE rejection score is calculated, then this score is compared with the calibrated global threshold, and finally, the final judgment result for UAV open set recognition is output.

[0075] Input: Training set Validation set Test samples Training epoch budget; Output: Open set prediction Initialize trainable parameters θ; Set up the training set Each batch M in; The following steps are performed during the training phase: training set In each batch M, samples are converted into grouped samples through a fixed-step long-term sampling process. ; Grouped and fused feature maps are extracted using an adaptive multi-scale feature extraction module. ; Obtain sample representation Normalized representation Known Class And cosine response ; calculate Construct a reliable pool and anchor pool ; Depend on Estimating category prototypes ; Synthetic boundary virtual points and difficult virtual points ; Filter candidate points to obtain the retain set , and ; Calculate joint loss ; Updated via backpropagation θ; Estimating the energy target from training statistics and distribution scale ; exist Use SNE(·) Calibrate global rejection threshold And record .

[0076] The following steps are performed during the testing phase: structure Grouped samples, obtained , and ; Candidate authorization identities are determined from known classes logits according to equation (19). ; Calculate the energy response of the test sample according to equation (13). ; The cosine interval between the prediction class and the competition class is calculated according to equation (20). ; Calculate according to formula (21) ,in ; Determine according to formula (22) This allows us to predict whether a drone belongs to an authorized individual or an unauthorized drone.

[0077] 2. Example 2.1 Experimental Setup Experimental Scenario: The experiment was conducted on the DroneRFb-DIR dataset, which includes three open set settings, denoted as follows: , and ,in This indicates the number of non-cooperative drones in the test set. The training and validation sets contain only authorized drone samples. The training, validation, and test sets contain 6420, 452, and 456 authorized drone signal samples, respectively. In the test set, for... , and The number of unauthorized drone individual samples were 564, 1070, and 1580, respectively.

[0078] Comparison Method: To verify the effectiveness of the proposed method, this invention compares the proposed RFGuard framework with three representative open set identification methods: SoftMax, OpenMax, and VOS. SoftMax uses the maximum SoftMax probability as the baseline method for open set identification results; OpenMax is a typical open set identification method based on score recalibration; and VOS is an open set identification baseline method based on virtual outlier synthesis.

[0079] Evaluation Metrics: This invention uses the Open-Set Classification Rate (OSCR) as the primary evaluation metric, which can jointly reflect the comprehensive performance of authorized drone recognition and non-cooperative drone rejection across thresholds. To further characterize the method's performance in different dimensions of open-set scenarios, this invention also introduces three supplementary metrics: 1) AUROC (Area Under the Receiver Operating Characteristic Curve): Characterizes the model's overall ability to distinguish between known and unknown classes of samples; 2) Macro-F1 (Macro-averaged F1-Score): Measures the balance between recognition accuracy and rejection performance across all categories under threshold decision-making. 3) HNA (Hybrid Normalized Accuracy): The harmonic average of the accuracy of authorized sample recognition and the accuracy of unauthorized sample rejection, which comprehensively reflects the overall performance of open set recognition.

[0080] The specific definitions and calculation methods for each indicator are as follows.

[0081] make and They represent the test set respectively. Authorized and unauthorized subsets in:

[0082] Among them, | | represents the cardinality of the set, 1[ The symbol ] represents the indicator function. The correct classification rate (CCR) of authorized samples and the false positive rate of incorrectly accepting unauthorized samples are defined as follows:

[0083] OSCR can adjust the scan decision threshold. The area obtained from the CCR–FPROSCR curve is calculated as follows:

[0084] AUROC is calculated by treating unauthorized samples as positive classes:

[0085] in, and From respectively and Independent extraction.

[0086] For open set evaluation with a threshold, this invention uses a calibrated threshold. The final prediction obtained Build a Confusion Matrix .forward Rows and columns correspond to authorized identities , No. ( Row and column correspondence .symbol The statistical true assessment serial number is And the prediction and evaluation sequence number is The number of test samples.

[0087] For the evaluation category number {1, , Precision, recall, and F1 are defined as follows:

[0088] The formula for calculating Macro-F1 is:

[0089] HNA is a weighted average of authorized sample accuracy (AKS) and unauthorized sample detection accuracy (AUS), which can be calculated as follows:

[0090] Specific parameter settings: In the data preprocessing stage, each I / Q signal segment is windowed using a HannWindow and the FFT size is adopted. The configuration with a frame shift (hopsize) of 512 is converted into a Short-Time Fourier Transform (STFT) time-frequency image, ultimately generating three sets of images with a size of [missing information]. The grouped samples. During the model training phase, RFGuard uses the Adam optimizer, with 110 epochs of training and an initial learning rate of... The weight decay coefficient is The training batch size was 24, and the evaluation batch size was 64. All experiments were conducted on an Ubuntu 24.04.3LTS system with a single NVIDIA GeForce RTX 3090 GPU. The experiments were implemented using Python 3.10.19 and PyTorch 2.9.1 + cu130, and CUDA 13.0 and cuDNN 9.13 were used to accelerate model training and inference.

[0091] 2.2 Analysis of Experimental Results (1) Algorithm effectiveness analysis First, to verify the effectiveness of the proposed RFGuard algorithm in open-set UAV RF fingerprinting tasks, this invention conducted comparative experiments with three representative methods—SoftMax, OpenMax, and VOS—under three different open-set settings. The results are shown in Table 1. To ensure fairness in the comparison, SoftMax and OpenMax adopted the same adaptive dynamic feature extraction backbone network as RFGuard, only employing the maximum SoftMax confidence and OpenMax score recalibration mechanisms respectively in the open-set decision stage. VOS was implemented according to its framework, using a DenseNet backbone and a virtual outlier synthesis strategy. Therefore, SoftMax and OpenMax benefit from stronger UAV time-frequency feature extraction capabilities and remain competitive in some scenarios. As shown in Table 2, RFGuard achieved the best performance in OSCR, AUROC, Macro-F1, and HNA metrics under all open-set settings. Specifically, in the U=2 scenario, its OSCR reaches 95.64%, an improvement of 4.52 percentage points compared to the second-best OpenMax method. As the openness increases to U=4 and U=6, RFGuard still maintains OSCR values ​​of 92.26% and 90.69%, respectively, improvements of 2.95 and 1.70 percentage points compared to the VOS method, demonstrating good robustness. In terms of the AUROC metric, RFGuard achieves 97.20%, 93.67%, and 92.01% in the three settings, indicating its stronger ability to distinguish between known and unknown class samples. Furthermore, RFGuard's leading performance in Macro-F1 and HNA metrics validates its ability to balance authorized recognition and unauthorized rejection.

[0092] In summary, the experimental results show that the proposed RFGuard algorithm can improve the performance of authorized identity recognition and unauthorized target rejection in different openness scenarios, significantly outperforming the comparison methods and verifying its effectiveness in complex open environments.

[0093] Table 1. Comparison results under different open set settings

[0094] (2) Analysis of ablation test results To evaluate the contribution of each core module to the open set recognition performance, this invention... Ablation experiments were conducted on RFGuard under open-set settings, and the results are shown in Table 2. The four modules M1–M4 correspond to: multi-scale time-frequency feature extraction module, known class hyperspherical representation module, unknown class boundary-aware generation module, and open-set decision module based on SNE scoring, respectively. To ensure consistency in the comparison logic, modules were replaced as follows when disabled: when M1 was disabled, the adaptive multi-scale branch was replaced by the basic feature extraction branch; when M2 was disabled, hyperspherical normalization and cosine classifiers were replaced by traditional linear classifiers; when M3 was disabled, boundary virtual points or difficult virtual points were not synthesized during training; when M4 was disabled, the open-set score used the maximum softmax probability (MSP), otherwise, the SNE score with threshold calibration was used as the open-set rejection score. Since both M3 and M4 are built upon the hyperspherical / cosine feature space output by M2, invalid configurations activating these two modules when M2 was disabled were excluded.

[0095] Table 2. Experimental results of ablation using the RFGuard algorithm.

[0096]

[0097] The ablation results show that RFGuard's performance improvement exhibits a clear modular progression, with the coupling of multi-scale time-frequency feature extraction (M1) and known-class hyperspherical representation (M2) being the main source of performance gain. When only M1 is enabled, the OSCR only slightly improves from the baseline of 81.24% to 81.86%; however, when both M1 and M2 are enabled simultaneously, the OSCR significantly improves to 91.11%, indicating that hyperspherical representation can effectively utilize multi-scale features to construct a more separable known-class feature space. Furthermore, introducing boundary-aware unknown-class modeling (M3) further improves the OSCR to 92.69% and HNA to 84.26%, demonstrating that boundary virtual point synthesis enhances the model's ability to distinguish open set boundaries. Ultimately, the complete model (M1+M2+M3+M4) achieved OSCR and HNA metrics of 95.64% and 91.98%, respectively, validating the open set rejection score estimation and threshold calibration mechanism based on SNE scoring, which can significantly improve the rejection capability of unauthorized targets while maintaining the recognition performance of authorized UAVs.

[0098] (3) Visualization results analysis To intuitively verify the recognition and rejection performance under different openness scenarios, Figure 2(a) , 2(b)Figure 2(c) shows the confusion matrix of RFGuard's recognition results in the prediction set under three open set settings (U=2,4,6). The matrix rows correspond to the real labels, the columns to the predicted labels, the values ​​in parentheses represent the sample size, and the cell color indicates the recognition ratio after row normalization. Regarding the performance of authorized drone recognition, the prediction results for authorized drone samples are highly concentrated on the diagonal position under all settings, indicating that the proposed method has stable and reliable recognition capabilities for known authorized drone individuals, and the recognition performance is not significantly affected by changes in the open set size. Regarding the performance of non-authorized rejection, as the number of non-authorized individuals increases from 2 to 6, RFGuard's rejection rate for non-cooperative drones decreases, but it still maintains a high rejection level in high-openness scenarios. Specifically, the false acceptance rate of non-cooperative drones is consistently kept at a low level, indicating that the decision mechanism based on SNE scoring and calibrating the global rejection threshold can effectively suppress the risk of non-cooperative drones being incorrectly identified as authorized individuals, significantly improving the rejection robustness in the open set environment while ensuring the accuracy of authorized recognition. Further comparison of the matrix distribution under different degrees of openness reveals that as the number of non-cooperative drones increases, the confusion between authorized drones remains within a very small range, and the interference of non-cooperative drones on the classification results of authorized drones does not show significant deterioration. This verifies the overall recognition and rejection stability of the proposed method in complex open scenarios.

[0099] Figure 3(a) , 3(b) Figure 3(c) shows the visualization results of t-SNE dimensionality reduction in the feature space at different stages, intuitively revealing the impact of hyperspherical representation and virtual point generation on feature distribution. The following conclusions can be drawn through observation and analysis: First, in the linear embedding without the M2 module (Figure a), the known class features are loosely distributed, with significant overlap between some categories, weak intra-class clustering, and blurred inter-class boundaries, making it difficult for the model to construct clear recognition and rejection boundaries. Second, after introducing the known class hyperspherical representation (M2) (Figure b), all known class features are mapped onto the hypersphere, significantly enhancing intra-class clustering and making the boundaries between different categories clearer, forming a compact and highly discriminative feature space, providing a stable feature foundation for authorized class recognition. Finally, after introducing boundary-aware virtual point synthesis (M2+M3) (Figure c), a large number of virtual boundary points are generated around the known class feature clusters. These points form an enclosure of the unknown class in the feature space, effectively expanding the open set rejection boundary, making it easier for unknown class samples to be judged as unauthorized, thereby directly improving the model's ability to reject unauthorized drones.

[0100] Addressing the practical needs of counter-drone operations in complex electromagnetic environments, this research focuses on the core issues of accurate identification of authorized drones and reliable rejection of unauthorized threat targets. Key technologies and architecture design for open-set identification of non-cooperative drones were developed. Experimental verification demonstrates that the proposed open-set perception architecture for non-cooperative drones, based on multi-scale representation learning, effectively solves challenges such as weak separability of drone radio frequency fingerprints, large intra-class dispersion, and difficulty in modeling unknown unauthorized targets in battlefield environments. Specifically, the adaptive multi-scale time-frequency feature extraction module dynamically captures fine-grained radio frequency fingerprints of drones at different scales, effectively mitigating the instability of feature extraction caused by complex electromagnetic interference. The known-class hyperspherical representation module significantly improves the intra-class compactness and inter-class separability of known authorized drone features, suppressing feature norm drift in complex environments. The combination of unknown-class boundary perception synthesis and open-set decision-making modules achieves effective regularization and calibrated reliable rejection of high-risk identity boundaries even without real unauthorized samples. Multi-openness experiments demonstrate that the proposed architecture significantly outperforms existing representative open-set baseline methods in terms of authorized UAV individual identification accuracy and unauthorized threat target denial reliability. This has significant practical implications for improving the battlefield's ability to identify and deny enemy UAVs, modified loitering munitions, and other aerial threats. However, the research still has certain limitations. The architecture's adaptability to complex battlefield scenarios with strong electromagnetic interference and densely intertwined multiple targets requires further optimization. Future research will focus on improving cross-domain generalization capabilities and real-time performance optimization to further promote the technology's practical deployment.

[0101] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A non-cooperative open-set perception method for unmanned aerial vehicles based on multi-scale representation learning, characterized in that, include: An adaptive multi-scale time-frequency feature extraction module is used to group the radio frequency signals of the UAV to obtain grouped time-frequency samples. Multiple convolutional kernels of different scales are used to extract multi-scale feature maps in parallel from the grouped time-frequency samples. The fusion weight of each scale feature map is determined through a learnable attention mechanism. The multi-scale feature maps are then weighted and fused to obtain a fused feature map. The known class hypersphere representation module is used to map the fused feature map to a unit hypersphere to obtain a normalized representation, and a cosine classifier is used to obtain the classification logits of the known class based on the normalized representation; During the training phase, an unknown class boundary awareness generation module is employed. Based on the normalized representation and the classification logits, a reliable sample pool and an anchor point sample pool are selected from the training samples. Based on the samples in the reliable sample pool, a class prototype is estimated for each known class. Based on the class prototype and the samples in the anchor point sample pool, boundary virtual points and difficult virtual points are synthesized to obtain a set of retained virtual points. A joint optimization loss function is then used to train the parameters of the adaptive multi-scale time-frequency feature extraction module and the known class hyperspherical representation module. During the inference phase, the fusion feature map of the sample to be identified is extracted using the adaptive multi-scale time-frequency feature extraction module, and the normalized representation and classification logits of the sample to be identified are obtained using the known class hyperspherical representation module. Then, the open set decision module is used to calculate the free energy score based on the classification logits and to calculate the cosine interval between the predicted category and the strongest competing category. The free energy score and the cosine interval are fused to construct the distribution normalized energy interval score. The distribution normalized energy interval score is compared with the preset global rejection threshold to identify authorized or non-cooperative drones. The formula for calculating the distribution-normalized energy interval score is as follows: in, Indicates input sample The fraction of free energy; These are energy targets associated with the predicted empowerment individuals. This corresponds to the intra-class dispersion scale. It is a small constant used for numerical stability. Used to control the contribution of the cosine interval term. To measure the angular advantage of a predicted granting individual relative to its strongest competing granting individual.

2. The non-cooperative UAV open set perception method based on multi-scale representation learning according to claim 1, characterized in that, The adaptive multi-scale time-frequency feature extraction module has multiple convolutional kernels of different scales, including fine-grained convolutional kernels, medium-scale convolutional kernels, and wide-domain convolutional kernels; the attention mechanism maps the global average pooling vector of feature maps at each scale to normalized fusion weights through fully connected layers and the Softmax function.

3. The non-cooperative UAV open set perception method based on multi-scale representation learning according to claim 1, characterized in that, The decision score of the cosine classifier is determined by the cosine similarity between the normalized representation and the learnable classifier parameter vectors of each category, and the classification logits are obtained by scaling the cosine similarity with a scaling factor.

4. The non-cooperative UAV open set perception method based on multi-scale representation learning according to claim 1, characterized in that, The selection criteria for the reliable sample pool and the anchor sample pool include: the target class confidence is greater than their respective confidence thresholds, and the cosine interval is greater than their respective cosine interval thresholds; wherein, the target class confidence is the cosine similarity between the sample and its class, and the cosine interval is the difference between the cosine similarity between the sample and its class and the highest cosine similarity between the sample and other known classes.

5. The non-cooperative UAV open set perception method based on multi-scale representation learning according to claim 1, characterized in that, The boundary virtual points are synthesized as follows: for each class prototype, the competing prototype with the highest cosine similarity is selected, and spherical linear interpolation is performed between the class prototype and the competing prototype, and tangent space perturbation is superimposed; the difficult virtual points are synthesized as follows: for the selected difficult samples, starting from the normalized representation of the sample, tangent noise is superimposed in the direction towards the outside of the class prototype, the direction towards the most difficult competing prototype, and tangential noise.

6. The non-cooperative UAV open set perception method based on multi-scale representation learning according to claim 1, characterized in that, A joint loss function is constructed based on the reserved set of virtual points and training samples. The joint loss function includes cross-entropy loss for known authorized drone identity classification, known class representation loss, and virtual point unknown region regularization term.

7. The non-cooperative UAV open set perception method based on multi-scale representation learning according to claim 1, characterized in that, angular dominance The calculation method is as follows: make Indicates sample Normalized representation and the first The cosine similarity between the normalized classifier weights of the authorized individuals, the cosine margin is defined as: in, , Representing samples respectively The cosine similarity between the normalized representation of the predicted authorized individual and the normalized classifier weights of authorized individual j.

8. The non-cooperative UAV open set perception method based on multi-scale representation learning according to claim 1, characterized in that, The method for calculating the free energy fraction is as follows: For any sample, the normalized representation The logits relative to the k-th known class are calculated using a cosine classifier, and are represented as follows: And define the free energy fraction: in, This represents the number of authorized drone categories.

Images