A method and system for open scene unmanned aerial vehicle radio frequency individual identification based on TFG-OIDNet

Abstract

This invention relates to a method and system for open-scene UAV radio frequency individual identification based on TFG-OIDNet, comprising: Step 1: preprocessing the raw I / Q signals of the UAV radio frequency; Step 2: performing known individual category discrimination; Step 3: calculating the prototype fit score representing the degree of fit of the sample to the known category space; Step 4: calculating the distribution score representing the degree of offset of the sample in the file-level feature distribution space; Step 5: standardization processing, and determining the operating point threshold and gate interval based on the known recall rate of the target; Step 6: calculating the final fusion score; Step 7: realizing UAV radio frequency individual identification. This invention improves the accuracy and stability of known individual identification.

Description

Technical Field

[0001] This invention relates to a method and system for radio frequency individual identification of unmanned aerial vehicles (UAVs) in open scenarios based on TFG-OIDNet, belonging to the fields of UAV radio frequency signal processing, wireless spectrum sensing, pattern recognition, and open set identification technology. Background Technology

[0002] As a low-altitude flight platform, unmanned aerial vehicles (UAVs) have been widely used in recent years in scenarios such as inspection, surveying, logistics, security, and emergency rescue. With the continuous increase in the number of UAVs, the demand for UAV target perception, identification, and security supervision in open environments is becoming increasingly prominent. Against this backdrop, UAV identification methods based on radio frequency (RF) signals have gradually become an important technological approach in the field of UAV target perception because they do not rely on optical imaging conditions and can operate in non-line-of-sight and complex lighting environments. Compared with coarse-grained identification based on UAV type, UAV RF individual identification focuses more on the fine-grained differences between different individuals within the same type of device, possessing higher application value but also facing greater technical challenges.

[0003] Most existing UAV RFID methods are based on the closed-set classification assumption, meaning that samples appearing during the testing phase are assumed to come from a known set of categories from the training phase. Under this condition, related methods typically preprocess and perform time-frequency transformation on the raw I / Q signals to construct a time-spectrum graph, which is then input into a neural network. Convolutional networks, residual networks, or modern visual backbone networks are then used to extract features and output category labels. These methods can achieve certain recognition results within the known category range, but their primary goal remains improving classification accuracy among known categories, with insufficient consideration given to the problem of rejecting unknown individuals in open scenarios.

[0004] Furthermore, compared to drone type recognition, the separable information relied upon for individual identification is weaker and more subtle. These differences typically do not manifest as significant overall structural changes, but rather as local texture perturbations, weak response patterns, and non-uniform distributions in the time and frequency dimensions of the time-frequency spectrum. Therefore, using only a general feature extraction structure and a single global statistical approach can easily lead to insufficient representation of fine-grained identity information, making it difficult to learn stable and separable individual identity representations. In open scenarios, this insufficient representation further amplifies the risk of confusion between unknown samples and known categories.

[0005] In real-world open environments, if closed-set discrimination is still used, unknown drones appearing during the testing phase are often forcibly categorized into a known class, leading to misrecognition of unknown individuals and reducing system reliability and security. Especially in complex scenarios, strong local similarities may exist between different unknown subclasses and some known classes, making it easier for unknown samples to cross the decision threshold and be incorrectly received. Furthermore, environmental differences, local subspace offsets, and file-level distribution variations also affect the separability between known and unknown samples, making it difficult for a single discrimination branch to simultaneously achieve both known recognition and unknown rejection capabilities.

[0006] Therefore, there is an urgent need to propose a method for UAV radio frequency individual identification in open scenarios: on the one hand, it can establish more discriminative feature representations for fine-grained, weakly differentiable identity information in the time-spectrum image; on the other hand, based on the identification of known individuals, it can further combine the degree of fit of the sample to the known category space and the distribution offset information under cross-environment conditions to stably reject unknown individuals, thereby improving the accuracy, robustness and engineering practicality of UAV radio frequency individual identification in open scenarios. Summary of the Invention

[0007] To address the shortcomings of existing UAV radio frequency individual identification methods, which are mainly geared towards closed-set classification scenarios and have difficulty effectively handling the problem of misreception of unknown individuals in open environments, this paper proposes a UAV radio frequency individual identification method, system, electronic equipment, and storage medium based on TFG-OIDNet for open scenarios.

[0008] The method of this invention aims to solve the following technical problems simultaneously: First, the separable information between individual UAVs is weak and hidden, and conventional backbone networks have difficulty in reliably extracting the subtle time-frequency differences that have identity significance; Second, the distribution of identity discrimination information in the time and frequency dimensions is non-uniform, and single path statistics or single global pooling methods are prone to compression and loss of fine-grained discrimination information; Third, traditional closed-set methods cannot reliably reject unknown UAVs that have not been seen during the training phase, and are prone to misclassifying unknown samples into a known category, thereby reducing the reliability and security of the recognition system.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: A method for open-scene UAV radio frequency individual identification based on TFG-OIDNet includes the following steps: Step 1: Read the raw I / Q signals of the UAV radio frequency, preprocess the raw I / Q signals, and convert them into a fixed-size time-spectrum representation; Step 2: Input the time-spectrum into a known individual discrimination network, extract the identity features of the samples, and complete the known individual category discrimination; Step 3: Construct an identity prototype scoring branch (IPS Branch) based on the identity features and known individual category prototypes to obtain a prototype fit score that characterizes the degree of fit of the sample to the known category space; Step 4: Construct an environment routing file-level ViM module (ERF-ViM) based on the identity features, file-level descriptors, and environment cluster centers to obtain a distribution score that characterizes the degree of offset of the sample in the file-level feature distribution space; Step 5: Standardize the prototype fit score and distribution score on the development set, and determine the operating point threshold and gate interval based on the known recall rate of the target. Step 6: Input the standardized prototype fit score and distribution score into the dual-branch gated decision fusion module DGDF. Use the prototype fit score as the dominant signal and introduce the distribution score for correction when the sample falls into the gate control interval to obtain the final fusion score. Step 7: Compare the final fusion score with the pre-defined open scene decision threshold. When the final fusion score meets the reception conditions, output the corresponding known drone individual identity; otherwise, determine it as an unknown individual.

[0010] According to a preferred embodiment of the present invention, Step 1 involves converting the raw I / Q signal of the UAV radio frequency into a fixed-size time-spectrum representation; including: The input samples are raw I / Q signals stored in .mat files, which include in-phase components I and quadrature components Q. First, the in-phase components I and quadrature components Q are combined into a complex baseband signal. Then, mean removal is performed, followed by amplitude normalization. A short-time Fourier transform is then performed on the normalized complex signal to obtain a two-dimensional time-frequency amplitude spectrum. Finally, frequency centering, logarithmic amplitude transformation, time dimension clipping or padding, and normalization are performed on the two-dimensional time-frequency amplitude spectrum to generate a time-frequency representation of uniform size, i.e., a fixed-size time-frequency spectrum representation.

[0011] According to a preferred embodiment of the present invention, in Step 2, the time-spectrum image is input into the known individual discrimination network to extract the identity features of the sample and complete the known individual category discrimination; including: The known individual discrimination network includes the anisotropic fingerprint backbone module AFB, the time-frequency dual-axis statistical discrimination neck module TFAxisNeck, and the identity prototype discrimination head IPH; AFB is used to extract hierarchical deep features from the input time-frequency spectrum, and the perception of weak individual differences and non-uniform time-frequency structure is enhanced by anisotropic receptive field modeling and gating enhancement mechanism. TFAxisNeck is used to compress, map and fuse feature maps along the global path, time axis path and frequency axis path respectively, in order to obtain identity features that retain both overall response information and time-frequency dual axis discrimination information. Using IPH, corresponding prototype centers are established for each known individual category. Based on the similarity and distance relationship between sample identity features and category prototypes, the identity of known individuals is determined, and discrimination information related to category prototypes is output.

[0012] According to a preferred embodiment of the present invention, in Step 3, the construction process of the Identity Prototype Scoring Branch (IPS Branch) includes: Let the normalized identity features of the sample output by TFAxisNeck be: , No. The normalized prototype vectors corresponding to the known categories, where... For known category indexes, The temperature coefficient is Then the sample and the first The similarity between the normalized prototype vectors corresponding to known categories is defined as: (1); In equation (1), This is a temperature coefficient used to adjust the scale of the similarity distribution. , Represents the total number of known categories; sample The prototype fit score for a known category space is defined as follows: (2); in, This represents the prototype fit score output by the IPS Branch.

[0013] According to a preferred embodiment of the present invention, in Step 4, an environment routing file-level ViM module ERF-ViM is constructed based on identity features, file-level descriptors, and environment cluster centers to obtain a distribution score characterizing the degree of offset of a sample in the file-level feature distribution space; including: Let the sample The global score is , No. The local score under each environmental cluster is Then, the global score and the local score are standardized separately: (3); (4); in, and Let represent the mean and standard deviation of the global scores, respectively. and They represent the first The mean and standard deviation of local scores within each environmental cluster. To prevent extremely small constants with a denominator of zero; let the sample... The file-level descriptor is , It is the global score. The results after standardization It is the first Local scores under each environmental cluster The standardization results; No. Each environmental cluster center is The temperature parameters of the soft router are Then the sample Assigned to the Soft route weights for each environment cluster Defined as: (5); in, Indicates the total number of environment clusters; the soft-local score is obtained based on the soft route weights. : (6); Let the sample The route distance is The dominant environment cluster index is Then the normalized measure of the route distance Defined as: (7); in, and These represent the mean and standard deviation of the routing distances corresponding to the dominant environment cluster, respectively. Let the basic fusion weight be The route distance adjustment factor is The upper and lower bounds of the fusion weights are respectively and Then adaptive fusion weights Defined as: (8); Where clip represents the clipping function; And based on this, a patch-level fusion score is obtained. : (9); set up This indicates a file-level aggregation operation, and the file-level ViM score is... Defined as: (10); After file-level aggregation, a routing penalty term is introduced to obtain the final ERF-ViM score, which is a distribution score characterizing the degree of offset of a sample in the file-level feature distribution space. : (11); in, Indicates the routing penalty coefficient. Indicates sample The routing distance, i.e., the distance between the sample file-level descriptor and the center of its dominant environment cluster; Indicates the distance to the route The result after standardization.

[0014] According to a preferred embodiment of the present invention, the specific implementation process of Step 5 and Step 6 includes: Let the prototype fit score output by the IPS Branch be... The distribution score of the ERF-ViM output is Then, the two scores are standardized on the development set respectively: (12); (13); in, and Let denot represent the mean and standard deviation of the IPS Branch scores on the development set, respectively. and Let represent the mean and standard deviation of the ERF-ViM score on the development set, respectively; and let the pre-defined known recall target be . Then the working point threshold Known recall rate based on target on the development set The threshold selection operation is determined as follows: (14); in, This means selecting a prototype from the standardized prototype score distribution of the development set that meets the preset known recall target. The threshold; This represents the operating point threshold corresponding to the target recall constraint, and is used to characterize the baseline position of the known sample acceptance and rejection boundary; Let the width of the gated interval be Then the gating indicator variable Defined as: (15); in, Indicates characteristic functions; Indicates sample Prototype fit score The result after standardization; When the sample is within the gated interval, the distribution score of the ERF-ViM branch is introduced for additive correction to obtain the final fusion score. : (16); in, Indicates the fusion coefficient; Indicates sample Distribution of scores The result after standardization; Finally, the decision threshold for open scenarios is determined based on the calibration of the development set. Complete known / unknown decisions: (17); in, Indicates sample The judgment result in the open scenario.

[0015] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the above-described open-scene UAV radio frequency individual identification method based on TFG-OIDNet.

[0016] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described open-scene UAV radio frequency individual identification method based on TFG-OIDNet.

[0017] An open-scenario UAV radio frequency individual identification system based on TFG-OIDNet includes: The data preprocessing module is configured to: read the raw I / Q signals of the UAV, preprocess the raw I / Q signals of the UAV, and convert them into a fixed-size time-spectrum representation; the known individual discrimination module is configured to: input the time-spectrum into the known individual discrimination network, extract the identity features of the samples, and complete the known individual category discrimination; the prototype scoring module is configured to: construct an identity prototype scoring branch (IPS Branch) based on identity features and known individual category prototypes, and obtain a prototype fit score that characterizes the degree of fit of the sample to the known category space; the distribution scoring module is configured to: construct an environment routing file-level ViM module (ERF-ViM) based on identity features, file-level descriptors, and environment cluster centers, and obtain a distribution score that characterizes the degree of offset of the sample in the file-level feature distribution space; The gated fusion decision module is configured to: standardize the prototype fit score and distribution score on the development set, and determine the operating point threshold and gate interval based on the known recall rate of the target; input the standardized prototype fit score and distribution score into the dual-branch gated decision fusion module DGDF, using the prototype fit score as the dominant signal, and introducing the distribution score for correction when the sample falls into the gate interval, to obtain the final fusion score; The drone radio frequency individual identification module is configured to compare the final fusion score with a pre-calibrated open scene decision threshold, and output the corresponding known drone individual identity when the final fusion score meets the reception conditions; otherwise, it is determined to be an unknown individual.

[0018] Compared with the prior art, the present invention has the following beneficial effects: First, this invention uses a known individual discrimination network composed of AFB, TFAxisNeck and IPH to jointly model the individual identity differences of UAVs from three levels: front-end feature extraction, multi-dimensional statistical modeling and discrimination space constraints. This can more effectively extract weak, hidden and directional time-frequency identity features, and improve the accuracy and stability of known individual identification.

[0019] Second, this invention constructs an IPS Branch, enabling the model to directly characterize the degree of fit of samples to the known category space from the prototype discrimination space, providing an interpretable scoring path for open scene discrimination, and providing an effective basis for subsequent rejection of unknown individuals.

[0020] Third, this invention constructs ERF-ViM, which supplements the discriminative information at the distribution level from the perspectives of file-level distribution, environmental routing, and local subspace offset, thereby enhancing the distribution separability between known and unknown samples under complex environments and cross-scene conditions.

[0021] Fourth, this invention constructs a DGDF-gated fusion mechanism, which adaptively corrects the prototype fit score and distribution score near the discrimination boundary, avoiding the instability caused by simple averaging or fixed weighting. This reduces the risk of unknown individuals being mistakenly identified as known categories, and improves the robustness and security of overall discrimination in open scenarios. Fifth, this invention unifies the identification of known individuals and the rejection of unknown individuals in open scenarios within the same identity representation framework. Without relying on additional complex classification networks, it can achieve stable identification of known individuals and effective differentiation of unknown individuals, which has strong engineering practical value and promotion prospects. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall structure of the known individual discrimination network of the present invention; Figure 2 This is a schematic diagram of the dual-branch open set scoring and gating fusion framework of the present invention; Figure 3 This is a schematic diagram of the overall process of the open-scene UAV radio frequency individual identification method based on TFG-OIDNet of the present invention; Figure 4 This is a schematic diagram of the t-SNE visualization of the open set feature space of this invention; Figure 5 This is a schematic diagram of the open set extended confusion matrix in the open set scenario of this invention. Detailed Implementation

[0023] This invention, in conjunction with the accompanying drawings and specific embodiments, describes the open-scene UAV radio frequency individual identification method based on TFG-OIDNet proposed by this invention. It should be noted that the following embodiments are only used to clearly illustrate the technical concept, implementation process, and key modules of this invention, and do not constitute a limitation on the scope of protection of this invention. Equivalent substitutions, improvements, or adjustments made by those skilled in the art based on the disclosure of this invention without departing from the spirit of this invention should all fall within the scope of protection of this invention.

[0024] Example 1 An open-scene UAV radio frequency individual identification method based on TFG-OIDNet, such as Figure 3As shown, this method takes the original RF I / Q signal as input and first converts it into a fixed-size time-spectrum graph. Then, the time-spectrum graph is input into a known individual discrimination network to extract discriminative identity features and distinguish between known individuals. Based on this, the identity prototype scoring branch IPSBranch and the environment routing file-level ViM module ERF-ViM are constructed in parallel to generate open-set scores from the perspectives of prototype space fit and file-level distribution offset, respectively. Finally, the dual-branch gated decision fusion module DGDF is used to standardize, gate, and fuse the two scores, outputting the final decision result as to whether the test sample belongs to a known or unknown individual. This overall process does not simply concatenate closed-set identification and open-set rejection, but rather simultaneously completes the discrimination of known individuals and the distinction of unknown individuals within a unified identity representation space. The specific steps are as follows: Step 1: Read the raw I / Q signals of the UAV stored in .mat files, preprocess the raw I / Q signals, and convert them into a fixed-size time-spectrum representation; Step 2: Input the time-spectrum into the known individual discrimination network, extract the identity features of the samples, and complete the known individual category discrimination; Step 3: Construct the identity prototype scoring branch (IPS Branch) based on the identity features and known individual category prototypes to obtain the prototype fit score, which represents the degree of fit of the sample to the known category space; Step 4: Construct the environment routing file-level ViM module (ERF-ViM) based on the identity features, file-level descriptors, and environment cluster centers to obtain the distribution score, which represents the degree of offset of the sample in the file-level feature distribution space; Step 5: Standardize the prototype fit score and distribution score on the development set, and determine the operating point threshold and gate interval based on the known recall rate of the target; Step 6: Input the standardized prototype fit score and distribution score into the dual-branch gated decision fusion module (DGDF), using the prototype fit score as the dominant signal, and introducing the distribution score for correction when the sample falls into the gate interval to obtain the final fused score;

[0025] Step 7: Compare the final fusion score with the pre-defined open scene decision threshold. When the final fusion score meets the reception conditions, output the corresponding known drone individual identity; otherwise, determine it as an unknown individual.

[0026] Example 2 The difference between the open-scene UAV radio frequency individual identification method based on TFG-OIDNet described in Example 1 and the method described in Example 1 is that: in Step 1, the raw I / Q signal of the UAV radio frequency is converted into a fixed-size time-spectrum representation; including: The input samples are raw I / Q signals stored in .mat files, including in-phase I and quadrature Q components. First, the in-phase I and quadrature Q components are combined into a complex baseband signal. Then, mean removal is performed to eliminate the interference of DC bias on the subsequent spectrum structure. Next, amplitude normalization is applied to reduce scale differences caused by link fading and environmental changes. Then, a short-time Fourier transform is performed on the normalized complex signal to obtain a two-dimensional time-frequency amplitude spectrum. Finally, frequency centering, logarithmic amplitude transformation, time dimension pruning or padding, and normalization are performed on the two-dimensional time-frequency amplitude spectrum to generate a uniform-size time-frequency representation, i.e., a fixed-size time-frequency spectrum representation. This serves as the input for subsequent networks. Compared to the original IQ steady-state form, the time-frequency spectrum retains stronger individual fingerprint information and is more suitable as an input representation for individual UAV identification in open scenarios.

[0027] In Step 2, the time-spectrum image is input into the known individual discrimination network to extract the identity features of the samples and complete the known individual category discrimination; including: The design goal of the known individual discrimination network is to learn more stable, compact, and highly separable identity representations within a known set of drone individual categories. For example... Figure 1 As shown, the known individual discrimination network includes an anisotropic fingerprint backbone module AFB, a time-frequency dual-axis statistical discrimination neck module TFAxisNeck, and an identity prototype discrimination head IPH. This network is a dedicated discrimination network built to address the challenges of "weak inter-class differences, discrete distribution of identity information in a time-frequency two-dimensional space, and the need to provide stable identity representations for subsequent unknown individual discrimination in open scenarios" in UAV radio frequency individual identification tasks. The network is composed of three collaborative parts: AFB, TFAxisNeck, and IPH. AFB enhances the extraction capability of fine-grained time-frequency fingerprint features and anisotropic structures; TFAxisNeck jointly aggregates global statistical information, time-axis discrimination information, and frequency-axis discrimination information; and IPH constructs an identity discrimination space with prototype constraints.

[0028] Anisotropic receptive field (AFF) is used to extract deep features from the input time-frequency spectrogram. Anisotropic receptive field modeling and gating enhancement mechanisms are employed to enhance the perception of subtle individual differences and non-uniform time-frequency structures. The process involves: AFB first receiving the pre-processed time-frequency spectrogram input and then progressively extracting low-level texture features, mid-level local structural features, and high-level semantic features through multi-layer convolutional mapping. During deep feature extraction, anisotropic receptive field modeling and gating enhancement mechanisms are introduced to enhance the network's response to non-uniform structures in the time and frequency directions. This allows the network to not only retain overall semantic information but also more fully respond to local time-frequency textures, weak difference stripes, and directional detail patterns related to individual identity, resulting in a deep feature map for identity determination. AFB introduces anisotropic enhancement mechanisms on top of the hierarchical convolutional backbone, enabling the feature extraction process to not only retain high-level semantic information but also more fully respond to local time-frequency textures and weak directional difference patterns related to individual identity. Due to the weak inter-individual differences and concealed distribution of UAVs, AFB can enhance the perception of non-uniform time-frequency structures and fine-grained local perturbations.

[0029] TFAxisNeck is used to compress, map, and fuse feature maps along the global, time-axis, and frequency-axis paths to obtain identity features that simultaneously retain overall response information and time-frequency dual-axis discriminative information. TFAxisNeck is used for time-frequency dual-axis statistical modeling of deep feature maps output by AFB. Specifically, TFAxisNeck constructs global, time-axis, and frequency-axis paths to compress, map, and fuse feature maps: the global path compresses and maps the feature map as a whole to retain the overall discriminative semantics of the samples; the time-axis path aggregates along the frequency dimension to extract the variation patterns of features in the time dimension to retain identity information related to temporal perturbations; the frequency-axis path aggregates along the time dimension to extract the energy distribution differences of features in the frequency dimension to retain identity information related to spectral structure. Subsequently, the statistical features obtained from the three paths are mapped, concatenated, or weighted and fused, and then normalized to form a unified identity feature representation. The global path preserves the overall discriminative semantics, the time-axis path depicts the local variation patterns of identity information in the time dimension, and the frequency-axis path describes the structural differences in different frequency bands. These three feature maps are then fused and normalized to obtain a unified identity feature representation. This structure can explicitly preserve the distribution differences of identity information in the time-frequency two-dimensional space, alleviating the problem of insufficient expression of fine-grained identity features by single-path statistical methods.

[0030] IPH (Integrated Prototype Center) is used to establish corresponding prototype centers for each known individual category. Based on the similarity and distance relationships between sample identity features and category prototypes, the identity of known individuals is determined, and discrimination information related to the category prototypes is output. IPH is used to construct an identity discrimination space with prototype constraints. Specifically, a corresponding category prototype center is established for each known UAV individual category, and the identity features output by TFAxisNeck are mapped to the prototype discrimination space. By calculating the similarity and distance relationships between sample identity features and each category prototype, the identity determination results of known individuals are output. Compared with ordinary linear classifiers, IPH places greater emphasis on structural constraints in the discrimination space, which can enhance inter-class separation between different individuals and compress the intra-class distribution of samples of the same type for subsequent open scene scoring construction. Compared with ordinary linear classifiers, IPH can enhance inter-class separation and intra-class compactness, providing a more stable identity representation foundation for subsequent open set scoring construction.

[0031] The feature representation of the known individual discrimination network structure is as follows. Let the input spectrogram after preprocessing and time-frequency transformation be denoted as:

[0032] ; in, Indicates the length of the time dimension. This represents the length of the frequency dimension. The goal of a known individual discrimination network is to learn a mapping from the input spectrogram to the identity category:

[0033] ; in, This represents the predicted identity category output by the network. Let the front-end feature extraction mapping corresponding to AFB be... Then, the feature map extracted from the input spectrum by the front-end backbone can be represented as:

[0034] ; Let the global statistics branch, the time axis statistics branch, and the frequency axis statistics branch be denoted as follows: , , and The fusion mapping is denoted as Then the identity features output by TFAxisNeck can be written as: ; This formula shows that the final identity features are not generated by a single path, but are obtained by comprehensively utilizing the overall response, time direction discrimination information and frequency direction discrimination information, thus explicitly preserving the distribution differences of identity information in the time-frequency two-dimensional space.

[0035] The prototype similarity and distance calculation in IPH are as follows. The prototype vectors corresponding to the known individual categories are: The normalized sample features are The normalized category prototype is Then the similarity between a sample and its class prototype can be expressed as:

[0036] ; The corresponding prototype distance can be further expressed as: ; By utilizing both similarity and distance information, IPH is able to construct a more compact and separable identity feature space while maintaining classification and discrimination capabilities.

[0037] The prototype fit score of IPSBranch is as follows. Let the normalized identity features output by TFAxisNeck be... , No. The class prototype is The temperature coefficient is Then the sample With the The similarity between prototypes of each category is defined as:

[0038] ; Furthermore, the sample The prototype fit score for a known category space is defined as follows: ; in, This represents the total number of known categories. The larger the value, the more likely the sample is to form a stable match with a known class prototype; conversely, the smaller the value, the more likely the sample is to deviate from the known class space.

[0039] like Figure 2 As shown, this embodiment illustrates the scoring construction and fusion process in open scenarios. Since unknown individuals may be forcibly categorized into a known class during the testing phase, after the identity features of known individuals are output by the discrimination network, this invention further constructs a dual-branch open-set scoring and gating fusion mechanism to improve the ability to reject unknown individuals in open scenarios. The dual-branch open-set scoring and gating fusion mechanisms include IPSBranch, ERF-ViM, and DGDF.

[0040] In Step 3, the construction process of the Identity Prototype Scoring Branch (IPS Branch) includes: IPSBranch operates on a prototype discriminative space, taking the normalized identity features output by TFAxisNeck as input. It constructs a prototype fit score by calculating the similarity between a sample and each known category prototype. A higher score indicates a more likely stable match between the sample and a known category prototype; a lower score indicates a more likely deviation from the known category space. IPSBranch thus provides a direct, interpretable, and computationally inexpensive basic open-set scoring path.

[0041] Let the normalized identity features of the sample output by TFAxisNeck be: , No. The normalized prototype vectors corresponding to each known category specifically refer to the normalized results of the category prototype centers established by IPH for each known individual drone category in Step 2; the normalized identity features are... This refers to the sample identity features output by TFAxisNeck and normalized. For known category indexes, The temperature coefficient is Then the sample and the first The similarity between the normalized prototype vectors corresponding to known categories is defined as:

[0042] (1); In equation (1), The temperature coefficient is used to adjust the scale of the similarity distribution, so that the similarity score maintains appropriate numerical sensitivity during the training and inference phases. , Represents the total number of known categories; sample The prototype fit score for a known category space is defined as follows:

[0043] (2); in, This represents the prototype fit score output by the IPS Branch. When The larger the value, the better the sample size. The more likely it is to form a stable match with a known category prototype; when The smaller the sample size, the better. The more likely it is to belong to an unknown individual.

[0044] In Step 4, an environment routing file-level ViM module, ERF-ViM, is constructed based on identity features, file-level descriptors, and environment cluster centers to obtain a distribution score characterizing the degree of sample offset in the file-level feature distribution space; including: ERF-ViM operates based on file-level feature distribution. This module first standardizes the global score and local scores within the environment cluster; then, it constructs soft-route weights based on the distance between the sample file-level descriptor and the center of the environment cluster, and calculates the soft-local score accordingly; next, it fuses the global score and soft-local score using adaptive fusion weights; finally, after file-level aggregation, a routing penalty term is added to obtain the final distribution score. ERF-ViM can provide supplementary evidence for open-set discrimination from the perspectives of environmental differences, local distribution shifts, and file-level uncertainties, thereby compensating for the instability that may occur with a single prototype score in complex environments and near boundary samples.

[0045] Let the sample The global score is , This represents the global distribution score obtained based on the global representation of the sample, used to describe the degree of global deviation of the sample relative to the overall known distribution; it is obtained by mapping the global features or global residual representation of the sample through a score. The local score under each environmental cluster is , Indicates the sample in the th Each environment cluster corresponds to a distribution score in its local subspace, used to characterize the degree of local shift of a sample under specific environmental conditions; it is calculated from the sample's shift in the local representation space corresponding to that environment cluster. The global score and local score are then standardized separately.

[0046] (3); (4); in, and Let represent the mean and standard deviation of the global scores, respectively. and They represent the first The mean and standard deviation of local scores within each environmental cluster. To prevent extremely small constants with a denominator of zero; let the sample... The file-level descriptor is , This represents a file-level representation vector obtained by aggregating features from multiple segments within the same test file. It is not an instantaneous feature of a single patch, but rather a statistical representation of the entire file. It is the global score. The standardized results are used to eliminate the influence of different scoring scales and statistical ranges, enabling them to be integrated with local scores for subsequent calculations on a unified scale. It is the first Local scores under each environmental cluster The standardized results are used to improve the comparability of local scores between different environmental clusters.

[0047] No. Each environmental cluster center is , This represents the result obtained after clustering the file-level descriptors of the training or development set. An environmental prototype center. The soft router temperature parameters are... Then the sample Assigned to the Soft route weights for each environment cluster Defined as:

[0048] (5); in, Indicates the total number of environment clusters; the soft-local score is obtained based on the soft route weights. : (6); Let the sample The route distance is The dominant environment cluster index is Then the normalized measure of the route distance Defined as: (7); in, and These represent the mean and standard deviation of the routing distances corresponding to the dominant environment cluster, respectively. Let the basic fusion weight be The route distance adjustment factor is The upper and lower bounds of the fusion weights are respectively and Then adaptive fusion weights Defined as: (8); Here, clip represents the clipping function; it is used to limit the fusion weights within preset upper and lower bounds to avoid abnormal amplification or reduction.

[0049] And based on this, a patch-level fusion score is obtained. : (9); set up This indicates a file-level aggregation operation, and the file-level ViM score is... Defined as: (10); After file-level aggregation, a routing penalty term is introduced to obtain the final ERF-ViM score, which is a distribution score characterizing the degree of offset of a sample in the file-level feature distribution space. : (11); in, Indicates the routing penalty coefficient. Indicates sample The routing distance, i.e., the distance between the sample file-level descriptor and the center of its dominant environment cluster; Indicates the distance to the route The result after standardization. It reflects the degree of standardized shift of the sample relative to the center of the dominant environmental cluster.

[0050] DGDF is used to complete the gated fusion decision. Instead of simple averaging or fixed weighting, DGDF uses the prototype fit score output by IPSBranch as the dominant signal. Only when a sample is close to the boundary neighborhood of the operating point threshold is the distribution score of the ERF-ViM branch introduced for additive correction. Thus, when the sample is far from the discrimination boundary, the system maintains the dominant role of the prototype score; when the sample is near the boundary, the system introduces distribution evidence to correct the final decision, thereby reducing the risk of misreceiving unknown individuals and improving overall discrimination stability.

[0051] The specific implementation process of Step 5 and Step 6 includes: Let the prototype fit score output by the IPS Branch be... The distribution score of the ERF-ViM output is Then, the two scores are standardized on the development set respectively: (12); (13); in, and Let denot represent the mean and standard deviation of the IPS Branch scores on the development set, respectively. and Let represent the mean and standard deviation of the ERF-ViM score on the development set, respectively; and let the pre-defined known recall target be . Then the working point threshold Known recall rate based on target on the development set The threshold selection operation is determined as follows: (14); in, This means selecting a prototype from the standardized prototype score distribution of the development set that meets the preset known recall target. The threshold; This represents the operating point threshold corresponding to the target recall constraint, and is used to characterize the baseline position of the known sample acceptance and rejection boundary; Let the width of the gated interval be Then the gating indicator variable Defined as: (15); in, Indicates characteristic functions; Indicates sample Prototype fit score The results are standardized to eliminate differences in numerical scale and statistical range between the prototype scores, enabling them to be compared and integrated with the distributed scores on a uniform scale.

[0052] When the sample is within the gated interval, the distribution score of the ERF-ViM branch is introduced for additive correction to obtain the final fusion score. : (16); in, Indicates the fusion coefficient; Indicates sample Distribution of scores The results are standardized; used to eliminate differences in numerical scale and statistical range of the distributed scores, making them comparable to the standardized prototype scores. Gating fusion is performed at a unified scale. Ultimately, the decision threshold for open scenarios is determined based on the development set calibration. Complete known / unknown decisions:

[0053] (17); in, Indicates sample The judgment result in the open scenario. This represents the final open-scene decision threshold, which is determined by the fusion score on the development set. The statistical distribution is obtained and used to distinguish between known and unknown individual samples. Since the distribution range of the fusion score may differ under different datasets, different scenario conditions, and different model parameter settings, therefore... It is not fixed to a certain constant in advance, but is adaptively selected according to the preset judgment target within the actual distribution range of the development set fusion score.

[0054] This embodiment describes the training and inference phases. During the training phase, a known individual discrimination network is used as the main component. Discriminative identity representations are learned collaboratively through AFB, TFAxisNeck, and IPH, making similar samples more clustered in the feature space and maintaining clearer boundaries between different individuals. After training, the model can not only discriminate the identities of known individuals but also output identity features with prototype constraints, providing a foundation for subsequent open-set scoring.

[0055] During the inference phase, the test sample first passes through a known individual discrimination network to obtain identity features and category prototype information; then, two scoring branches, IPSBranch and ERF-ViM, are constructed in parallel; next, DGDF performs standardization and fusion under development set calibration statistics and gating interval constraints; finally, the predicted result for known identity or the result for unknown identity is output. For file-level decision-making, multiple segments within the same file can be aggregated for scoring to improve the overall stability of the decision.

[0056] See Figures 4 to 5 This embodiment illustrates the application effect of the method of the present invention in the task of radio frequency individual identification of unmanned aerial vehicles (UAVs) in open scenarios. In specific applications, the radio frequency sample of the UAV under test is preprocessed and converted into a time-spectrum image, which is then input into a known individual discrimination network for identity feature extraction. Subsequently, corresponding scores are generated by the identity prototype scoring branch (IPS Branch) and the environment routing file-level ViM module (ERF-ViM), respectively. The final open decision result is output by the dual-branch gated decision fusion module (DGDF) to determine whether the sample belongs to a known individual or an unknown individual. Figure 4 As shown, the sample features extracted by the method of this invention form a relatively clear distribution structure in the feature space, with a good separation trend between known and unknown individual samples. This indicates that the present invention can learn discriminative identity representations and provide effective support for known recognition and unknown rejection in open scenarios. Figure 5 As shown, in open-set scenarios, the method of this invention can effectively distinguish between known and unknown individuals, and the situation where unknown samples are mistakenly received as known categories is effectively suppressed. This demonstrates that the dual-branch open-set scoring and gating fusion mechanism constructed in this invention can improve the accuracy and reliability of UAV radio frequency individual identification in open scenarios. In summary, this embodiment shows that the method of this invention can effectively identify known UAV individuals and effectively reject unknown UAV individuals in open scenarios, and has good application value.

[0057] Example 3 A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the open-scene UAV radio frequency individual identification method based on TFG-OIDNet described in Embodiment 1 or 2.

[0058] Example 4 A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the open-scene UAV radio frequency individual identification method based on TFG-OIDNet as described in Embodiment 1 or 2.

[0059] Example 5 An open-scenario UAV radio frequency individual identification system based on TFG-OIDNet includes: The data preprocessing module is configured to: read the raw I / Q signals of the UAV stored in .mat files, preprocess the raw I / Q signals, and convert them into a fixed-size time-spectrum representation; the known individual discrimination module is configured to: input the time-spectrum into the known individual discrimination network, extract the identity features of the samples, and complete the known individual category discrimination; the prototype scoring module is configured to: construct an identity prototype scoring branch (IPS Branch) based on identity features and known individual category prototypes to obtain a prototype fit score characterizing the degree of fit of the sample to the known category space; the distribution scoring module is configured to: construct an environment routing file-level ViM module (ERF-ViM) based on identity features, file-level descriptors, and environment cluster centers to obtain a distribution score characterizing the degree of offset of the sample in the file-level feature distribution space; The gated fusion decision module is configured to: standardize the prototype fit score and distribution score on the development set, and determine the operating point threshold and gate interval based on the known recall rate of the target; input the standardized prototype fit score and distribution score into the dual-branch gated decision fusion module DGDF, using the prototype fit score as the dominant signal, and introducing the distribution score for correction when the sample falls into the gate interval, to obtain the final fusion score; The drone radio frequency individual identification module is configured to compare the final fusion score with a pre-calibrated open scene decision threshold, and output the corresponding known drone individual identity when the final fusion score meets the reception conditions; otherwise, it is determined to be an unknown individual.

Claims

1. A method for radio frequency identification of unmanned aerial vehicles (UAVs) in open scenarios based on TFG-OIDNet, characterized in that, The process includes the following steps: Step 1: Read the raw I / Q signals of the UAV radio frequency, preprocess the raw I / Q signals, and convert them into a fixed-size time-spectrum representation; Step 2: Input the time-spectrum into the known individual discrimination network, extract the identity features of the samples, and complete the known individual category discrimination; Step 3: Construct the identity prototype scoring branch (IPS Branch) based on the identity features and known individual category prototypes to obtain the prototype fit score, which represents the degree of fit of the sample to the known category space; Step 4: Construct the environment routing file-level ViM module (ERF-ViM) based on the identity features, file-level descriptors, and environment cluster centers to obtain the distribution score, which represents the degree of offset of the sample in the file-level feature distribution space; Step 5: Standardize the prototype fit score and distribution score on the development set, and determine the operating point threshold and gate interval based on the known recall rate of the target; Step 6: Input the standardized prototype fit score and distribution score into the dual-branch gated decision fusion module (DGDF), using the prototype fit score as the dominant signal, and introducing the distribution score for correction when the sample falls into the gate interval to obtain the final fused score; Step 7: Compare the final fusion score with the pre-defined open scene decision threshold. When the final fusion score meets the reception conditions, output the corresponding known drone individual identity; otherwise, determine it as an unknown individual.

2. The open-scene UAV radio frequency individual identification method based on TFG-OIDNet according to claim 1, characterized in that, Step 1 involves converting the raw I / Q signals of the UAV radio frequency into a fixed-size time-spectrum representation; this includes: The input samples are raw I / Q signals stored in .mat files, which include in-phase components I and quadrature components Q. First, the in-phase components I and quadrature components Q are combined into a complex baseband signal. Then, mean removal is performed, followed by amplitude normalization. A short-time Fourier transform is then performed on the normalized complex signal to obtain a two-dimensional time-frequency amplitude spectrum. Finally, frequency centering, logarithmic amplitude transformation, time dimension clipping or padding, and normalization are performed on the two-dimensional time-frequency amplitude spectrum to generate a time-frequency representation of uniform size, i.e., a fixed-size time-frequency spectrum representation.

3. The open-scene UAV radio frequency individual identification method based on TFG-OIDNet according to claim 1, characterized in that, In Step 2, the time-spectrum image is input into the known individual discrimination network to extract the identity features of the samples and complete the known individual category discrimination; including: The known individual discrimination network includes the anisotropic fingerprint backbone module AFB, the time-frequency dual-axis statistical discrimination neck module TFAxisNeck, and the identity prototype discrimination head IPH; AFB is used to extract hierarchical deep features from the input time-frequency spectrum, and the perception of weak individual differences and non-uniform time-frequency structure is enhanced by anisotropic receptive field modeling and gating enhancement mechanism. TFAxisNeck is used to compress, map and fuse feature maps along the global path, time axis path and frequency axis path respectively, in order to obtain identity features that retain both overall response information and time-frequency dual axis discrimination information. Using IPH, corresponding prototype centers are established for each known individual category. Based on the similarity and distance relationship between sample identity features and category prototypes, the identity of known individuals is determined, and discrimination information related to category prototypes is output.

4. The open-scene UAV radio frequency individual identification method based on TFG-OIDNet according to claim 1, characterized in that, In Step 3, the construction process of the Identity Prototype Scoring Branch (IPS Branch) includes: Let the normalized identity features of the sample output by TFAxisNeck be: , No. The normalized prototype vectors corresponding to the known categories, where... For known category indexes, The temperature coefficient is Then the sample and the first The similarity between the normalized prototype vectors corresponding to known categories is defined as: （1）； In equation (1), This is a temperature coefficient used to adjust the scale of the similarity distribution. , Represents the total number of known categories; sample The prototype fit score for a known category space is defined as follows: （2）； in, This represents the prototype fit score output by the IPS Branch.

5. The open-scene UAV radio frequency individual identification method based on TFG-OIDNet according to claim 1, characterized in that, In Step 4, an environment routing file-level ViM module, ERF-ViM, is constructed based on identity features, file-level descriptors, and environment cluster centers to obtain a distribution score characterizing the degree of sample offset in the file-level feature distribution space; including: Let the sample The global score is , No. The local score under each environmental cluster is Then, the global score and the local score are standardized separately: （3）； （4）； in, and Let represent the mean and standard deviation of the global scores, respectively. and They represent the first The mean and standard deviation of local scores within each environmental cluster. To prevent extremely small constants with a denominator of zero; let the sample... The file-level descriptor is , It is the global score. The results after standardization It is the first Local scores under each environmental cluster The standardization results; No. Each environmental cluster center is The temperature parameters of the soft router are Then the sample Assigned to the Soft route weights for each environment cluster Defined as: （5）； in, Indicates the total number of environment clusters; the soft-local score is obtained based on the soft route weights. : （6）； Let the sample The route distance is The dominant environment cluster index is Then the normalized measure of the route distance Defined as: （7）； in, and These represent the mean and standard deviation of the routing distances corresponding to the dominant environment cluster, respectively. Let the basic fusion weight be The route distance adjustment factor is The upper and lower bounds of the fusion weights are respectively and Then adaptive fusion weights Defined as: （8）； Where clip represents the clipping function; And based on this, a patch-level fusion score is obtained. : （9）； set up This indicates a file-level aggregation operation, and the file-level ViM score is... Defined as: （10）； After file-level aggregation, a routing penalty term is introduced to obtain the final ERF-ViM score, which is a distribution score characterizing the degree of offset of a sample in the file-level feature distribution space. : （11）； in, Indicates the routing penalty coefficient. Indicates sample The routing distance, i.e., the distance between the sample file-level descriptor and the center of its dominant environment cluster; Indicates the distance to the route The result after standardization.

6. A method for radio frequency identification of unmanned aerial vehicles (UAVs) in open scenarios based on TFG-OIDNet according to any one of claims 1-5, characterized in that, The specific implementation process of Step 5 and Step 6 includes: Let the prototype fit score output by the IPS Branch be... The distribution score of the ERF-ViM output is Then, the two scores are standardized on the development set respectively: （12）； （13）； in, and Let denot represent the mean and standard deviation of the IPS Branch scores on the development set, respectively. and Let represent the mean and standard deviation of the ERF-ViM score on the development set, respectively; and let the pre-defined known recall target be . Then the working point threshold Known recall rate on the development set that meets the target The threshold selection operation is determined as follows: （14）； in, This means selecting a prototype from the standardized prototype score distribution of the development set that satisfies the preset known recall target. The threshold; This represents the operating point threshold corresponding to the target recall constraint, and is used to characterize the baseline position of the known sample acceptance and rejection boundary; Let the width of the gated interval be Then the gating indicator variable Defined as: （15）； in, Indicates characteristic functions; Indicates sample Prototype fit score The result after standardization; When the sample is within the gated interval, the distribution score of the ERF-ViM branch is introduced for additive correction to obtain the final fusion score. : （16）； in, Indicates the fusion coefficient; Indicates sample Distribution of scores The result after standardization; Finally, the decision threshold for open scenarios is determined based on the calibration of the development set. Complete known / unknown decisions: （17）； in, Indicates sample The judgment result in the open scenario.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the open-scene UAV radio frequency individual identification method based on TFG-OIDNet as described in any one of claims 1-6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the open-scene UAV radio frequency individual identification method based on TFG-OIDNet as described in any one of claims 1-6.

9. An open-scenario UAV radio frequency individual identification system based on TFG-OIDNet, characterized in that, include: The data preprocessing module is configured to: read the raw I / Q signals of the UAV radio frequency, preprocess the raw I / Q signals of the UAV radio frequency, and convert them into a fixed-size time-spectrum representation; The known individual discrimination module is configured to: input the time-spectrum image into the known individual discrimination network, extract the identity features of the sample, and complete the known individual category discrimination; the prototype scoring module is configured to: construct the identity prototype scoring branch (IPS Branch) based on the identity features and the known individual category prototype, and obtain the prototype fit score that represents the degree of fit of the sample to the known category space; The distribution scoring module is configured to: construct an environment routing file-level ViM module ERF-ViM based on identity features, file-level descriptors, and environment cluster centers to obtain a distribution score that characterizes the degree of offset of a sample in the file-level feature distribution space; The gated fusion decision module is configured to: standardize the prototype fit score and distribution score on the development set, and determine the operating point threshold and gate interval based on the known recall rate of the target; input the standardized prototype fit score and distribution score into the dual-branch gated decision fusion module DGDF, using the prototype fit score as the dominant signal, and introducing the distribution score for correction when the sample falls into the gate interval, to obtain the final fusion score; The drone radio frequency individual identification module is configured to compare the final fusion score with a pre-calibrated open scene decision threshold, and output the corresponding known drone individual identity when the final fusion score meets the reception conditions; otherwise, it is determined to be an unknown individual.

Images