Unmanned aerial vehicle radio open set rejection and anomaly alerting method and system

Abstract

The application provides a kind of unmanned aerial vehicle radio open set recognition and abnormality alarm method and system, belong to unmanned aerial vehicle radio signal perception safety technical field, the application adopts the deep learning network of double representation branch decoupling, decouples network representation into two kinds of branches of identity discrimination and physical agent, and the discrimination information used for category identification is output simultaneously with the predicted physical agent vector in forward inference;Based on the discrimination information, the open set discrimination score is calculated, and the unknown target recognition decision is completed;Combined with the reference physical agent vector calculated from the original signal, the consistency deviation is used to complete the abnormality discrimination and alarm of known target.The application improves the recognition stability in complex environment, can accurately identify unknown targets and countermeasure abnormal signals, the judgment result is traceable, and meets the high reliability demand of low-altitude supervision.

Description

Technical Field

[0001] This invention relates to the field of radio signal sensing security technology for unmanned aerial vehicles (UAVs), and in particular to a method and system for radio open-set rejection and anomaly alarm for UAVs. Background Technology

[0002] With the gradual opening of low-altitude airspace and the large-scale application of drones in logistics, urban patrols, emergency rescue, and public safety, the monitoring system faces multiple challenges, including an increase in the number of flying targets, a more complex operating environment, and rising potential risks. Abnormal drone behavior is becoming increasingly covert and diverse, encompassing unauthorized flights, intrusions, and evasion of regulations, as well as adversarial behaviors such as signal spoofing, identity impersonation, replay deception, and deliberate interference. Such anomalies can easily disrupt airspace order, increase risks in key areas, and overburden emergency response.

[0003] Current drone surveillance technologies generally rely on radio signal sensing to detect and identify targets. The common process involves collecting and extracting features from the drone's radio signals, then classifying or identifying the target based on these features. While this method offers advantages such as being non-contact, having wide coverage, and flexible deployment, it still has limitations in complex propagation environments and under various operating conditions: First, factors such as multipath propagation in urban areas, obstruction, signal-to-noise ratio fluctuations, and frequency shifts can cause signal features to drift with the environment and time, leading to unstable identification results. Second, new drone models, new firmware protocols, or modified equipment often appear in surveillance scenarios that were not covered during the training phase. Closed-category identification strategies often force unknown targets into known categories, easily causing misjudgments and missed detections. Third, adversarial attacks against surveillance models are increasing. Attackers can create highly similar deceptive signals through signal replay, modulation camouflage, and feature simulation, thereby misleading the identification system and reducing the reliability of surveillance.

[0004] Furthermore, while existing deep learning-based identification methods can achieve good recognition performance under certain conditions, they generally lack verifiable criteria for practical application. If the model output lacks a constraint mechanism consistent with the statistical characteristics of radio signals, it is prone to producing erroneous outputs with excessively high confidence when channel conditions change abruptly or when subjected to deceptive interference. This makes it difficult to meet the requirements of regulatory systems for reliability, controllable false alarm rates, and traceable criteria. Therefore, there is an urgent need for a method and system for UAV radio signal identification and anomaly alarm for low-altitude airspace regulation. This system should be able to simultaneously perform fingerprint representation, open-set rejection decision, and consistency anomaly discrimination under a unified radio sensing link, enabling reliable identification and alarming of unknown targets and abnormal signals. It should also maintain stable performance under complex propagation environments and various operating conditions to improve the real-time performance, deployability, and scalability of regulatory systems. Summary of the Invention

[0005] This invention provides a method and system for radio open set rejection and anomaly alarm for unmanned aerial vehicles (UAVs) to address the deficiencies in the prior art.

[0006] In a first aspect, the present invention provides a method for radio open set rejection and anomaly alarm of unmanned aerial vehicles (UAVs), comprising: preprocessing the collected UAV radio signals to form a standardized input tensor; The input tensor is fed into the trained dual-representation branch decoupled deep learning network, which outputs discriminant information for known category identification and a predicted physical surrogate vector for describing the physical statistical characteristics of the signal. Based on the discrimination information, the open set discrimination score is calculated, and the target is determined to belong to the known category set according to the preset open set threshold. If the target is determined to be of an unknown type, output a rejection result and trigger a regulatory alert. If the target is determined to belong to a known category set, then based on the original signal segment corresponding to the input tensor, a reference physical proxy vector corresponding one-to-one with the components of the predicted physical proxy vector is calculated. Calculate the consistency deviation between the predicted physical proxy vector and the reference physical proxy vector, and perform anomaly detection on the target signal based on a preset consistency threshold; Based on the open set rejection determination result and the anomaly discrimination result, the final identification result or regulatory alarm information is output.

[0007] According to the UAV radio open set rejection and anomaly alarm method provided by the present invention, the dual-representation branch decoupled deep learning network includes: A shared feature extraction unit is used to receive the input tensor and extract basic time-frequency texture features, local spectral structure features and propagation perturbation-related features to form a shared feature representation. The identity discrimination representation branch is used to receive the shared feature representation, extract discrimination features related to UAV identity recognition, category boundary distinction and open set rejection, and output the discrimination features to the classification head; the classification head performs mapping processing on the received discrimination features to obtain the discrimination information; the discrimination information includes a known category score vector for open set rejection decision and a category probability vector for known category recognition; The physical proxy representation branch is used to extract statistical physical features related to received intensity changes, overall spectrum shift, and spectrum spread from the shared feature representation, and output the statistical physical features to the physical proxy prediction head; the physical proxy prediction head maps the received statistical physical features and outputs the predicted physical proxy vector.

[0008] According to the UAV radio open set rejection and anomaly alarm method provided by the present invention, the reference physical proxy vector includes the energy change rate, spectral centroid and spectral bandwidth calculated from the original signal segment corresponding to the input tensor, and the predicted physical proxy vector includes predicted values ​​that correspond one-to-one with the energy change rate, spectral centroid and spectral bandwidth.

[0009] According to the UAV radio open-set rejection and anomaly alarm method provided by the present invention, the training method of the dual-representation branch decoupled deep learning network includes: acquiring radio signal samples of known UAV categories and corresponding category labels; preprocessing each radio signal sample to obtain a training input tensor, and simultaneously calculating the reference physical proxy vector corresponding to each radio signal sample; constructing a classification supervision loss based on the difference between the known category prediction result obtained by forward inference and the category label; constructing a physical consistency loss based on the difference between the predicted physical proxy vector obtained by forward inference and the calculated reference physical proxy vector; and iteratively optimizing the network parameters using the weighted sum of the classification supervision loss and the physical consistency loss as the total loss.

[0010] According to the UAV radio open-set rejection and anomaly alarm method provided by the present invention, the collected UAV radio signals are preprocessed to form a standardized input tensor. Specifically, the method includes: digitizing the collected UAV radio signals to form a complex baseband sampling sequence; performing sliding slicing on the complex baseband sampling sequence to obtain multiple signal segments of equal length; performing time-frequency transformation on each signal segment to generate a two-dimensional time-frequency power map; and performing amplitude normalization and size alignment on the time-frequency power map to form the input tensor of fixed size.

[0011] According to the UAV radio open set rejection and anomaly alarm method provided by the present invention, the open set discrimination score is calculated based on the discrimination information, specifically by: obtaining the known category score vector contained in the discrimination information; and obtaining the open set discrimination score by performing logarithmic summation and exponential operation on the known category score vector.

[0012] According to the UAV radio open set rejection and anomaly alarm method provided by the present invention, during the network training phase, the input tensor used for training is perturbed while keeping its corresponding category label and the reference physical proxy vector unchanged, so as to improve the robustness of the network in complex propagation environments.

[0013] Secondly, the present invention also provides a UAV radio open-collection rejection and anomaly alarm system, comprising: The signal preprocessing module is configured to preprocess the acquired UAV radio signals to form a standardized input tensor. The network inference module is configured to input the input tensor into a trained dual-representation branch decoupled deep learning network, and output discriminant information for known category identification, and a predicted physical surrogate vector for describing the physical statistical characteristics of the signal. The open set rejection determination module is configured to calculate the open set discrimination score based on the discrimination information, and determine whether the target belongs to the known category set according to the preset open set threshold. The unknown alarm module is configured to output a rejection result and trigger a regulatory alarm if the target is determined to be of an unknown type. The physical quantity calculation module is configured to, if the target is determined to belong to a known category set, calculate a reference physical proxy vector that corresponds one-to-one with the components of the predicted physical proxy vector based on the original signal segment corresponding to the input tensor. The consistency anomaly detection module is configured to calculate the consistency deviation between the predicted physical proxy vector and the reference physical proxy vector, and to perform anomaly detection on the target signal according to a preset consistency threshold. The result output module is configured to output the final identification result or regulatory alarm information based on the open set rejection determination result and the anomaly discrimination result.

[0014] Thirdly, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of any of the above-described UAV radio open set rejection and anomaly alarm methods.

[0015] Fourthly, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the UAV radio open set rejection and anomaly alarm method as described above.

[0016] The UAV radio open-set rejection and anomaly alarm method and system provided by this invention have the following advantages compared with the prior art: (1) This invention constructs a "dual-representation branch decoupled deep learning network," which internally separates the representation learning of identity discrimination information and physical statistical information. Simultaneously, a "physical proxy consistency constraint mechanism" is introduced, which forces the network to maintain consistency between the predicted physical proxy quantities (such as energy change rate and spectral centroid) and the reference quantities directly calculated from the signal during the training phase through a physical consistency loss function. This design allows the identity discrimination representation branch to focus more on stable identity features when dealing with signal feature drift caused by factors such as urban multipath, occlusion, and signal-to-noise ratio fluctuations, while the physical proxy representation branch learns statistical laws independent of environmental changes. The noise and frequency offset perturbations applied during the training phase further enhance the model's generalization ability. Therefore, the system can maintain more stable recognition performance and more consistent physical statistical predictions under complex conditions, effectively reducing misjudgments caused by environmental interference.

[0017] (2) This invention explicitly introduces an "open set rejection decision mechanism." Unlike traditional closed-category recognition systems that force any input into a known category, this method constructs an "open set discrimination score" based on the known category score vector output by the identity discrimination representation branch and sets a threshold. When this score indicates that the input signal does not belong to the feature distribution range of any known category, the system determines it as an "unknown type" and outputs a rejection warning. This mechanism directly solves the inherent problem of "forcibly classifying unknown targets into known categories" pointed out in the background technology, enabling the monitoring system to effectively discover and warn of new unmanned aerial vehicle targets that are not registered or not covered in the model training, thus improving the completeness of the monitoring.

[0018] (3) This invention creatively uses "physical proxy consistency" as one of the core criteria for anomaly detection. During online inference, the system calculates the physical proxy quantity predicted by the network and the reference physical proxy quantity directly calculated from the original signal segment in parallel, and calculates the consistency deviation between the two. When the target is identified as a known type, but this deviation exceeds the threshold, it is judged as "consistency anomaly" and an alarm is triggered. This provides a verification standard that is objectively measurable and traceable and does not rely on the black-box output inside the model. This mechanism can effectively identify adversarial attacks such as signal replay and modulation spoofing that attempt to deceive the identity recognition model, because such deceptive signals often have difficulty maintaining consistency with the real signal in both identity features and real physical statistical features (such as signal energy change trends and spectrum morphology), thus exposing anomalies and greatly improving the system's defense capability against deceptive behavior and the credibility of alarm results. Attached Figure Description

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

[0020] Figure 1 This is a flowchart illustrating the UAV radio open-set rejection and anomaly alarm method provided by the present invention. Figure 2 This is a schematic diagram of the structure of the UAV radio open set rejection and anomaly alarm system provided by the present invention; Figure 3 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0022] It should be noted that, in the description of the embodiments of the present invention, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. Those skilled in the art can understand the specific meaning of the above terms in the present invention according to the specific circumstances.

[0023] The terms "first," "second," etc., used in this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more.

[0024] The following is combined with Figures 1-3 This invention describes the method and system for unmanned aerial vehicle (UAV) radio open set rejection and anomaly alarm provided in embodiments of the present invention.

[0025] Figure 1 This is a flowchart illustrating the UAV radio open-set rejection and anomaly alarm method provided by the present invention, as shown below. Figure 1 As shown, including but not limited to the following steps: Step 101: Preprocess the collected UAV radio signals to form a standardized input tensor.

[0026] Optionally, step 101 of the present invention can be implemented as follows: (1) The collected UAV radio signals are digitally processed to form a complex baseband sampling sequence: This invention continuously collects and digitizes UAV radio signals within the monitored airspace using a radio receiving device, forming a complex baseband sampling sequence:

[0027] in, Indicates the first The complex baseband values ​​of each sampling point, with the real part and imaginary part corresponding to the in-phase component and quadrature component, respectively; Indicates the total number of sampling points; Represents the complex field. The sampling rate is denoted as... (Unit: Hz), used to determine the time resolution and the frequency scale for subsequent time-frequency analysis.

[0028] (2) The complex baseband sampling sequence is slicing to obtain multiple signal segments of equal length:

[0029] in, Indicates the first One sample segment; For fragment indexing; The window length (unit: sampling points) determines the duration of coverage for a single inference iteration. (Unit: seconds); The sliding step size (unit: sampling points) determines the degree of overlap between adjacent segments and the update frequency. (Unit: seconds). This slicing method converts continuous signals into equal-length sample sequences that can be processed online, meeting real-time monitoring requirements.

[0030] (3) Perform time-frequency transformation on each of the signal segments to generate a two-dimensional time-frequency power map; normalize the amplitude and align the size of the time-frequency power map to form the input tensor of a fixed size.

[0031] This invention performs time-frequency transformation on each signal segment and constructs a two-dimensional time-frequency power map as the network input representation, denoted as . This invention relates to... Amplitude normalization and size alignment are performed to obtain a fixed-size input tensor, ensuring that the input scale is consistent under different acquisition conditions and different signal-to-noise ratios and meeting the requirements of end-to-end network inference.

[0032] Step 102: Input the input tensor into the trained dual-representation branch decoupled deep learning network, and output the discriminant information for known category identification and the predicted physical surrogate vector for describing the physical statistical characteristics of the signal.

[0033] (1) Decoupling of shared feature extraction and dual representation branch construction This invention uses the fixed-size input tensor obtained in step 101 as the input to the deep learning network, and the input tensor is denoted as... .in, Indicates the first A sliding segment index, This represents the network input after normalization and size alignment of the two-dimensional time-frequency power map corresponding to the segment. To address the problem that existing single shared representations, when simultaneously undertaking category discrimination and physical agent prediction tasks, easily lead to the coupling and interference between identity discrimination information and statistical physical information, this invention constructs a dual-representation branch decoupled deep learning network.

[0034] The network first uses a shared feature extraction unit to extract basic features from the input time-frequency map to obtain low-level features (shared feature representation) containing modulation texture, local spectral structure and propagation perturbation information. Then, the network further performs dual representation branch decoupling on the basis of the low-level features, dividing the subsequent representation process into an identity discrimination representation branch for identity discrimination and a physical agent representation branch for physical agent prediction.

[0035] This invention constructs a feature extraction backbone network and outputs a shared fingerprint representation:

[0036] in, Represents the input tensor The fingerprint representation vector obtained by the shared feature extraction unit serves as the shared fingerprint representation, which includes the modulation texture features, spectral structure features, and propagation disturbance related features of the UAV signal. The dimension of the fingerprint representation vector. express A real vector space.

[0037] The shared feature extraction unit is implemented using a convolutional neural network structure and is used to process the input tensor. Multi-layer feature extraction is performed. Specifically, the shared feature extraction unit includes several convolutional layers, nonlinear activation layers, and downsampling layers. By performing layer-by-layer convolution operations on the input two-dimensional time-frequency power map, local time-frequency texture features and spectral structure features of the signal are extracted. The convolutional layers can use two-dimensional convolution operations, the nonlinear activation layers can use the ReLU activation function, and the downsampling layers can be implemented using max pooling or stride convolution. Further, after the convolutional feature extraction is completed, the high-dimensional feature map is compressed into a fixed-dimensional feature vector through global average pooling or fully connected mapping, thereby forming a shared fingerprint representation. The identity discrimination representation branch is based on shared fingerprint representation. The structure is constructed, specifically implemented as a set of fully connected network structures. The identity discrimination representation branch is connected through at least one fully connected layer. A mapping transformation is performed, and a nonlinear activation function is used to extract feature representations related to category discrimination. Subsequently, the classification head uses a linear mapping layer to further map these feature representations, outputting a category score vector. And calculate the class probability vector using the softmax function. This enables the output of discriminant information required for known category recognition and open set rejection discrimination.

[0038] The physical proxy representation branch is also based on shared fingerprint representation. The structure is constructed, specifically implemented as a set of regression network structures. The physical agent representation branch is connected to at least one fully connected layer. A mapping transformation is performed, and feature representations related to the statistical physical properties of the signal are extracted. The physical surrogate prediction head uses a linear mapping layer to map the feature representations into a low-dimensional vector output, thereby obtaining the physical surrogate quantity prediction vector. This includes predicted values ​​for the rate of energy change, the spectral centroid, and the spectral bandwidth.

[0039] Through the above structure As a shared representation, it realizes unified feature encoding of the input signal, while identity discrimination representation and physical agent representation are in Based on this, separate modeling is performed using a branch decoupling approach, thereby reducing mutual interference between the category discrimination task and the physical consistency modeling, and improving the overall discrimination stability and statistical consistency representation ability of the network.

[0040] (2) Identity discrimination representation branch construction and known category prediction In the dual-representation branch decoupling network, this invention sets up an identity discrimination representation branch for using shared fingerprint representations. The discriminative features related to the drone's identity category are extracted and used to perform known category prediction. The shared fingerprint representation... The unified feature representation obtained by the feature extraction backbone network of the input signal includes the modulation texture features, spectral structure features, and propagation-related features of the UAV signal; the identity discrimination representation branch is in Based on this, further discriminative features are extracted to enhance the separability between different UAV categories. These discriminative features are automatically extracted through supervised learning during the training phase. Specifically, the system uses known category labels to supervise the output of the classification head, minimizing the classification loss to ensure the identity discrimination branch operates within the shared fingerprint representation. Based on this, automatic learning of feature representations is beneficial for distinguishing different drone categories. The identity discrimination representation branch outputs discriminative features:

[0041] in, This represents the identity discrimination representation branch mapping. Indicates shared fingerprint representation The extracted discriminative features.

[0042] This invention constructs a classification head to map the discriminative features, obtaining a category score vector and a category probability vector:

[0043] in, Represents the category header mapping, For category score vectors, The dimension is equal to the number of known categories. , its first Each component Indicates that the input belongs to the first... Unnormalized scores for each known category.

[0044]

[0045] in, Represents the normalization function. For class probability vectors, and , This indicates that the input was judged as the first. The probability of a class is used to output the known class identification result and subsequent open set rejection decision.

[0046] In this invention, the classification head, as the task output of the identity discrimination representation branch, primarily serves known category identification and open set rejection decision. The identity discrimination representation branch, through shared fingerprint representation... Further mapping processing is performed to extract discriminative features related to the drone's identity. Specifically, for differences in launch hardware, the network extracts steady-state spectral morphology and frequency shift features from the time-frequency power map through convolutional operations; for differences in modulation methods, the network models the structural pattern differences of the signal in the time-frequency plane through multi-layer convolution; for differences in signal time-frequency texture, the network extracts local texture features through multi-scale convolution and nonlinear mapping. Based on this, the identity discrimination representation branch forms discriminative features with class distinguishing capabilities through fully connected mapping, and inputs them into the classification head for class score calculation, thereby improving the stability of known class recognition and enhancing the network's ability to distinguish the boundaries of unknown targets.

[0047] (3) Physical surrogate representation branch construction and physical surrogate prediction In the dual-representation branch decoupling network, this invention further sets up a physical proxy representation branch for using shared fingerprint representations. Statistical physical features related to the statistical physical properties of the signal are extracted and used to perform physical proxy quantity prediction. The shared fingerprint representation... The unified feature representation obtained from the input signal through the feature extraction backbone network includes modulation texture features, spectral structure features, and propagation perturbation-related features; the physical proxy representation branch is in Based on this, features related to the statistical regularity of the signal are further extracted to characterize the overall variation characteristics of the signal in the time-frequency domain. Specifically, for the statistical regularities corresponding to the energy change rate, spectral centroid, and spectral bandwidth, the physical proxy representation branch uses convolutional mapping and fully connected mapping to represent the shared fingerprint. The transformation yields a statistical physical feature representation; this statistical physical feature representation is then input into the physical surrogate prediction head, outputting a predicted physical surrogate quantity. During the training phase, the energy change rate, spectral centroid, and spectral bandwidth directly calculated from the original time-frequency power spectrum are used as reference physical surrogate quantities. The difference between the predicted physical surrogate quantity and the reference physical surrogate quantity is constrained by physical consistency loss, thus making the parameter update process of the physical surrogate representation branch constrained by the reference physical surrogate quantity. The physical surrogate representation branch outputs statistical physical features:

[0048] in, This represents the physical agent representation branch mapping. Indicates shared fingerprint representation The extracted statistical physical features.

[0049] The system constructs a physical proxy prediction head to map the statistical physical features and outputs a physical proxy quantity prediction vector:

[0050] in, This indicates the physical agent prediction head mapping. For physical agent prediction vectors; The predicted value representing the rate of change of energy. The predicted value representing the centroid of the spectrum. This represents the predicted value of the spectral bandwidth. This prediction header influences the network's output class classification results. Simultaneously, it further outputs a physical proxy quantity that can be directly calculated from the receiving signal end, thus enabling the network output to simultaneously possess both category discrimination results and physical statistical prediction results. The physical proxy quantity can be obtained by calculating the original time-frequency power spectrum and can be compared with the predicted value for consistency. It is used to construct physical consistency constraints during the training phase and to perform anomaly detection and open set recognition during the inference phase.

[0051] (4) Formation of forward inference output of dual-representation branch decoupling network This invention generates four types of core outputs after completing one forward propagation of a single input segment. ,in, and Used for known category identification and open set rejection. This is used to calculate the consistency deviation with the calculated physical proxy quantity (referencing the physical proxy quantity vector). This invention employs a dual-representation branch decoupling method, mapping the "identity discrimination task" and the "physical proxy task" to different representation branches within the same network framework. This allows the network's forward inference process to simultaneously possess two distinct yet complementary output capabilities: one type of output reflects evidence that the target belongs to a known category set, and the other type reflects the consistency state between the target signal and statistical physical laws. Through this structure, the network can more stably handle the three objectives of category recognition, unknown rejection, and anomaly verification in complex low-altitude propagation environments, thus forming a dedicated deep learning network architecture for low-altitude surveillance scenarios.

[0052] According to the aforementioned scheme, this embodiment of the invention provides a specific configuration for a dual-representation branch decoupled deep learning network: The input tensor first enters the shared feature extraction unit; the shared feature extraction unit adopts a 7-layer convolutional neural network. The first layer has 64 3×3 convolutional kernels with a stride of 1, the second layer has 64 3×3 convolutional kernels with a stride of 1 followed by 2×2 max pooling, the third layer has 128 3×3 convolutional kernels with a stride of 1, the fourth layer has 128 3×3 convolutional kernels with a stride of 1 followed by 2×2 max pooling, the fifth layer has 256 3×3 convolutional kernels with a stride of 1, the sixth layer has 256 3×3 convolutional kernels with a stride of 1, and the seventh layer has 256 3×3 convolutional kernels with a stride of 1 followed by global average pooling. All convolutional layers use the ReLU activation function and are supplemented with batch normalization, finally outputting a shared fingerprint representation vector with dimension d=256.

[0053] The shared fingerprint representation vector is input in parallel to the identity discrimination representation branch and the physical surrogate representation branch. The identity discrimination representation branch receives the shared fingerprint representation vector and consists of two fully connected layers: the first layer has 256 neurons followed by a ReLU activation function, and the second layer has 128 neurons followed by a ReLU activation function. These layers are used to extract discriminative features and output them to the classification head. The classification head is a linear mapping layer that outputs a class score vector, which is then processed by a Softmax function to obtain a class probability vector. This enables the identity discrimination representation branch to output prediction results for known classes. The physical surrogate representation branch also receives the shared fingerprint representation vector and consists of two fully connected layers: the first layer has 256 neurons followed by a ReLU activation function, and the second layer has 128 neurons followed by a ReLU activation function. These layers are used to extract statistical physical features and output them to the physical surrogate prediction head. The physical surrogate prediction head is a linear mapping layer that outputs a 3-dimensional predicted physical surrogate quantity vector, corresponding to the predicted values ​​of energy change rate, spectral centroid, and spectral bandwidth, respectively. This enables the physical surrogate representation branch to output the predicted results of signal physical statistics.

[0054] Step 103: Based on the discrimination information, calculate the open set discrimination score, and determine whether the target belongs to the known category set according to the preset open set threshold.

[0055] Specifically, the open set discrimination score is calculated based on the discrimination information by: obtaining the known category score vector contained in the discrimination information; and obtaining the open set discrimination score by performing logarithmic summation and exponential operation on the known category score vector.

[0056] Specifically, this invention reads known category score vectors. And construct an open-set discriminant score to characterize the "strength of evidence belonging to a known category set". :

[0057] in, The number of known categories; For the first Class score; This is a scalar score for open-set discrimination. Using this score, the invention aggregates multi-category outputs into a single piece of evidence for subsequent threshold discrimination.

[0058] Optionally, the present invention is based on a first preset false alarm rate. Determine the open set threshold on the validation data. .in, The threshold represents the upper probability limit of "a known category sample being misclassified as unknown"; By analyzing known samples in the validation set The statistical distribution is determined through quantile estimation or equivalent statistical procedures, such that the known sample satisfies... .

[0059] Step 104: If the target is determined to be of an unknown type, output the rejection result and trigger a regulatory alarm.

[0060] This invention will With threshold Comparison: When When the target is identified as an unknown type, the system will output a rejection result.

[0061] Step 105: If the target is determined to belong to a known category set, then based on the original signal segment corresponding to the input tensor, calculate the reference physical proxy vector that corresponds one-to-one with the components of the predicted physical proxy vector.

[0062] when When this occurs, the present invention determines that the current target belongs to a known category set and performs consistency anomaly detection.

[0063] Here, the reference physical proxy vector The calculation is briefly explained, and it is compared with the predicted physical agent vector output by the network. The components correspond one-to-one in meaning, thus providing a unified basis for subsequent training consistency constraints and inference consistency anomaly detection.

[0064] (1) Calculate the time-frequency power spectrum This invention relates to the first The complex time-frequency spectrum of a signal segment is power-processed to convert it into a non-negative power intensity. This gives statistical operations such as energy accumulation, frequency-weighted averaging, and second-order moments clear physical meaning and numerical stability. The output time-frequency power spectrum is as follows:

[0065] in, Spectral value when it is a complex number; This represents the power intensity at the corresponding time and frequency point; For frequency point index or frequency value, This serves as a time frame index; the power spectrum acts as a unified input for three types of reference quantities: energy, centroid, and bandwidth.

[0066] (2) Calculate the reference quantity of energy change rate

[0067] This embodiment is used to construct a reference quantity that reflects the trend of changes in the received signal strength of adjacent segments. When the distance, obstruction, transmission power, or link status of the UAV changes, the overall strength of the received signal will exhibit an observable change pattern; the energy change rate reference quantity describes this change in a relative form, making it comparable at different absolute intensity scales.

[0068] This invention output and the network output Corresponding. This invention calculates the first... Total energy of the segments: ; in, The total energy scalar of the fragment; The number of time frames within the segment; Let be the set of frequency points; this formula represents the accumulation of power over the entire time frequency range, used to characterize the overall received strength level of this segment. This invention calculates the reference quantity for the normalized rate of change of energy: ; in, This represents the total energy of the previous segment; To maintain stability and avoid an excessively small denominator, this reference value describes the relative change in received intensity between adjacent segments and is correlated with the predicted value. One-to-one correspondence.

[0069] (3) Calculate the reference value of the spectral centroid. With spectral bandwidth reference

[0070] This invention calculates the spectral centroid for each time frame:

[0071] in, Represents time frame The center frequency position; the numerator is the frequency-weighted power sum, and the denominator is the power sum; this quantity is used to characterize the overall shift trend of energy on the frequency axis. The system defines the segment-level centroid reference quantity as the average of the time frame centroids: in, The centroid reference value for the segment spectrum and the prediction value correspond.

[0072] This invention calculates the spectral bandwidth for each time frame:

[0073] in, Represents time frame The degree of spectral spread; this quantity is used to characterize the "expansion / contraction" of energy distribution along the frequency axis. The system defines the segment-level bandwidth reference as the average of the time frame bandwidth: in, The reference value for the segment's spectral bandwidth, and the predicted value. correspond.

[0074] This invention ultimately forms a reference physical proxy vector:

[0075] in, The reference vector is directly calculated from the signal, and They correspond in terms of components; subsequently, consistency constraints or consistency anomaly detection can be performed based on the differences between the two.

[0076] Step 106: Calculate the consistency deviation between the predicted physical proxy vector and the reference physical proxy vector, and perform anomaly detection on the target signal according to the preset consistency threshold.

[0077] When the target is determined to be in a known category set, the present invention further reads the predicted physical agent vector. And read the reference physical agent vector calculated from the signal. This invention constructs a consistency deviation. To characterize the difference between the predicted and reference values:

[0078] in, Represents the L2 norm; This is a consistency deviation scalar; a larger value indicates a greater inconsistency between the model output and the signal statistical reference. This invention is based on a second preset false alarm rate. Determine the consistency threshold on normal validation data .in, The threshold represents the upper probability limit of "a known normal sample being misclassified as an anomaly"; By analyzing normal samples in the validation set The statistical distribution is determined through quantile estimation or equivalent statistical procedures to ensure that the normal sample meets the following conditions. This invention will With threshold Comparison: When When the system determines that the target has a consistency anomaly, it outputs an anomaly alarm message; when When the target is deemed normal, the system outputs the known category recognition result.

[0079] Step 107: Based on the open set rejection determination result and the anomaly discrimination result, output the final identification result or regulatory alarm information.

[0080] When the target is determined to be of an unknown type, the present invention outputs an unknown rejection result and marks the target as an unknown target for subsequent regulatory handling.

[0081] When the target is determined to belong to a known category set and the anomaly detection result is normal, the present invention outputs the category identification result, wherein the identification category can be the category index corresponding to the category with the highest output category probability, and the corresponding confidence information is output.

[0082] When the target is determined to belong to a known category set and the anomaly detection result is anomaly, this invention outputs an anomaly alarm message. The alarm message includes at least a timestamp, target identifier, and detection score. Consistency deviation Threshold and This is so that the low-altitude monitoring system can record alarms, issue risk warnings, or take further action.

[0083] Based on the above embodiments, as an optional embodiment, the UAV radio open-set rejection and anomaly alarm method provided by the present invention includes the following training method for the dual-representation branch decoupled deep learning network: acquiring radio signal samples of known UAV categories and corresponding category labels; preprocessing each radio signal sample to obtain a training input tensor, and simultaneously calculating a reference physical proxy vector corresponding to each radio signal sample; constructing a classification supervision loss based on the difference between the known category prediction result obtained by forward inference and the category label; constructing a physical consistency loss based on the difference between the predicted physical proxy vector obtained by forward inference and the calculated reference physical proxy vector; and iteratively optimizing the network parameters using the weighted sum of the classification supervision loss and the physical consistency loss as the total loss.

[0084] During the training phase, this invention updates network parameters based on supervised labels and physical proxy quantities. The known definition of class-supervised loss is:

[0085] in To classify and monitor losses; The number of known categories; This is the category indicator variable for the sample, with the component corresponding to the true category taking the value 1; The loss is the class probability output in step three; this loss enables the network to form a discriminative boundary for known classes.

[0086] Physical consistency loss is defined as:

[0087] in This represents a loss of physical consistency. This is the physical agent prediction vector output from step three; This refers to the physical proxy vector obtained from the signal calculation in step four. It is a 2-norm; this loss makes the fingerprint representation consistent with the energy change and spectral statistics in a statistical sense, thereby improving robustness and verifiability.

[0088] The total loss function is defined as:

[0089] in For the overall training objective; is a weighting coefficient used to balance the classification ability and the physical consistency constraint strength; this definition ensures that the network learns both "classification" and "physical consistency" simultaneously.

[0090] In this invention, the above training process corresponds to a dual-representation branch decoupled network structure: wherein, the classification supervision loss... Primarily applied to the identity discrimination representation branch, it enhances the distinguishing ability between known categories and stabilizes the category boundaries upon which open-set rejection depends; physical consistency loss. Primarily applied to the physical agent representation branch, this invention strengthens the correspondence between network output and the statistical physical properties of the signal. Since the two types of losses correspond to two different representation objectives, this invention further enhances the representation separation effect within the network through objective decoupling during the training phase. This allows identity discrimination representation to focus more on class separability, and physical agent representation to focus more on statistical consistency, thereby reducing the mutual constraints between the two tasks within the same feature space.

[0091] Furthermore, to improve the model's generalization ability to changes in propagation conditions, a perturbation is applied to the input tensor during the training phase while keeping the class labels unchanged:

[0092] in, The input after perturbation; For perturbation operators; Indicates the noise intensity disturbance parameter; This represents the frequency offset perturbation parameter; this mechanism is used to cover conditions with increased noise and frequency offset, keeping the network output stable. The perturbation sample and the original sample both participate in the above process. The training process improves the robustness of the model.

[0093] Furthermore, since the present invention adopts a dual-representation branch decoupling method to separate the identity discrimination task and the physical agent prediction task into separate models, under the condition of input perturbation, the identity discrimination representation branch can still maintain a relatively stable category discrimination output, and the physical agent representation branch can still maintain a prediction output that matches the signal statistical properties, thereby improving the overall robustness of the entire system in complex propagation environment, multipath interference environment and frequency offset perturbation environment.

[0094] Figure 2 This is a schematic diagram of the structure of the UAV radio open-set rejection and anomaly alarm system provided by the present invention, as shown below. Figure 2 As shown, it includes: a signal preprocessing module 210, a network inference module 220, an open set rejection determination module 230, an unknown alarm module 240, a physical quantity calculation module 250, a consistency anomaly discrimination module 260, and a result output module 270.

[0095] The signal preprocessing module 210 is configured to preprocess the acquired UAV radio signals to form a standardized input tensor; The network inference module 220 is configured to input the input tensor into a trained dual-representation branch decoupled deep learning network, and output discriminant information for known category recognition, and a predicted physical surrogate vector for describing the physical statistical characteristics of the signal. The open set rejection determination module 230 is configured to calculate the open set discrimination score based on the discrimination information, and determine whether the target belongs to the known category set according to the preset open set threshold. The unknown alarm module 240 is configured to output a rejection result and trigger a regulatory alarm if the target is determined to be of an unknown type. The physical quantity calculation module 250 is configured to calculate a reference physical proxy vector that corresponds one-to-one with the components of the predicted physical proxy vector based on the original signal segment corresponding to the input tensor if the target is determined to belong to a known category set. The consistency anomaly detection module 260 is configured to calculate the consistency deviation between the predicted physical proxy vector and the reference physical proxy vector, and to perform anomaly detection on the target signal according to a preset consistency threshold. The result output module 270 is configured to output the final identification result or regulatory alarm information based on the open set rejection determination result and the anomaly discrimination result.

[0096] It should be noted that the UAV radio open set rejection and anomaly alarm system provided in this embodiment of the invention can execute the UAV radio open set rejection and anomaly alarm method described in any of the above embodiments during actual operation, which will not be elaborated in this embodiment.

[0097] In summary, the UAV radio open-set rejection and anomaly alarm method and system provided by this invention have the following advantages compared with the prior art: (1) This invention constructs a "dual-representation branch decoupled deep learning network," which internally separates the representation learning of identity discrimination information and physical statistical information. Simultaneously, a "physical proxy consistency constraint mechanism" is introduced, which forces the network to maintain consistency between the predicted physical proxy quantities (such as energy change rate and spectral centroid) and the reference quantities directly calculated from the signal during the training phase through a physical consistency loss function. This design allows the identity discrimination representation branch to focus more on stable identity features when dealing with signal feature drift caused by factors such as urban multipath, occlusion, and signal-to-noise ratio fluctuations, while the physical proxy representation branch learns statistical laws independent of environmental changes. The noise and frequency offset perturbations applied during the training phase further enhance the model's generalization ability. Therefore, the system can maintain more stable recognition performance and more consistent physical statistical predictions under complex conditions, effectively reducing misjudgments caused by environmental interference.

[0098] (2) This invention explicitly introduces an "open set rejection decision mechanism." Unlike traditional closed-category recognition systems that force any input into a known category, this method constructs an "open set discrimination score" based on the known category score vector output by the identity discrimination representation branch and sets a threshold. When this score indicates that the input signal does not belong to the feature distribution range of any known category, the system determines it as an "unknown type" and outputs a rejection warning. This mechanism directly solves the inherent problem of "forcibly classifying unknown targets into known categories" pointed out in the background technology, enabling the monitoring system to effectively discover and warn of new unmanned aerial vehicle targets that are not registered or not covered in the model training, thus improving the completeness of the monitoring.

[0099] (3) This invention creatively uses "physical proxy consistency" as one of the core criteria for anomaly detection. During online inference, the system calculates the physical proxy quantity predicted by the network and the reference physical proxy quantity directly calculated from the original signal segment in parallel, and calculates the consistency deviation between the two. When the target is identified as a known type, but this deviation exceeds the threshold, it is judged as "consistency anomaly" and an alarm is triggered. This provides a verification standard that is objectively measurable and traceable and does not rely on the black-box output inside the model. This mechanism can effectively identify adversarial attacks such as signal replay and modulation spoofing that attempt to deceive the identity recognition model, because such deceptive signals often have difficulty maintaining consistency with the real signal in both identity features and real physical statistical features (such as signal energy change trends and spectrum morphology), thus exposing anomalies and greatly improving the system's defense capability against deceptive behavior and the credibility of alarm results.

[0100] Figure 3 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 3As shown, the electronic device may include a processor 310, a communications interface 320, a memory 330, and a communication bus 340. The processor 310, communications interface 320, and memory 330 communicate with each other via the communication bus 340. The processor 310 can call logical instructions from the memory 330 to execute a UAV radio open-set rejection and anomaly alarm method.

[0101] Furthermore, the logical instructions in the aforementioned memory 330 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0102] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the UAV radio open set rejection and anomaly alarm method provided in the above embodiments.

[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for radio open-set rejection and anomaly alarm for unmanned aerial vehicles (UAVs), characterized in that, include: The collected UAV radio signals are preprocessed to form a standardized input tensor; The input tensor is fed into the trained dual-representation branch decoupled deep learning network, which outputs discriminant information for known category identification and a predicted physical surrogate vector for describing the physical statistical characteristics of the signal. Based on the discrimination information, the open set discrimination score is calculated, and the target is determined to belong to the known category set according to the preset open set threshold. If the target is determined to be of an unknown type, output a rejection result and trigger a regulatory alert. If the target is determined to belong to a known category set, then based on the original signal segment corresponding to the input tensor, a reference physical proxy vector corresponding one-to-one with the components of the predicted physical proxy vector is calculated. Calculate the consistency deviation between the predicted physical proxy vector and the reference physical proxy vector, and perform anomaly detection on the target signal based on a preset consistency threshold; Based on the open set rejection determination result and the anomaly discrimination result, the final identification result or regulatory alarm information is output.

2. The method for unmanned aerial vehicle (UAV) radio open-set rejection and anomaly alarm according to claim 1, characterized in that, The dual-representation branch decoupled deep learning network includes: A shared feature extraction unit is used to receive the input tensor and extract basic time-frequency texture features, local spectral structure features and propagation perturbation-related features to form a shared feature representation. The identity discrimination representation branch is used to receive the shared feature representation, extract discrimination features related to UAV identity recognition, category boundary distinction and open set rejection, and output the discrimination features to the classification head; the classification head performs mapping processing on the received discrimination features to obtain the discrimination information; the discrimination information includes a known category score vector for open set rejection decision and a category probability vector for known category recognition; The physical proxy representation branch is used to extract statistical physical features related to received intensity changes, overall spectrum shift, and spectrum spread from the shared feature representation, and output the statistical physical features to the physical proxy prediction head; the physical proxy prediction head maps the received statistical physical features and outputs the predicted physical proxy vector.

3. The method for unmanned aerial vehicle (UAV) radio open-set rejection and anomaly alarm according to claim 1, characterized in that, The reference physical surrogate vector includes the energy change rate, spectral centroid, and spectral bandwidth calculated from the original signal segment corresponding to the input tensor, and the predicted physical surrogate vector contains predicted values ​​that correspond one-to-one with the energy change rate, spectral centroid, and spectral bandwidth.

4. The method for unmanned aerial vehicle (UAV) radio open-set rejection and anomaly alarm according to claim 1, characterized in that, The training method for the dual-representation branch decoupling deep learning network includes: Obtain radio signal samples and corresponding category labels for known drone categories; For each radio signal sample, preprocessing is performed to obtain the input tensor for training, and the reference physical proxy vector corresponding to each radio signal sample is calculated at the same time. The classification supervision loss is constructed based on the difference between the known category prediction results obtained from forward inference and the category labels; The physical consistency loss is constructed by the difference between the predicted physical agent vector obtained by forward inference and the calculated reference physical agent vector. The network parameters are iteratively optimized and trained using the weighted sum of the classification supervision loss and the physical consistency loss as the total loss.

5. The method for unmanned aerial vehicle (UAV) radio open-set rejection and anomaly alarm according to claim 1, characterized in that, The collected UAV radio signals are preprocessed to form a standardized input tensor, specifically including: The collected UAV radio signals are digitally processed to form a complex baseband sampling sequence; The complex baseband sampling sequence is sliced ​​by sliding to obtain multiple signal segments of equal length; Perform time-frequency transformation on each of the signal segments to generate a two-dimensional time-frequency power map; The time-frequency power graph is normalized in amplitude and aligned in size to form the input tensor of fixed size.

6. The method for unmanned aerial vehicle (UAV) radio open-set rejection and anomaly alarm according to claim 1, characterized in that, Based on the discrimination information, the open set discrimination score is calculated as follows: Obtain the known category score vector contained in the discrimination information; The open set discrimination score is obtained by performing logarithmic summation and exponential operation on the known category score vector.

7. The method for unmanned aerial vehicle (UAV) radio open-set rejection and anomaly alarm according to claim 4, characterized in that, During the network training phase, the input tensor used for training is perturbed while keeping its corresponding category label and the reference physical proxy vector unchanged, in order to improve the robustness of the network in complex propagation environments.

8. A UAV radio open-set rejection and anomaly alarm system, used to implement the UAV radio open-set rejection and anomaly alarm method according to any one of claims 1 to 7, characterized in that, include: The signal preprocessing module is configured to preprocess the acquired UAV radio signals to form a standardized input tensor. The network inference module is configured to input the input tensor into a trained dual-representation branch decoupled deep learning network, and output discriminant information for known category identification, and a predicted physical surrogate vector for describing the physical statistical characteristics of the signal. The open set rejection determination module is configured to calculate the open set discrimination score based on the discrimination information, and determine whether the target belongs to the known category set according to the preset open set threshold. The unknown alarm module is configured to output a rejection result and trigger a regulatory alarm if the target is determined to be of an unknown type. The physical quantity calculation module is configured to, if the target is determined to belong to a known category set, calculate a reference physical proxy vector that corresponds one-to-one with the components of the predicted physical proxy vector based on the original signal segment corresponding to the input tensor. The consistency anomaly detection module is configured to calculate the consistency deviation between the predicted physical proxy vector and the reference physical proxy vector, and to perform anomaly detection on the target signal according to a preset consistency threshold. The result output module is configured to output the final identification result or regulatory alarm information based on the open set rejection determination result and the anomaly discrimination result.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the UAV radio open set rejection and anomaly alarm method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the UAV radio open set rejection and anomaly alarm method as described in any one of claims 1 to 7.

Images