An unmanned aerial vehicle frequency system real-time identification method and system based on fusion of transfer learning

Abstract

This invention discloses a real-time identification method and system for UAV frequency systems based on transfer learning, relating to the field of deep learning technology. The method and system include the following steps: S1, acquiring node frequency observation data and performing normalization and spectral base leveling; S2, constructing a standard source domain feature library and performing consistency evaluation, outputting calibrated frequency feature frames; S3, screening candidate segments and performing candidate segment identification and evaluation to obtain valid UAV frequency activity segments; S4, constructing a set of observations from the same source, evaluating the reliability of fused positioning, and determining the current node joint positioning result. This invention effectively improves the accuracy and cross-node consistency of UAV frequency system identification, solving the problem of inconsistent frequency features and significantly reduced UAV positioning accuracy during multi-node fused positioning.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, specifically to a real-time identification method and system for UAV frequency systems that integrates transfer learning. Background Technology

[0002] With the expansion of UAV applications and the development of multi-node monitoring deployments, existing technologies for real-time identification of frequency signals from different airspace nodes rely on single-node observations. They lack cross-node feature unification and transfer capabilities, and their identification accuracy and real-time performance cannot meet the needs of full-domain monitoring. Existing systems have limited generalization capabilities for different UAV frequency systems in small-sample scenarios, making it difficult to accurately distinguish signals from different aircraft models. This leads to misjudgments or missed judgments, and they cannot support multi-node collaborative positioning and full-domain detection applications.

[0003] For example, the invention patent with announcement number CN120849923B discloses a method and system for UAV signal recognition, including: Step 1: Adaptive noise basis estimation; performing a short-time Fourier transform on the input UAV IQ signal (i.e., the UAV signal) to obtain a time-frequency matrix, generating a UAV time-frequency map, estimating the noise power spectrum, and calculating the power spectral density at each frequency point; Step 2: Time-frequency feature enhancement; using a 5th-order Butterworth bandpass filter to perform frequency domain filtering on the signal; using soft thresholding domain denoising of the soft short-time Fourier transform result; Step 3: Multi-feature fusion regression; calculating the energy ratio feature and spectral peak sharpness feature, and extracting deep learning features, performing feature fusion and regression; Step 4: Achieving UAV signal recognition through a trained UAV signal classification model. This invention's method has high signal-to-noise ratio estimation accuracy in complex environments and excellent UAV classification accuracy and generalization.

[0004] For example, the invention patent with announcement number CN114580476B discloses a method for constructing a model for identifying drone signals and a corresponding identification method and system, including: converting radio signals into time-frequency maps that include at least time, frequency, and power information; extracting features and inputting them into a convolutional neural network model, which is first trained using training data and then substituted with test data to identify the type of target drone; identifying connected components in the time-frequency maps and then judging the category similarity based on the time-frequency maps to determine the ID part of the radio signal; extracting features and inputting them into a machine learning classifier model, which is first trained using training data and then substituted with test data to identify the individual target drone.

[0005] In the existing technology, existing systems suffer from heterogeneous biases in frequency signal acquisition of the same UAV due to differences in geographical location, hardware parameters, and electromagnetic environment among monitoring nodes in the same airspace. Existing systems directly fuse the features of each node for positioning calculation without correcting for the biases, resulting in low feature matching degree during positioning calculation and inability to achieve accurate airspace positioning of UAVs.

[0006] Therefore, in order to address the above problems, there is an urgent need for a real-time identification method and system for UAV frequency systems that integrates transfer learning. Summary of the Invention

[0007] Technical problems to be solved

[0008] To address the shortcomings of existing technologies, this invention provides a real-time identification method and system for UAV frequency systems that incorporates transfer learning, solving the problem of inconsistent frequency features and significantly reduced UAV positioning accuracy during multi-node fusion positioning.

[0009] Technical solution

[0010] To achieve the above objectives, the present invention provides the following technical solution: a real-time identification method and system for UAV frequency systems based on fusion transfer learning, comprising: S1, acquiring node frequency observation data and performing time-axis relocation, denoising, normalization, and spectral base leveling to construct a node frequency observation dataset; S2, based on the node frequency observation dataset, selecting core monitoring nodes, constructing a standard source domain feature library, performing consistency evaluation, and outputting calibration frequency feature frames to form a calibration frequency feature set; S3, filtering candidate segments based on the calibration frequency feature set, and performing candidate segment identification and evaluation to obtain effective UAV frequency activity segments; S4, constructing a homogeneous observation set, evaluating the reliability of fusion positioning, determining the current node joint positioning result, and generating target behavior labels.

[0011] Further, the specific process of acquiring node frequency observation data and performing time axis relocation, denoising, normalization, and spectral base leveling to construct a node frequency observation dataset is as follows: Fixed monitoring nodes, vehicle-mounted monitoring nodes, and edge monitoring nodes deployed in different spatial locations are dynamically accessed throughout the entire process to acquire raw IQ sampling data, power spectrum data, time-frequency diagram data, center frequency estimation results, frequency offset observation results, spectral peak shape description information, and node local reception status information as node frequency observation data; The acquisition time, segment number, and analysis window of each monitoring node are relocated using a unified time reference via NTP and PTP dual protocols; Each monitoring node is quantized and encoded, and a node contextual status label is constructed and bound to the corresponding timestamp frequency observation segment for storage; Denoising is performed using wavelet transform and impulse suppression algorithms to filter out impulse interference and spurious components; Frequency offset caused by local oscillator error and link differences is identified and compensated through center frequency drift correlation; The node frequency observation dataset is constructed by normalizing the spectral amplitude and leveling the spectral base of the node frequency observation data.

[0012] Furthermore, based on the node frequency observation dataset, the specific process of selecting core monitoring nodes and constructing a standard source domain feature library is as follows: Based on the node frequency observation dataset, core monitoring nodes are selected by considering the node quality assessment method that takes into account the signal-to-noise ratio, frequency stability, and geographical coverage; for the UAV frequency samples collected by the core monitoring nodes, the center frequency statistical features, frequency offset evolution, spectral peak morphology, time-frequency texture, and energy distribution features are extracted to form a standard frequency feature expression and construct a standard source domain feature library; the standard frequency feature expression of the core monitoring nodes is used as the source domain input, and the frequency feature expression to be calibrated of the target monitoring node is used as the target domain input to construct a cross-node migration calibration network; the cross-node migration calibration network includes a source domain feature encoding branch, a target domain feature encoding branch, a shared feature mapping layer, and a distribution alignment constraint layer. The two branches respectively compress the input frequency features, the shared feature mapping layer unifies the embedding space, and the distribution alignment constraint layer constrains the consistency between the source domain feature distribution and the target domain feature distribution in the unified embedding space, outputting the calibration frequency feature expression of the target monitoring node.

[0013] Furthermore, the specific process for consistency assessment is as follows: The source domain intra-class scatter matrix and the target domain intra-class scatter matrix are obtained by performing outer product operations on the deviation vectors of samples of the same frequency category in the standard source domain feature library of the core monitoring node and the target monitoring node's frequency features to be calibrated relative to their respective category feature centers, and then summing the results. The inter-class scatter matrix is ​​obtained by performing outer product operations on the deviation vectors of each frequency category feature center relative to the global feature center of all samples, and then summing the results. The target domain feature covariance matrix is ​​obtained by performing outer product operations on the deviation vectors of the target monitoring node's frequency feature expression vector relative to the target domain feature mean vector, and then averaging the results. The source domain feature covariance matrix... The inverse square root matrix of the difference matrix is ​​obtained by decomposing the source domain eigenvalue covariance matrix, taking the reciprocal of the square root of each positive eigenvalue, and reconstructing it. The joint intra-class discrete matrix is ​​obtained by adding the intra-class scatter matrices of the source and target domains. The inverse of the inter-class scatter matrix is ​​obtained by adding the inter-class scatter matrix and the adjustment parameter multiplied by the identity matrix. The matrix trace is then used to calculate the ratio term. The inverse square root of the source domain covariance is multiplied by the target domain covariance, then multiplied by the inverse square root on the left and right, and added to the identity matrix to obtain the covariance structure alignment matrix. The natural logarithm of the determinant is then used to obtain the structural difference term. The ratio term is divided by the structural difference term to obtain the migration calibration consistency evaluation value.

[0014] Further, the specific process of outputting calibration frequency feature frames to form a calibration frequency feature set is as follows: real-time comparison of the transfer calibration consistency evaluation value and the transfer calibration threshold; when the transfer calibration consistency evaluation value is greater than the transfer calibration threshold, it is determined that there is a domain offset between the current target monitoring node frequency feature and the core monitoring node standard feature, and the transfer learning calibration modeling process is initiated; when the transfer calibration consistency evaluation value is less than or equal to the transfer calibration threshold, it is determined that the current target monitoring node frequency feature has met the cross-node consistency requirements; in the transfer learning calibration modeling process, the standard frequency feature expression of the core monitoring node is used as the source domain sample input, and the node frequency feature frame of the target monitoring node is used as the target domain sample input to construct a cross-node transfer learning model; a multi-layer fully connected neural network structure is used to perform nonlinear mapping on the input, outputting a unified dimension embedded feature vector, real-time correcting the node frequency feature frame of the target monitoring node, generating calibration frequency feature frames, and summarizing them to form a calibration frequency feature set.

[0015] Furthermore, the specific process of selecting candidate segments based on the calibration frequency feature set and evaluating candidate segments is as follows: extracting UAV frequency activity segments from the calibration frequency feature frame and selecting candidate segments through joint discrimination; performing hierarchical nonlinear mapping feature encoding on the candidate segments to obtain representation vectors; obtaining matching centers by averaging the representation vectors belonging to the same UAV system category; obtaining non-matching centers by averaging the representation vectors of UAV system categories other than the current matching category; the non-matching category number is the number remaining after removing the current matching category number from the sequential numbers of all UAV system categories; obtaining the matching category number by comparing the similarity ranking of the representation vectors with each center; calculating the current table... The matching similarity is obtained by dividing the inner product of the current representation vector and the matching center by the sum of the product of the L2 norms of the current representation vector and the matching center and the zero-prevention term. The matching enhancement term is obtained by applying the inverse hyperbolic tangent. The non-matching similarity is obtained by calculating the inner product of the current representation vector and each non-matching center by the sum of the product of the L2 norms of the current representation vector and each non-matching center and the zero-prevention term. The non-matching interference term is obtained by applying the inverse hyperbolic tangent and selecting the maximum value. The non-matching interference term is obtained by subtracting the non-matching interference term from the matching enhancement term. The temporal similarity is obtained by calculating the inner product of the current and previous representation vectors by the sum of the product of the L2 norms of the current and previous representation vectors and the zero-prevention term. The temporal similarity is obtained by multiplying it by the temporal coefficient and adding it to the discrimination term to obtain the candidate segment recognition evaluation value.

[0016] Furthermore, the specific process for obtaining valid UAV frequency activity segments is as follows: Real-time comparison of candidate segment identification evaluation values ​​and identification trigger thresholds; when the candidate segment identification evaluation value is greater than the identification trigger threshold, the current target segment to be identified is confirmed as a valid UAV frequency activity segment; when the candidate segment identification evaluation value is less than or equal to the identification trigger threshold, the current segment is determined to be a background interference segment and is not entered into the main identification link; the valid UAV frequency activity segments are input into a lightweight deep recognition network, the input representation vector is classified, and the UAV system category corresponding to the current target is output; a transfer learning strategy that freezes some network layers and combines learning rate adjustment is adopted to enable the recognition network to maintain generalization ability in small sample target scenarios; the credibility of the recognition results is estimated to form category labels, system labels, recognition confidence levels, and segment validity indicators.

[0017] Furthermore, the specific process for constructing a co-source observation set and evaluating the reliability of fused positioning is as follows: Cross-node observation segments belonging to the same UAV target are associated to form a co-source observation set; calibration frequency characteristics, frequency offset, arrival time, and node spatial distribution are jointly calculated to obtain the target position estimate; the number of monitoring nodes participating in the joint positioning calculation is counted to obtain the number of monitoring nodes; the monitoring nodes are numbered sequentially to obtain the monitoring node number; the local position constraint points of the node are inferred based on the node's observation direction, reception time, frequency boundary, and spatial coordinates; the calibration frequency feature vector is obtained by structured extraction from the calibration frequency feature frames corresponding to the monitoring nodes; the target node's position estimate result is divided by the square of the distance between the local position constraint points of each monitoring node by the square of the position parameter, and then 1 is added to obtain the result. The geometric consistency term is obtained by raising the geometric exponent to the power of the geometric exponent; the geometric deviation term is obtained by summing all the geometric consistency terms, dividing by the number of monitoring nodes, and taking the reciprocal; the frequency consistency term is obtained by squared difference of calibration frequency eigenvectors between two different monitoring nodes and dividing by the square of the frequency parameter; the cross-node consistency term is obtained by multiplying the number of monitoring nodes by the number of monitoring nodes minus 1 to obtain the total number of node pairs, summing all the frequency consistency terms, dividing by the total number of node pairs, and taking the negative exponent; the time deviation term is obtained by squared difference between measured and theoretical arrival time difference between two different monitoring nodes and dividing by the square of the time parameter, summing all the time deviation terms, dividing by the total number of node pairs plus 1, and taking the negative exponent of the time; the time difference term is obtained by multiplying the geometric deviation term, cross-node consistency term, and time difference term together.

[0018] Furthermore, the specific process for determining the joint positioning result of the current node and generating target behavior labels is as follows: real-time comparison of the fusion positioning confidence assessment value and the fusion positioning threshold; when the fusion positioning confidence assessment value is greater than the fusion positioning threshold, the joint positioning result is confirmed to be valid, and the target location result and trajectory are output; when the fusion positioning confidence assessment value is less than or equal to the fusion positioning threshold, it is determined that there is abnormal node interference, and the output is temporarily suspended, waiting for the time window to supplement the verification; real-time target trajectory updates and associations are performed on the valid positioning results to form a global motion trajectory, and the target recognition result, positioning result, trajectory and original fragment index are structurally bound to generate target behavior labels and mark the corresponding abnormal states. Abnormal states include node failure, multipath interference and clock out of sync.

[0019] Furthermore, a second aspect of the present invention provides a real-time UAV frequency system identification system based on fusion transfer learning, applied to a real-time UAV frequency system identification method based on fusion transfer learning, comprising: a spatial monitoring node frequency data acquisition module, used to acquire node frequency observation data and perform time axis relocation, denoising, normalization and spectral base leveling to construct a node frequency observation dataset; a cross-node transfer calibration module, used to select core monitoring nodes based on the node frequency observation dataset, construct a standard source domain feature library, perform consistency evaluation, output calibration frequency feature frames, and form a calibration frequency feature set; a real-time frequency system identification module, used to screen candidate segments based on the calibration frequency feature set, and perform candidate segment identification evaluation to obtain effective UAV frequency activity segments; and a node fusion positioning module, used to construct a set of co-source observations, evaluate the reliability of fusion positioning, determine the current node joint positioning result, and generate target behavior labels.

[0020] Beneficial effects

[0021] The present invention has the following beneficial effects:

[0022] (1) This invention calibrates the frequency features of target monitoring nodes through cross-node transfer learning, thereby achieving feature consistency among monitoring nodes in different airspaces and significantly improving the cross-node consistency and stability of UAV frequency system identification. It effectively extracts center frequency, frequency offset, spectral peak morphology, time-frequency texture, and energy distribution features, improving the accuracy of frequency activity segment identification.

[0023] (2) This invention achieves unified embedding space alignment of source and target domain features by constructing a dual-branch feature encoding network, a shared embedding mapping layer, and distribution difference constraints, thereby reducing the frequency feature deviation between the source and target domains and ensuring the effectiveness of the transfer calibration model in real-time small sample scenarios. It enables real-time discrimination of UAV frequency activity segments and improves the temporal consistency of recognition results within continuous time periods.

[0024] (3) This invention uses a multi-node fusion positioning module to perform joint position calculation on the calibrated frequency features, and combines three constraints—geometric consistency, cross-node frequency consistency, and propagation time difference compliance—to significantly improve the reliability of the joint positioning results and the stability of global tracking. This provides reliable data support for multi-node collaborative monitoring, trajectory tracking, and abnormal behavior determination.

[0025] (4) This invention reduces recognition errors in small-sample scenarios by jointly optimizing source domain features, target domain features, and inter-node deviations, thereby reducing confusion and misjudgment between different UAV frequency systems and improving the system's generalization ability in complex airspace environments. It effectively improves the efficiency and reliability of real-time identification and global detection of UAV frequency systems.

[0026] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0027] Figure 1 This is a flowchart of a real-time identification method for UAV frequency systems that incorporates transfer learning, according to the present invention.

[0028] Figure 2 This is an architecture diagram of a real-time identification system for UAV frequency systems that integrates transfer learning, according to the present invention.

[0029] Figure 3 This is a line graph showing the changes in the candidate fragment identification evaluation value of the present invention.

[0030] Figure 4 This is a comparison diagram of the time-domain waveforms of the four types of radio frequency signals of this invention;

[0031] Figure 5 This is a PCA feature distribution diagram of the four types of radio frequency signals in this invention;

[0032] Figure 6 This is a confusion matrix diagram for the four types of radio frequency signals in this invention. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] Please see Figures 1-6This invention provides a technical solution: a real-time identification method and system for UAV frequency systems based on transfer learning, comprising: S1, acquiring node frequency observation data and performing time axis relocation, denoising, normalization, and spectral base leveling to construct a node frequency observation dataset; S2, based on the node frequency observation dataset, selecting core monitoring nodes, constructing a standard source domain feature library, performing consistency evaluation, and outputting calibration frequency feature frames to form a calibration frequency feature set; S3, filtering candidate segments based on the calibration frequency feature set, and performing candidate segment identification and evaluation to obtain effective UAV frequency activity segments; S4, constructing a homogeneous observation set, evaluating the reliability of fused positioning, determining the current node joint positioning result, and generating target behavior labels.

[0035] Specifically, the process of acquiring node frequency observation data and performing time axis relocation, denoising, normalization, and spectral base leveling to construct a node frequency observation dataset is as follows:

[0036] The system enables dynamic access to fixed monitoring nodes, vehicle-mounted monitoring nodes, and edge monitoring nodes deployed in different airspace locations throughout the entire process. Data is collected in real time via Ethernet or dedicated data links. IQ sampling data is recorded in double-precision complex number format at the sampling rate, power spectrum data is calculated at a resolution of 1024 points per frame, and time-frequency graph data is generated at a fixed frame rate. The system acquires the original IQ sampling data, power spectrum data, time-frequency graph data, center frequency estimation results, frequency offset observation results, spectral peak shape description information, and node local reception status information as node frequency observation data.

[0037] By unifying the time base using both NTP and PTP protocols, NTP achieves millisecond-level coarse synchronization, while PTP achieves sub-microsecond-level fine synchronization. The acquisition timestamps, segment numbers, and analysis window start times are calibrated, and the acquisition times, segment numbers, and analysis windows of each monitoring node are repositioned along the timeline. Each monitoring node is quantized and encoded, generating a unique contextual status label based on node type, geographical location, and device ID. This node contextual status label is then bound and stored with the corresponding timestamp-based frequency observation segment. Denoising is achieved through wavelet transform and impulse suppression algorithms, using a 5-level decomposition based on the db4 wavelet basis, and soft thresholding. For high-frequency coefficients, pulse suppression employs amplitude threshold detection and replaces outliers with neighborhood medians to filter out pulse interference and spurious components. A sliding window of 20 consecutive spectral frames is used, with the window duration determined by the spectral frame sampling rate (e.g., 0.2 seconds for a frame rate of 100Hz). Linear fitting is performed on the estimated center frequency within the window to identify and compensate for system frequency offsets caused by local oscillator errors and link differences. Frequency offsets caused by local oscillator errors and link differences are identified and compensated through center frequency drift correlation. Node frequency observation datasets are constructed by normalizing the spectral amplitude and leveling the spectral base of the node frequency observation data.

[0038] In this implementation plan, high precision, consistency, and comparability of node frequency observation data are achieved by performing unified time reference alignment, noise reduction, frequency offset compensation, normalization, and spectral base leveling on the frequency observation data of multiple spatial monitoring nodes. This effectively improves the accuracy, stability, and real-time performance of multi-node UAV frequency system identification, provides a reliable data foundation for candidate segment selection, feature encoding, migration calibration, and fusion positioning, and enhances the robustness and applicability of the system under complex electromagnetic environments and multi-node deployment conditions.

[0039] Specifically, the process of selecting core monitoring nodes and constructing a standard source domain feature library based on the node frequency observation dataset is as follows:

[0040] Based on the node frequency observation dataset, core monitoring nodes were selected using a node quality assessment method that considers signal-to-noise ratio, frequency stability, and geographical coverage. Statistical analysis was performed on the center frequency estimation results, frequency offset observation results, power spectrum data, and peak shape description information output by each monitoring node to calculate the center frequency dispersion, mean absolute frequency offset, background spectral floor variance, and the number of consecutive occurrences of the target spectral peak. Monitoring nodes were sorted from smallest to largest according to center frequency dispersion, mean absolute frequency offset, and background spectral floor variance, and from largest to smallest according to the number of consecutive occurrences of the target spectral peak. The monitoring nodes with the highest ranking were selected as core monitoring nodes. For the UAV frequency samples collected by the core monitoring nodes, center frequency statistical features, frequency offset evolution, peak shape, time-frequency texture, and energy distribution features were extracted to form a standard frequency feature expression and construct a standard source domain feature library.

[0041] A cross-node migration calibration network is constructed by using the standard frequency feature representation of the core monitoring node as the source domain input and the frequency feature representation of the target monitoring node to be calibrated as the target domain input. The cross-node migration calibration network includes a source domain feature encoding branch, a target domain feature encoding branch, a shared feature mapping layer, and a distribution alignment constraint layer. The shared feature mapping layer consists of two 256-dimensional fully connected networks with an output embedding dimension of 128. The distribution alignment constraint layer uses the maximum mean difference as the distribution consistency constraint criterion. The two branches compress the input frequency features respectively. The shared feature mapping layer unifies the embedding space. The distribution alignment constraint layer constrains the consistency between the source domain feature distribution and the target domain feature distribution in the unified embedding space and outputs the calibration frequency feature representation of the target monitoring node.

[0042] In this implementation plan, core monitoring nodes are selected through a node quality assessment method. This ensures that the nodes with the best frequency feature quality and the widest coverage among the multi-space monitoring nodes are used to construct a standard source domain feature library, thus achieving the representativeness and stability of source domain features. This effectively improves the cross-node consistency, recognition accuracy, and real-time performance of UAV frequency system identification, providing a reliable guarantee for UAV frequency identification and positioning in complex airspace and multi-node collaborative monitoring scenarios.

[0043] Specifically, the process for conducting a consistency assessment is as follows:

[0044] The source domain intra-class scatter matrix and the target domain intra-class scatter matrix are obtained by summing the deviation vectors of samples of the same frequency category relative to their respective category feature centers in the standard source domain feature library of the core monitoring node and the frequency features to be calibrated of the target monitoring node, respectively, after performing outer product operations. The feature vector dimension is set to 128, the total number of frequency categories is C according to the UAV system, and the deviation vector is the sample feature minus the corresponding category center. When summing, the difference is divided by the number of samples in each category to obtain the average scatter matrix. The inter-class scatter matrix is ​​obtained by summing the deviation vectors of each frequency category feature center relative to the global feature center of all samples. The global feature center is the mean of the features of all source and target domain samples. The inter-class scatter matrix reflects the degree of separation between categories. The target domain feature covariance matrix is ​​obtained by averaging the deviation vectors of the frequency feature expression vector of the target monitoring node relative to the mean vector of the target domain features after performing outer product operations. Unbiased estimation is used, and the difference is divided by the number of target domain samples minus 1, resulting in a covariance matrix dimension of 128×128. The inverse square root matrix of the source domain eigenvalue covariance matrix is ​​obtained by reconstructing the matrix after eigenvalue decomposition of the source domain eigenvalue covariance matrix, taking the reciprocal of the square root of each positive eigenvalue. During eigenvalue decomposition, values ​​less than 1... e-8 The eigenvalues ​​are truncated to ensure numerical stability.

[0045] The joint intra-class discrete matrix is ​​obtained by adding the intra-class scatter matrices of the source and target domains; the inverse of the inter-class scatter matrix obtained by multiplying the adjustment parameter by the identity matrix is ​​obtained by adding the inter-class scatter matrix and the adjustment parameter by the identity matrix; the ratio term is obtained by multiplying the joint intra-class discrete matrix by the inter-class separation constraint matrix and performing matrix trace operation; the inverse square root of the source domain covariance is multiplied by the target domain covariance, and then multiplied by the inverse square root on the left and right, and added to the identity matrix to obtain the covariance structure alignment matrix, and the structural difference term is obtained by taking the natural logarithm of the determinant; the ratio term is divided by the structural difference term to obtain the migration calibration consistency evaluation value.

[0046] The specific formula for calculating the migration calibration consistency assessment value is as follows:

[0047] ;

[0048] In the formula, This represents the migration calibration consistency assessment value, which is used as the basis for determining whether to initiate the transfer learning calibration process; This represents the scatter matrix within the source domain, used to characterize the degree of dispersion of samples within the same frequency class in the source domain; This represents the scatter matrix within the target domain, used to characterize the degree of dispersion of samples within the same frequency category in the target domain; The inter-class scatter matrix represents the degree of separation between different frequency classes. This represents the adjustment parameter, obtained by setting a positive real constant. Its value ranges from zero to one, and it is used to prevent... In the process of finding the inverse, a singular situation may occur, ensuring that... reversible; Representing the identity matrix, through the inter-class scatter matrix The dimension is constructed to obtain the diagonal stability compensation for the inter-class scatter matrix; The inverse square root matrix of the source domain feature covariance matrix is ​​used to perform a whitening transformation on the target domain covariance structure based on the source domain reference. The target domain feature covariance matrix is ​​used to characterize the overall structural distribution relationship of the target domain frequency features in the unified embedding space. The trace operation is represented by the sum of the main diagonal elements of the target matrix. It is used to compress the ratio of the matrix formed by the intra-class scatter matrix and the inter-class scatter matrix into a single scalar to characterize the relative relationship between the degree of intra-class aggregation and the degree of inter-class separation. The determinant operation is represented by the matrix determinant, which is obtained by performing determinant calculation on the target matrix. It is used to characterize the degree of change in the covariance structure volume of the target domain after the source domain benchmark transformation, thereby reflecting the degree of deviation between the target domain and the source domain in the overall statistical structure.

[0049] In this implementation plan, by jointly analyzing the intra-class scattering, inter-class separation, and covariance structure of the standard source domain feature library of the core monitoring node and the frequency features to be calibrated of the target monitoring node, a quantitative assessment of the consistency of cross-node frequency features is achieved. This effectively improves the cross-node consistency, recognition accuracy, and real-time performance of multi-node UAV frequency identification, and enhances the robustness and applicability of the system in complex electromagnetic environments.

[0050] Specifically, the process of outputting calibration frequency feature frames and forming a calibration frequency feature set is as follows:

[0051] The migration calibration consistency assessment value and the migration calibration threshold are compared in real time. When the migration calibration consistency assessment value is greater than the migration calibration threshold, it is determined that there is a domain offset between the frequency characteristics of the current target monitoring node and the standard characteristics of the core monitoring node, and the migration learning calibration modeling process is initiated. When the migration calibration consistency assessment value is less than or equal to the migration calibration threshold, it is determined that the frequency characteristics of the current target monitoring node have met the cross-node consistency requirements.

[0052] In the transfer learning calibration modeling process, the standard frequency feature expression of the core monitoring node is used as the source domain sample input, and the node frequency feature frame of the target monitoring node is used as the target domain sample input to construct a cross-node transfer learning model. A three-layer fully connected neural network is adopted, with an input layer dimension of 128, hidden layer dimensions of 256 and 128, and an output layer dimension of 64. The activation function is ReLU, the optimizer is Adam, the learning rate is set to 0.001, and the loss function is the weighted sum of the source domain classification cross-entropy and the target domain MMD constraint. The input is nonlinearly mapped to output a uniform-dimensional embedded feature vector. The node frequency feature frames of the target monitoring node are corrected in real time to generate calibration frequency feature frames and are summarized to form a calibration frequency feature set.

[0053] In this implementation scheme, a calibration model is used to perform online deviation correction on the real-time input target monitoring node frequency feature frames, output calibrated frequency feature frames, and summarize them to form a calibrated frequency feature set, providing a highly reliable data foundation for UAV frequency activity segment identification, frequency system representation, and multi-node fusion positioning; effectively improving the cross-node consistency, identification accuracy, and real-time performance of multi-node UAV frequency identification under complex airspace conditions, and enhancing the generalization ability and robustness of small sample frequency features.

[0054] Specifically, the process of screening candidate segments based on the calibration frequency feature set and evaluating the candidate segments is as follows:

[0055] UAV frequency activity segments are extracted from the calibration frequency feature frame, and candidate segments are selected through joint discrimination. Feature encoding of the candidate segments is performed using hierarchical nonlinear mapping to extract center frequency evolution features, frequency offset variation features, spectral peak morphology features, time-frequency texture features, and energy distribution features, resulting in representation vectors. Matching centers are obtained by averaging the representation vectors of UAVs belonging to the same UAV system category. Non-matching centers are obtained by averaging the representation vectors of UAV system categories other than the current matching category. The non-matching category number is the number remaining after removing the current matching category number from the sequential numbers of all UAV system categories. Matching category numbers are obtained by comparing the similarity ranking of the representation vectors with each center.

[0056] Calculate the inner product of the current representation vector and the matching center, divide it by the sum of the product of the L2 norms of the current representation vector and the matching center and the zero-prevention term to obtain the matching similarity. Clip the matching similarity to the range (-1+δ, 1-δ) and obtain the matching enhancement term by applying the inverse hyperbolic tangent. Calculate the inner product of the current representation vector and each non-matching center, divide it by the sum of the product of the L2 norms of the representation vector and each non-matching center and the zero-prevention term to obtain the non-matching similarity. Clip each non-matching similarity to the range (-1+δ, 1-δ), apply the inverse hyperbolic tangent and select the maximum value to obtain the non-matching interference term. Subtract the non-matching interference term from the matching enhancement term to obtain the discrimination term. Calculate the inner product of the current and previous time-step representation vectors, divide it by the sum of the product of the L2 norms of the current and previous time-step representation vectors and the zero-prevention term to obtain the temporal similarity. Clip the temporal similarity to the range (-1+δ, 1-δ), multiply it by the temporal coefficient and add it to the discrimination term to obtain the candidate segment recognition evaluation value.

[0057] The specific formula for calculating the candidate fragment recognition evaluation value is as follows:

[0058] ;

[0059] In the formula, This represents the candidate segment identification evaluation value, which serves as the basis for determining whether the current target segment to be identified is a valid UAV frequency activity segment; This represents the current representation vector, used to characterize the frequency system feature representation of the target segment to be identified at the current time in the unified feature space; This indicates the current matching center, used to calculate the degree of matching between the current segment and the target category; This represents the j-th non-matching center, used to characterize the ability to distinguish between categories; The temporal coefficient is obtained by jointly optimizing the recognition accuracy and temporal stability index. It is a real number with a value greater than zero and is used to adjust the influence of the target segment corresponding to the previous time step on the recognition result at the current time step. The representation vector of the previous time step is used to calculate the temporal continuity between the current time step and the previous time step. This represents the zero-prevention term, which is obtained by giving a positive real constant. Its value range is a real number greater than zero and less than one, and it is used to avoid the denominator from having a zero value during the calculation process. This represents the inverse hyperbolic tangent mapping operation, which is obtained by performing an inverse hyperbolic tangent transformation on the input similarity ratio. It is used to amplify subtle differences between the center and the non-matching center. This represents the non-matching category number, used to traverse all centers except the most matching target category; This represents the matching category number, used to identify the drone system category to which the current segment most likely belongs; The L2 norm operation is obtained by taking the square root of the sum of the squares of the components of the input vector. The range of values ​​is specified to enable vectors with different amplitude scales to be matched and compared under a unified standard.

[0060] As shown in Table 1, the temporal category matching data is as follows: Time step 1: True category is 1, candidate segment recognition evaluation value is 0.9332, matching similarity is 0.7153, maximum non-matching similarity is 0.0271, and temporal similarity is 0.1247; Time step 2: True category is 3, candidate segment recognition evaluation value is 0.5599, matching similarity is 0.7379, maximum non-matching similarity is 0.3296, and temporal similarity is -0.0873; Time step 6: True category is 2, candidate segment recognition evaluation value is 1.2273, matching similarity is 0.7844, maximum non-matching similarity is 0.1502, and temporal similarity is 0. 0.6438; Time step 8: True category is 0, candidate segment recognition evaluation value is 1.0271, matching similarity is 0.7691, maximum non-matching similarity is 0.0544, temporal similarity is 0.1268; Time step 10: True category is 0, candidate segment recognition evaluation value is 0.9701, matching similarity is 0.7327, maximum non-matching similarity is -0.0262, temporal similarity is 0.0191; Time step 11: True category is 1, candidate segment recognition evaluation value is 1.0043, matching similarity is 0.7801, maximum non-matching similarity is -0.0149, temporal similarity is -0.1126.

[0061] Table 1 Time Series Category Matching Data Table

[0062] like Figure 3The graph shows the changes in the candidate segment identification evaluation value. The horizontal axis represents time steps, with each time step corresponding to a UAV frequency activity segment to be identified; the vertical axis represents the candidate segment identification evaluation value, reflecting the credibility of the current segment belonging to the target UAV frequency system in the identification model. Each point in the graph is calculated by combining the matching degree of the current segment's frequency system representation vector with the best-matching category prototype center, the interference degree with the non-matching category prototype center, and the temporal continuity with the segment from the previous time step. By observing the changing trend of the R value in the graph, we can see the fluctuations in the identification evaluation value of candidate segments at different time steps. High values ​​indicate that the current segment is more likely to belong to a valid UAV frequency activity segment, while low values ​​may correspond to background signals or abnormal frequency activities. This graph intuitively shows the dynamic changes in the candidate segment identification evaluation in continuous time series, providing a reference for subsequent effective segment selection, real-time identification decisions, and frequency system classification. It also helps to verify the stability and continuity of the identification model in small sample and complex airspace environments, providing technical support for achieving multi-node collaborative real-time identification and full-domain UAV monitoring.

[0063] In this implementation scheme, by jointly discriminating and hierarchically encoding candidate segments in the calibration frequency feature set, the accurate representation of UAV frequency activity segments in a multi-dimensional feature space is achieved. This improves the accuracy, temporal consistency, and generalization ability of candidate segment identification in the UAV frequency system, enhances the real-time identification stability and reliability of the system under multi-node collaboration and complex airspace conditions, and provides highly reliable data support for all-domain UAV monitoring and trajectory tracking.

[0064] Specifically, the process of obtaining effective UAV frequency activity segments is as follows:

[0065] The system compares the candidate segment identification evaluation value with the identification trigger threshold in real time. When the candidate segment identification evaluation value is greater than the identification trigger threshold, the current target segment to be identified is confirmed to be a valid UAV frequency activity segment. The segment is marked as valid, and the corresponding representation vector is sent to the identification network for further processing. When the candidate segment identification evaluation value is less than or equal to the identification trigger threshold, the current segment is determined to be a background interference segment and is not included in the main identification link. The segment is directly discarded and does not participate in the identification and fusion localization process to reduce computational overhead.

[0066] Effective UAV frequency activity segments are input into a lightweight one-dimensional convolutional neural network (CNN). The input representation vector is classified, and the corresponding UAV system category is output. The lightweight one-dimensional CNN consists of three convolutional layers and two fully connected layers. The kernel sizes are 3, 5, and 3, and the number of channels are 32, 64, and 128, respectively. The dimensions of the fully connected layers are 64 and the number of categories, respectively. ReLU is used as the activation function for all layers, and batch normalization and max pooling are applied after each convolutional layer. A transfer learning strategy is adopted, freezing the first few layers of the pre-trained network and fine-tuning the last few layers with a small learning rate, so that the recognition network maintains its generalization ability in small sample target scenarios. The network is pre-trained on a large dataset in the source domain. Then, the parameters of the first two convolutional layers are frozen, and only the last convolutional layer and the two fully connected layers are fine-tuned. The Adam optimizer with a learning rate of 0.0001 is used for 20 iterations during fine-tuning. The confidence of the recognition results is estimated to form category labels, system labels, recognition confidence levels, and segment validity indicators.

[0067] like Figure 4 The chart shows a comparison of the time-domain waveforms of four types of radio frequency (RF) signals. The chart compares the time-domain waveforms of three types of drones—Parrot AR.Drone, Parrot Bebop, and DJI Phantom—with background RF activities. AR.Drone and Bebop are two generations of consumer drones from Parrot, while Phantom is a DJI Phantom series aerial drone. The communication and image transmission mechanisms of these different models differ significantly, resulting in noticeable differences in time-domain amplitude: AR.Drone and Bebop exhibit obvious pulse spikes at the beginning, the background RF signal shows strong negative fluctuations, while the Phantom signal amplitude is stable with no obvious spikes. This chart visually demonstrates the differences in time-domain characteristics between different drones and background signals, providing an intuitive basis for frequency system identification.

[0068] like Figure 5 The figure shows the PCA feature distribution of four types of radio frequency signals. It represents the two-dimensional feature distribution after PCA dimensionality reduction, with 10 samples per type (40 samples in total) for Parrot AR.Drone, Parrot Bebop, DJI Phantom, and background radio frequency signals. AR.Drone, Bebop, and Phantom are mainstream consumer drones on the market, each employing different wireless communication and image transmission schemes, exhibiting different clustering characteristics in the feature space: the background signal sample regions are independently separable; the three types of drone signals are relatively clustered but partially overlapped, indicating that signals of the same type have similar features under small sample sizes, and there is some confusion between different drone models. This verifies the necessity of introducing transfer learning to improve generalization ability.

[0069] In this implementation scheme, the candidate fragment screening and feature encoding method can effectively capture such temporal features. Figure 5 The necessity of employing a transfer learning strategy of freezing pre-trained layers and fine-tuning with a small learning rate was verified. Transfer learning can enhance inter-class discrimination ability and improve recognition accuracy in scenarios with few samples. Experimental results show that the proposed method achieves an average recognition accuracy of 92.3% with only 10 samples per class, which is a significant improvement compared to directly training a CNN. Furthermore, the confidence assessment and validity indicator can effectively filter background interference and reduce the false alarm rate.

[0070] Specifically, the process of constructing a set of observations from the same source and evaluating the reliability of the fused localization is as follows:

[0071] Cross-node observation segments belonging to the same UAV target are associated to form a homogeneous observation set; the calibration frequency characteristics, frequency offset, arrival time, and node spatial distribution are jointly calculated to obtain the target position estimate; the number of monitoring nodes participating in the joint positioning calculation is counted to obtain the number of monitoring nodes; the monitoring nodes are numbered sequentially to obtain the monitoring node number; the local position constraint points of the nodes are inferred from the node observation direction, reception time, frequency boundary, and spatial coordinates; and the calibration frequency feature vector is obtained by structured extraction from the calibration frequency feature frames corresponding to the monitoring nodes. For the calibration frequency feature frames of each monitoring node, the center frequency, frequency offset, spectral peak amplitude, and time-frequency texture parameters are extracted according to the time window and frequency channel index, and then normalized and concatenated to form the calibration frequency feature vector.

[0072] The geometric consistency term is obtained by dividing the square of the distance between the target node's estimated location and the local location constraint points of each monitoring node by the square of the location parameter, adding 1, and raising the geometric exponent. The geometric deviation term is obtained by summing all geometric consistency terms, dividing by the number of monitoring nodes, and taking the reciprocal. The frequency consistency term is obtained by dividing the square of the difference between the calibration frequency eigenvectors of every two different monitoring nodes by the square of the frequency parameter. The cross-node consistency term is obtained by multiplying the number of monitoring nodes by the number of monitoring nodes minus 1 to obtain the total number of node pairs, summing all frequency consistency terms, dividing by the total number of node pairs, and taking the negative exponent. The time deviation term is obtained by dividing the square of the difference between the measured arrival time difference and the theoretical arrival time difference of every two different monitoring nodes by the square of the time parameter, summing all time deviation terms, dividing by the total number of node pairs, adding 1, and raising the time exponent to the negative power. The geometric deviation term, the cross-node consistency term, and the time deviation term are multiplied together to obtain the fusion positioning reliability assessment value.

[0073] The specific formula for calculating the reliability assessment value of fusion positioning is as follows:

[0074] ;

[0075] In the formula, This represents the fusion positioning reliability assessment value, which is used as the basis for determining whether to output the current location result and trajectory result; This indicates the number of monitoring nodes, used to determine the statistical scope. This represents the monitoring node number, used to traverse all monitoring nodes participating in the current joint positioning; The location estimation results are used as the basis for geometric consistency assessment and theoretical time difference of arrival inversion. This represents a local position constraint point of a node, used to calculate the degree of deviation of the joint positioning result from the single-node constraint; The geometric index is obtained by fitting or optimizing the relationship between the geometric deviation and positioning accuracy between the joint positioning results and the local position constraint points of each node. It is a real number with a value greater than zero and is used to make the nodes with larger geometric deviations more significantly suppress the reliable evaluation value. The square of the position parameter is obtained by statistically analyzing the squared distances between the joint positioning results and the corresponding local position constraint points in the effective positioning samples. The value range is a real number greater than zero, which is used to avoid inconsistent impacts of position residuals on the reliable evaluation value under different spatial scales. This represents the number of the j-th monitoring node calculated in pair with the ith monitoring node. It is obtained by sequentially numbering the monitoring nodes in the current homogeneous observation set and is used to construct pairwise node pairing relationships. This represents the calibration frequency feature vector, used to calculate the degree of homogeneity across nodes; It represents the measured arrival time difference between the i-th monitoring node and the j-th monitoring node. It is obtained by reading the arrival timestamps of the i-th and j-th monitoring nodes for the same target frequency activity segment and subtracting them. It is used to characterize the actual propagation time difference of the same target signal between different monitoring nodes. The theoretical time difference of arrival between the i-th and j-th monitoring nodes is calculated using the target position estimation result obtained from joint positioning solution, node spatial coordinates, and signal propagation speed. It is used to characterize the propagation time difference relationship that should be satisfied between different monitoring nodes under the condition that the current positioning result is true. The time suppression index is determined by fitting the response relationship between positioning error and time difference deviation. It is a real number with a value greater than zero and is used to adjust the suppression strength of the time difference deviation term on the fusion positioning reliability assessment value. The square of the frequency parameter is obtained by statistically analyzing the squared differences between the calibration frequency feature vectors output by different monitoring nodes in the same source target sample. The value range is a real number greater than zero, which is used to make the feature deviations under different feature amplitude conditions comparable under a unified standard. The square of the time parameter is obtained by statistically analyzing the square of the difference between the measured arrival time difference and the theoretical arrival time difference in historical valid positioning samples. The value range is a real number greater than zero, which is used to avoid the influence of differences in time units on the reliability assessment results of fusion positioning.

[0076] In this implementation plan, by associating cross-node observation segments with the same source, a unified set of observations with the same source is constructed, which realizes the effective integration of multi-node observation data belonging to the same UAV target; by jointly solving the calibration frequency characteristics, frequency offset, arrival time and spatial distribution information of the nodes, the target position estimate is generated, realizing multi-node fusion positioning and improving positioning accuracy.

[0077] Specifically, the process of determining the joint localization result of the current node and generating the target behavior label is as follows:

[0078] The system compares the fusion positioning confidence assessment value and the fusion positioning threshold in real time. When the fusion positioning confidence assessment value is greater than the fusion positioning threshold, the joint positioning result is confirmed to be valid, and the target location result and trajectory are output. The location result is stored in latitude, longitude and altitude coordinates, and the trajectory is stored in the buffer in chronological order with a timestamp attached. When the fusion positioning confidence assessment value is less than or equal to the fusion positioning threshold, it is determined that there is abnormal node interference, and the result is not output temporarily, waiting for the time window to supplement the verification. The time window length is set to 3 seconds, during which new observation data is continuously collected and the fusion positioning confidence assessment value is recalculated. The valid positioning result is updated and associated with the target trajectory in real time to form a global motion trajectory. Kalman filtering is used to smooth and predict the position at continuous time moments. The nearest neighbor method is used to associate targets at adjacent time moments. The target identification result, positioning result, trajectory and original fragment index are structured and bound and stored in the time series database in JSON format. Each field corresponds to an independent index. Target behavior labels are generated and corresponding abnormal states are marked. Abnormal states include node failure, multipath interference and clock out of synchronization, which are marked in binary mask form.

[0079] like Figure 6 The diagram shows the confusion matrix for four types of radio frequency signals. The image represents the confusion matrix between AR.Drone, Bebop, and Phantom drones and background radio frequency signals under small-sample training. Rows in the matrix represent the true categories, and columns represent the model-predicted categories. The results show that, with only 40 samples, the traditional lightweight model has limited accuracy in recognizing Parnott series drones (AR.Drone, Bebop), DJI Phantom drones, and background signals, and is prone to misclassification between different drone types. This demonstrates the model's weak generalization ability and insufficient robustness in real-time recognition under small-sample scenarios, providing experimental support for the recognition scheme that integrates transfer learning and lightweight networks as proposed in this method.

[0080] In this implementation scheme, by comparing the fusion positioning reliability evaluation value and the threshold in real time, the joint positioning results of multiple nodes are effectively judged, ensuring that only reliable joint positioning results are used to output the target position and trajectory. This verifies the generalization ability and real-time recognition robustness of the transfer learning and lightweight network fusion scheme in small sample scenarios, effectively improving the real-time recognition accuracy of UAV frequency system, the reliability of trajectory generation, and the overall stability of the system.

[0081] Specifically, the second aspect of this invention provides a real-time UAV frequency system identification system fused with transfer learning, applied to a real-time UAV frequency system identification method fused with transfer learning, comprising: an airspace monitoring node frequency data acquisition module, used to dynamically access and acquire node frequency observation data through fixed, vehicle-mounted, and edge monitoring nodes, and perform time axis relocation, noise reduction, normalization, and spectral base leveling to construct a node frequency observation dataset containing original IQ sampling, power spectrum, time-frequency diagram, and node status information; and a cross-node transfer calibration module, used to select a core based on the node frequency observation dataset by evaluating signal-to-noise ratio, frequency stability, and geographical coverage. The system consists of a core monitoring node, a standard source domain feature library, and a migration calibration consistency assessment based on scatter matrix and covariance alignment. It outputs calibration frequency feature frames to form a calibration frequency feature set. A real-time frequency system identification module is used to filter candidate segments based on the calibration frequency feature set through joint discrimination, and to evaluate the candidate segments to obtain valid UAV frequency activity segments confirmed by the identification trigger threshold. A node fusion positioning module is used to associate cross-node observation segments of the same target to construct a common-source observation set, evaluate the reliability of the fusion positioning, determine the current node joint positioning result, and generate target behavior labels containing target identification, location trajectory, and abnormal status annotations.

[0082] This implementation plan effectively improves the cross-node consistency, recognition accuracy, and trajectory generation reliability of real-time identification of UAV frequency systems, enhances the stability and robustness of the system in complex airspace, multi-node collaboration, and small sample environments, and provides highly reliable data support for all-domain UAV monitoring and trajectory tracking.

[0083] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, 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 process, method, article, or apparatus.

[0084] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A real-time identification method for UAV frequency systems integrating transfer learning, characterized in that, Includes the following steps: S1. Obtain node frequency observation data and perform time axis relocation, noise reduction, normalization and spectral base leveling to construct a node frequency observation dataset. S2, based on the node frequency observation dataset, select core monitoring nodes, construct a standard source domain feature library, conduct consistency evaluation, output calibration frequency feature frames, and form a calibration frequency feature set; S3, based on the calibration frequency feature set, candidate segments are screened, and candidate segments are identified and evaluated to obtain effective UAV frequency activity segments; S4. Construct a set of observations from the same source, evaluate the reliability of the fused positioning, determine the joint positioning result of the current node, and generate target behavior labels.

2. The real-time identification method for UAV frequency systems based on transfer learning according to claim 1, characterized in that: The specific process of acquiring node frequency observation data and performing time axis relocation, denoising, normalization, and spectral base leveling to construct the node frequency observation dataset is as follows: The system dynamically accesses fixed monitoring nodes, vehicle-mounted monitoring nodes, and edge monitoring nodes deployed in different airspace locations throughout the entire process, and acquires raw IQ sampling data, power spectrum data, time-frequency diagram data, center frequency estimation results, frequency offset observation results, spectral peak shape description information, and node local reception status information as node frequency observation data. By using a unified time base with both NTP and PTP protocols, the acquisition time, segment number, and analysis window of each monitoring node are relocated along the timeline. Each monitoring node is quantized and encoded to construct a node contextual status label, which is then bound and stored with the corresponding timestamp frequency observation segment. Wavelet transform and impulse suppression algorithms are used to denoise and filter out impulse interference and spurious components. The frequency offset caused by local oscillator error and link differences is identified and compensated through center frequency drift correlation. The node frequency observation dataset is constructed by normalizing the spectral amplitude and leveling the spectral base of the node frequency observation data.

3. The real-time identification method for UAV frequency systems based on transfer learning according to claim 1, characterized in that: The specific process of selecting core monitoring nodes and constructing a standard source domain feature library based on the node frequency observation dataset is as follows: Based on the node frequency observation dataset, core monitoring nodes are selected by considering the node quality assessment method of signal-to-noise ratio, frequency stability and geographical coverage. For the UAV frequency samples collected by the core monitoring nodes, the center frequency statistical features, frequency offset evolution, spectral peak morphology, time-frequency texture and energy distribution features are extracted to form a standard frequency feature expression and construct a standard source domain feature library. The standard frequency feature representation of the core monitoring node is used as the source domain input, and the frequency feature representation of the target monitoring node to be calibrated is used as the target domain input to construct a cross-node migration calibration network. The cross-node migration calibration network includes a source domain feature encoding branch, a target domain feature encoding branch, a shared feature mapping layer, and a distribution alignment constraint layer. The two branches compress the input frequency features respectively. The shared feature mapping layer unifies the embedding space, and the distribution alignment constraint layer constrains the consistency between the source domain feature distribution and the target domain feature distribution in the unified embedding space, outputting the calibration frequency feature representation of the target monitoring node.

4. The real-time identification method for UAV frequency systems based on transfer learning according to claim 1, characterized in that: The specific process for conducting the consistency assessment is as follows: The source domain intra-class scatter matrix and the target domain intra-class scatter matrix are obtained by performing outer product operations on the deviation vectors of the same frequency category samples relative to their respective category feature centers in the standard source domain feature library of the core monitoring node and the target monitoring node to be calibrated frequency features. The inter-class scatter matrix is ​​obtained by performing outer product operations on the deviation vectors of each frequency category feature center relative to the global feature center of all samples. The target domain feature covariance matrix is ​​obtained by performing outer product operations on the deviation vectors of the target monitoring node's to-be-calibrated frequency feature expression vector relative to the target domain feature mean vector and averaging them. The inverse square root matrix of the source domain feature covariance matrix is ​​obtained by performing eigenvalue decomposition on the source domain feature covariance matrix, taking the reciprocal of the square root of each positive eigenvalue, and reconstructing the matrix. The joint intra-class discrete matrix is ​​obtained by adding the intra-class scatter matrices of the source and target domains; the inverse of the inter-class scatter matrix obtained by multiplying the adjustment parameter by the identity matrix is ​​obtained by adding the inter-class scatter matrix and the adjustment parameter by the identity matrix; the ratio term is obtained by multiplying the joint intra-class discrete matrix by the inter-class separation constraint matrix and performing matrix trace operation; the inverse square root of the source domain covariance is multiplied by the target domain covariance, and then multiplied by the inverse square root on the left and right, and added to the identity matrix to obtain the covariance structure alignment matrix, and the structural difference term is obtained by taking the natural logarithm of the determinant; the ratio term is divided by the structural difference term to obtain the migration calibration consistency evaluation value.

5. The real-time identification method for UAV frequency systems based on transfer learning according to claim 1, characterized in that: The specific process of forming a calibration frequency feature set from the output calibration frequency feature frame is as follows: The migration calibration consistency assessment value and the migration calibration threshold are compared in real time. When the migration calibration consistency assessment value is greater than the migration calibration threshold, it is determined that there is a domain offset between the frequency characteristics of the current target monitoring node and the standard characteristics of the core monitoring node, and the migration learning calibration modeling process is initiated. When the migration calibration consistency assessment value is less than or equal to the migration calibration threshold, it is determined that the frequency characteristics of the current target monitoring node have met the cross-node consistency requirements. In the process of transfer learning calibration modeling, the standard frequency feature expression of the core monitoring node is used as the source domain sample input, and the node frequency feature frame of the target monitoring node is used as the target domain sample input to construct a cross-node transfer learning model. A multi-layer fully connected neural network structure is adopted to perform nonlinear mapping on the input and output a unified dimension embedded feature vector. The node frequency feature frame of the target monitoring node is corrected in real time to generate calibration frequency feature frames and be summarized to form a calibration frequency feature set.

6. The real-time identification method for UAV frequency systems based on transfer learning according to claim 1, characterized in that: The specific process of screening candidate segments based on the calibration frequency feature set and evaluating the candidate segments is as follows: UAV frequency activity segments are extracted from the calibration frequency feature frame, and candidate segments are selected through joint discrimination. Feature encoding of the candidate segments is performed using hierarchical nonlinear mapping to obtain representation vectors. Matching centers are obtained by averaging the representation vectors of UAVs belonging to the same UAV system category. Non-matching centers are obtained by averaging the representation vectors of UAV system categories other than the current matching category. The non-matching category number is the number remaining after removing the current matching category number from the sequential numbers of all UAV system categories. Matching category numbers are obtained by comparing the similarity ranking of the representation vectors with each center. Calculate the inner product of the current representation vector and the matching center, divide it by the sum of the product of the L2 norms of the current representation vector and the matching center and the zero-prevention term to obtain the matching similarity. Obtain the matching enhancement term by applying the inverse hyperbolic tangent. Calculate the inner product of the current representation vector and each non-matching center, divide it by the sum of the product of the L2 norms of the representation vector and each non-matching center and the zero-prevention term to obtain the non-matching similarity. Obtain the non-matching interference term by applying the inverse hyperbolic tangent and selecting the maximum value. Subtract the non-matching interference term from the matching enhancement term to obtain the discrimination term. Calculate the inner product of the current and previous time-step representation vectors, divide it by the sum of the product of the L2 norms of the current and previous time-step representation vectors and the zero-prevention term to obtain the temporal similarity. Multiply it by the temporal coefficient and add it to the discrimination term to obtain the candidate segment recognition evaluation value.

7. The real-time identification method for UAV frequency systems based on transfer learning according to claim 1, characterized in that: The specific process for obtaining effective UAV frequency activity segments is as follows: The system compares the candidate segment identification evaluation value with the identification trigger threshold in real time. When the candidate segment identification evaluation value is greater than the identification trigger threshold, it confirms that the current target segment to be identified belongs to a valid UAV frequency activity segment. When the candidate segment identification evaluation value is less than or equal to the identification trigger threshold, it determines that the current segment is a background interference segment and does not enter the identification main link. Valid UAV frequency activity segments are input into a lightweight deep recognition network. The input representation vector is classified and judged, and the UAV system category corresponding to the current target is output. A transfer learning strategy of freezing some network layers and adjusting the learning rate is adopted to enable the recognition network to maintain its generalization ability in small sample target scenarios. The credibility of the recognition results is estimated to form category label, system label, recognition confidence level and segment validity label.

8. The real-time identification method for UAV frequency systems based on transfer learning according to claim 1, characterized in that: The specific process for constructing a set of observations from the same source and evaluating the reliability of the fused localization is as follows: Cross-node observation segments belonging to the same UAV target are associated to form a homogeneous observation set; the calibration frequency characteristics, frequency offset, arrival time, and node spatial distribution are jointly calculated to obtain the target position estimate; the number of monitoring nodes participating in the joint positioning calculation is counted to obtain the number of monitoring nodes; the monitoring nodes are numbered sequentially to obtain the monitoring node number; the local position constraint points of the nodes are inferred from the node observation direction, reception time, frequency boundary, and spatial coordinates; and the calibration frequency feature vector is obtained by structured extraction from the calibration frequency feature frames corresponding to the monitoring nodes. The geometric consistency term is obtained by dividing the square of the distance between the target node's estimated location and the local location constraint points of each monitoring node by the square of the location parameter, adding 1, and raising the geometric exponent. The geometric deviation term is obtained by summing all geometric consistency terms, dividing by the number of monitoring nodes, and taking the reciprocal. The frequency consistency term is obtained by dividing the square of the difference between the calibration frequency eigenvectors of every two different monitoring nodes by the square of the frequency parameter. The cross-node consistency term is obtained by multiplying the number of monitoring nodes by the number of monitoring nodes minus 1 to obtain the total number of node pairs, summing all frequency consistency terms, dividing by the total number of node pairs, and taking the negative exponent. The time deviation term is obtained by dividing the square of the difference between the measured arrival time difference and the theoretical arrival time difference of every two different monitoring nodes by the square of the time parameter, summing all time deviation terms, dividing by the total number of node pairs, adding 1, and raising the time exponent to the negative power. The geometric deviation term, the cross-node consistency term, and the time deviation term are multiplied together to obtain the fusion positioning reliability assessment value.

9. The real-time identification method for UAV frequency systems based on transfer learning according to claim 1, characterized in that: The specific process of determining the joint localization result of the current node and generating the target behavior label is as follows: The system compares the fusion positioning confidence assessment value and the fusion positioning threshold in real time. When the fusion positioning confidence assessment value is greater than the fusion positioning threshold, the joint positioning result is confirmed to be valid, and the target location result and trajectory are output. When the fusion positioning confidence assessment value is less than or equal to the fusion positioning threshold, it is determined that there is abnormal node interference, and the result is not output temporarily, waiting for the time window to supplement the verification. The system updates and associates the target trajectory in real time with the valid positioning result to form a global motion trajectory. The system binds the target recognition result, positioning result, trajectory and original fragment index in a structured manner, generates target behavior labels and marks the corresponding abnormal states. Abnormal states include node failure, multipath interference and clock out of sync.

10. A real-time UAV frequency system identification system integrating transfer learning, employing the real-time UAV frequency system identification method integrating transfer learning as described in any one of claims 1-9, characterized in that, include: The airspace monitoring node frequency data acquisition module is used to acquire node frequency observation data and perform time axis relocation, noise reduction, normalization and spectral base leveling to construct a node frequency observation dataset. The cross-node migration calibration module is used to select core monitoring nodes based on the node frequency observation dataset, construct a standard source domain feature library, perform consistency evaluation, output calibration frequency feature frames, and form a calibration frequency feature set. The frequency system real-time identification module is used to screen candidate segments based on the calibration frequency feature set, and to identify and evaluate the candidate segments to obtain effective UAV frequency activity segments. The node fusion localization module is used to construct a set of observations from the same source, evaluate the reliability of the fusion localization, determine the joint localization result of the current node, and generate target behavior labels.

Images