Unmanned aerial vehicle radio signal cross-scene detection method based on transfer learning

Abstract

The application relates to the technical field of unmanned aerial vehicle signal detection, and discloses a method for cross-scene detection of unmanned aerial vehicle radio signals based on transfer learning. The method first collects unmanned aerial vehicle radio signals in a target area to generate original signal spectrum data, and then divides multiple signal detection scene categories according to spectrum distribution characteristics. The original data is processed by a multi-scale feature extraction network to generate a signal feature space with cross-scene invariance, which is input into the source domain and target domain channels of a double-channel residual transfer learning model. The feature alignment module in the model dynamically compensates for the feature distribution difference between the source domain and the target domain scene, and generates a cross-scene detection result through a decision tree integration mechanism. Finally, based on the matching degree of the result and a preset signal feature library, the identity of the unmanned aerial vehicle and the behavior type are output. The method can effectively cope with the signal feature difference in different scenes, and improve the adaptability and accuracy of unmanned aerial vehicle detection.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) signal detection technology, specifically to a method for cross-scene detection of UAV radio signals based on transfer learning. Background Technology

[0002] With the rapid development of drone technology, its applications in aerial photography, logistics, surveying and mapping are becoming increasingly widespread. However, this also brings security risks such as privacy leaks, making effective drone detection a pressing issue. Currently, drone detection mainly relies on radio signal analysis, capturing communication and navigation signals between the drone and ground stations to identify and track the drone.

[0003] Drone radio signals exhibit significant differences across various scenarios. For instance, urban environments are plagued by electromagnetic interference, and building obstructions can cause signal reflection and diffraction, resulting in distorted signal spectra. In contrast, while signal propagation paths are relatively simple in open suburban or mountainous areas, factors such as terrain undulations and vegetation blockage can still affect signal strength and spectral characteristics, leading to substantial differences from urban environments. Furthermore, varying weather conditions also impact radio signal propagation, further exacerbating signal instability.

[0004] Most existing signal detection methods are models trained in specific scenarios. When applied to new scenarios, the change in signal feature distribution due to scenario differences leads to a significant drop in detection accuracy. Traditional machine learning methods require the collection of large amounts of data for each new scenario for model training, which is not only time-consuming and labor-intensive but also difficult to implement in scenarios where data collection is challenging. Although some methods attempt to enhance model adaptability through feature engineering, manually designed features often fail to cover all scenario variations, resulting in limited generalization ability.

[0005] Drones come in a variety of types, and different brands and models use different radio signal frequency bands and modulation methods, which increases the difficulty of cross-scene detection. Existing methods are prone to confusion when processing signals from different types of drones, especially in complex scenes where signal characteristics are interfered with, making it even more difficult to accurately identify the drone's identity and behavior. Summary of the Invention

[0006] The purpose of this invention is to provide a method for cross-scene detection of UAV radio signals based on transfer learning, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, this invention provides a method for cross-scene detection of UAV radio signals based on transfer learning, the method comprising:

[0008] Collect radio signals from drones within the target area and generate raw signal spectrum data;

[0009] Based on the spectral distribution characteristics reflected by the original signal spectral data, multiple signal detection scenario categories are divided;

[0010] The original signal spectrum data is processed by a multi-scale feature extraction network to generate a signal feature space with cross-scene invariance.

[0011] Construct a dual-channel residual transfer learning model by inputting the signal feature space into the source domain channel and the target domain channel of the dual-channel residual transfer learning model;

[0012] The feature alignment module in the dual-channel residual transfer learning model is used to dynamically compensate for the feature distribution differences between the source domain scene and the target domain scene.

[0013] Based on the output after feature alignment, a decision tree ensemble mechanism is used to generate cross-scene detection results;

[0014] Based on the degree of matching between the cross-scene detection results and the preset signal feature library, the drone's identity and behavior type are output.

[0015] Preferably, the step of classifying multiple signal detection scene categories includes:

[0016] Extract the frequency domain energy distribution parameters, time domain pulse density parameters, and modulation type parameters from the original signal spectrum data;

[0017] The frequency domain energy distribution parameters, time domain pulse density parameters, and modulation type parameters are jointly analyzed using a spectral clustering algorithm.

[0018] Based on the intra-class similarity threshold and inter-class difference threshold in the clustering results, a set of signal detection scene categories including densely populated urban areas, open suburban areas, and complex mountainous areas is generated.

[0019] Preferably, the step of generating a signal feature space with cross-scene invariance includes:

[0020] The original signal spectrum data is spatially compressed using a group of convolutional kernels to generate a primary feature map.

[0021] Temporal dependencies in the primary feature map are extracted using a long short-term memory network.

[0022] By integrating the spatial dimension compression result with the temporal dependency, a dynamic signal feature fusion module is constructed.

[0023] Scene category weight coefficients are introduced into the signal feature dynamic fusion module to generate cross-scene invariant feature vectors.

[0024] Preferably, the step of constructing the dual-channel residual transfer learning model includes:

[0025] Set the number of deep residual network layers for the source domain channel and the target domain channel;

[0026] A feature transfer path is established between the source domain channel and the target domain channel, and the feature transfer path contains a learnable transfer weight matrix;

[0027] Configure the domain difference loss function in the feature alignment module, which is constructed based on the maximum mean difference criterion.

[0028] Preferably, the step of dynamically compensating for the feature distribution differences includes:

[0029] Calculate the difference in covariance matrices between the feature distributions of the source domain scene and the feature distributions of the target domain scene;

[0030] Generate feature distribution compensation coefficients based on the covariance matrix difference values;

[0031] An adaptive weighting mechanism is used to apply the feature distribution compensation coefficients to the output features of the target domain channel;

[0032] The compensation parameters of the feature alignment module are updated through gradient backpropagation.

[0033] Preferably, the method further includes an incremental training process:

[0034] When a new signal detection scenario category is added, the source domain channel parameters of the dual-channel residual transfer learning model are frozen.

[0035] Only update the deep residual network parameters and feature alignment module parameters of the target domain channel;

[0036] The learning rate of the target domain channel is adjusted using a momentum optimization algorithm until the model convergence error is lower than a preset threshold.

[0037] Preferably, the step of generating cross-scene detection results using a decision tree ensemble mechanism includes:

[0038] The feature-aligned output is then fed into a random forest classifier.

[0039] Construct multiple decision trees using the Gini coefficient splitting criterion;

[0040] Aggregate the output probability distributions of each decision tree to generate a cross-scene detection confidence vector;

[0041] The final detection result is determined based on the index of the maximum value in the confidence vector.

[0042] Preferably, the method further includes an interference signal suppression mechanism:

[0043] An interference feature filtering layer is added to the decision tree integration mechanism;

[0044] Extract the entropy features of the cross-scene detection confidence vector;

[0045] When the entropy value exceeds the dynamic interference threshold, the interference suppression filter is activated.

[0046] The interference suppression filter performs band-stop filtering on the input signal and then regenerates the detection result.

[0047] Preferably, the step of outputting the drone's identity and behavior type includes:

[0048] The cross-scene detection results are matched with a preset signal feature library using cosine similarity.

[0049] When the matching similarity reaches the authentication threshold, the registration information stored in the preset signal feature library is extracted;

[0050] The drone's identity is parsed based on the device code field in the registration information;

[0051] Identify takeoff, cruise, and hovering behavior types by combining signal temporal activity patterns.

[0052] Preferably, the method further includes a multimodal interference verification mechanism:

[0053] Simultaneously acquire the environmental background spectrum characteristics of the time window corresponding to the original signal spectrum data;

[0054] Construct the covariance feature matrix of UAV signal characteristics and environmental background spectrum characteristics;

[0055] The confidence level of signal-environment coupling interference is calculated using the covariance feature matrix.

[0056] When the confidence level of the signal-environment coupling interference reaches the interference verification threshold, the multi-source verification process is triggered.

[0057] The multi-source verification process calls historical spectrum data from neighboring detection nodes to perform spatiotemporal consistency verification.

[0058] The confidence weight coefficient of the cross-scene detection results is adjusted based on the verification results.

[0059] Compared with the prior art, the beneficial effects of the present invention are:

[0060] By collecting radio signals from drones within the target area and generating raw signal spectrum data, basic materials are provided for subsequent signal processing. Classifying signal detection scenarios into multiple categories based on spectrum distribution characteristics allows for a clearer understanding of the signal characteristics in different scenarios, making subsequent processing more targeted.

[0061] Multi-scale feature extraction networks can process raw signal spectral data to extract signal features at different scales and generate a signal feature space with cross-scene invariance. This reduces the impact of scene changes on the core features of the signal, making the signal features more stable and facilitating effective signal recognition in different scenarios.

[0062] The construction of the dual-channel residual transfer learning model involves inputting the signal feature space into the source domain channel and the target domain channel respectively. By leveraging the advantages of the residual structure, the key features of the signal can be effectively preserved. At the same time, through the idea of ​​transfer learning, the knowledge of the source domain is transferred to the target domain, enabling the model to function effectively in new scenarios without extensive retraining.

[0063] The feature alignment module dynamically compensates for the feature distribution differences between the source and target domain scenes. It can adjust the feature differences in different scenes in real time, so that the features of the source and target domains can be better matched in the same space, thereby reducing the detection bias caused by different scenes.

[0064] The decision tree ensemble mechanism generates cross-scene detection results based on the output after feature alignment. It integrates the advantages of multiple decision trees, can combine judgments from different perspectives, reduces the limitations of a single model, makes the detection results more reliable, and can better cope with complex and ever-changing scenarios.

[0065] Based on the matching of cross-scene detection results with a preset signal feature library, the system outputs the drone's identity and behavior type. It can associate the detection results with known information to achieve accurate identification of drones. Whether in cities, suburbs or other complex environments, it can adapt well to scene changes and improve the effective control of drones. Attached Figure Description

[0066] Figure 1 This is a schematic diagram illustrating the working principle of the UAV radio signal cross-scene detection method based on transfer learning described in this invention.

[0067] Figure 2 A flowchart for classifying signal detection scenarios;

[0068] Figure 3 A flowchart for dynamic compensation of the feature alignment module;

[0069] Figure 4 A flowchart for generating detection results using a decision tree ensemble mechanism;

[0070] Figure 5 This is a flowchart of the multimodal interference verification mechanism. Detailed Implementation

[0071] 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.

[0072] Please see Figure 1 This invention provides a method for cross-scene detection of UAV radio signals based on transfer learning, the method comprising:

[0073] Broadband radio frequency receiving equipment is used to collect UAV radio signals within the target area, with a signal sampling rate of no less than 100MHz, generating raw signal spectrum data containing time and frequency information. The raw data is preprocessed to form a standardized spectrum diagram, with a time resolution set to 10ms and a frequency resolution adjusted to 1kHz.

[0074] Feature analysis is performed on the raw signal spectrum data to extract frequency domain energy gradient, time domain pulse interval, and modulation cycle spectrum features. A spectral clustering algorithm is then used to classify the signal detection scene categories. During the clustering process, a dynamic similarity threshold is set to automatically generate three scene labels: densely populated urban areas, open suburban areas, and complex mountainous areas.

[0075] A multi-scale feature extraction network is constructed. The first layer uses a 7×7 convolutional kernel for spatial downsampling, and the second layer uses a bidirectional LSTM to extract temporal features. A scene category attention mechanism is introduced into the network output layer to generate cross-scene invariant feature vectors through a trainable weight coefficient matrix, with the feature dimension fixed at 256.

[0076] A dual-channel residual transfer learning model is designed, with the source domain channel employing a 12-layer ResNet structure and the target domain channel using an 8-layer ResNet structure. A cross-domain feature transfer path is established between the two channels, and the transfer weight matrix is ​​initialized as an identity matrix and dynamically updated through backpropagation.

[0077] Dynamic distribution compensation is implemented in the feature alignment module. The Jensen-Shannon divergence between the source and target domain features is calculated as a measure of difference, and adaptive compensation coefficients are generated and applied to the target domain feature mapping. The update frequency of the compensation coefficients is synchronized with the model training cycle.

[0078] Random forest is used as the decision tree ensemble mechanism, with 50 decision trees as base classifiers, and the maximum depth of each tree is limited to 10 layers. The classifier input is an aligned 256-dimensional feature vector, and the output dimension is consistent with the number of drone behavior types.

[0079] A pre-defined signal feature database is established to store the radio frequency fingerprint features of known drone models, including 12 parameters such as carrier frequency deviation and phase noise spectrum. The matching process uses cosine similarity calculation, and the authentication threshold is set to 0.85. When a match is successful, the drone ID and current flight mode from the registration information are output.

[0080] Example 1: See Figure 2 The process involves two key steps: scene classification for signal detection and generation of cross-scene invariant features. In the scene classification stage, a multi-parameter joint analysis method is used to process the raw signal spectrum data. The extraction of frequency domain energy distribution parameters employs windowed short-time Fourier transform technology, normalizing the energy distribution within each time window and calculating the statistical characteristics of the energy proportion of each frequency band. Time-domain pulse density analysis is achieved through an adaptive threshold detection algorithm. First, the signal envelope is smoothed, then the rising and falling edges of the pulses are detected, and the number of pulses per unit time and the coefficient of variation of their time intervals are statistically analyzed. Modulation type identification uses a high-order cyclic spectrum analysis method. By calculating the spectral correlation density function of the signal, the cyclic frequency characteristics unique to the modulation method are extracted. These characteristics can effectively distinguish UAV signals with different modulation types.

[0081] The spectral clustering algorithm employs a multi-stage optimization strategy. In the feature preprocessing stage, frequency domain energy, time domain pulse, and modulation type features are standardized to eliminate dimensional differences. An adaptive kernel function is used when constructing the similarity matrix, with the kernel width parameter automatically adjusted based on the density of the feature space. Constraints are introduced during clustering, requiring that the similarity of samples within the same category must exceed a preset threshold, and that the difference between samples from different categories must reach a minimum separation. During algorithm iteration, cluster centers are dynamically adjusted until the ratio of intra-cluster dispersion to inter-cluster distance reaches a stable state. The final output set of scene categories includes three typical environmental types, each with unique signal propagation characteristics. Urban densely populated areas are characterized by significant multipath effects and large signal delay spread; open suburban areas are characterized by low background noise and a dominant direct signal path; and complex mountainous areas are characterized by significant Doppler shift and drastic signal intensity fluctuations.

[0082] Cross-scene invariant feature generation is implemented using a deep neural network architecture. The spatial dimension compression module consists of multiple convolutional layers, with convolutional kernels of different sizes capturing local details and global structural features of the signal spectrum. Batch normalization is introduced after each convolutional operation to accelerate network convergence and improve feature stability. The temporal feature extraction module adopts a bidirectional long short-term memory network structure, learning the dependence of signal spectrum changes over time through a gating mechanism. Gradient pruning is used in the update process of the network hidden state to prevent gradient explosion during training. A cross-attention mechanism is designed in the feature fusion stage to dynamically weight and combine spatial and temporal features. The generation process of attention weights considers scene category information, converting scene labels into feature weight vectors through a trainable mapping function.

[0083] The scene category weight coefficients are generated using an end-to-end training approach. The weight generation network consists of stacked fully connected layers and non-linear activation functions. The input is the embedded representation of the scene category, and the output is the adjustment coefficients for each feature channel. During training, a multi-task learning strategy is employed, simultaneously optimizing the parameters of the feature extraction network and the weight generation network. Layer-level normalization is used for feature vector normalization to maintain scale consistency across all feature dimensions. The resulting cross-scene invariant feature vectors have clear physical meanings. The low-frequency components primarily reflect the carrier characteristics of the UAV, the high-frequency components mainly characterize the modulation details of the signal, and the time-varying components reflect the changing trends of the flight state.

[0084] In terms of specific implementation details, the extraction of frequency domain energy distribution parameters employs overlapping framing technology, with the frame shift set to one-quarter of the window length to improve temporal resolution. The temporal pulse detection algorithm uses a dual-threshold strategy to effectively distinguish between real signal pulses and noise fluctuations. During modulation type identification, the resolution setting of cyclic spectrum analysis comprehensively considers computational complexity and feature discriminative capability. The spectral clustering algorithm is implemented using an approximate solution method, with Nystrom extensions reducing the computational burden under large-scale data. The first layer of the convolutional neural network uses a large-size convolutional kernel to capture the wideband features of the signal spectrum. The number of hidden layer units in the Long Short-Term Memory network is dynamically adjusted according to the dimension of the input features to ensure a balance between temporal modeling capability and computational efficiency. In the attention mechanism, the query vector is obtained by transforming the scene category label through an embedding layer, with the key vector and value vector derived from the projections of spatial and temporal features, respectively.

[0085] The feature fusion module is designed with a trade-off between computational efficiency and feature representation capability. Cross-connection between spatial and temporal features is achieved using grouped convolutions to reduce the number of parameters. Attention weights are calculated using scaled dot products to prevent the vanishing gradient problem of the softmax function. Scene category information is introduced using conditional batch normalization, with scene labels as additional input to the normalization layer. Principal component analysis is used for feature vector dimensionality reduction, preserving most of the variance information. The computational complexity of the entire feature generation process is kept within the limits of real-time processing, meeting the speed requirements of practical applications.

[0086] The network training process employs a phased optimization strategy. First, the spatial feature extraction module and the temporal feature extraction module are trained separately, and then the entire feature generation network is jointly optimized. The loss function design comprehensively considers feature discriminative ability and cross-scene stability, balancing these two optimization objectives through weight parameters. The optimization algorithm uses adaptive moment estimation, with the learning rate dynamically adjusted according to the training progress. Regularization techniques include weight decay and dropout to prevent overfitting. The training data is organized using a scene-balanced sampling strategy to ensure a relatively balanced number of samples across different scenes.

[0087] The input / output interface design of the feature generation network takes system integration requirements into account. The input data format uses a standardized spectrogram representation, including both amplitude and phase channels. The output feature vector has a fixed dimension, facilitating a unified interface design for subsequent processing modules. The network's forward propagation process employs pipelined parallel computation to improve processing throughput. Memory management utilizes a block-based caching strategy to reduce data movement overhead. The entire implementation adopts a modular design, allowing each functional component to be updated and maintained independently.

[0088] In practical deployments, the classification results of signal detection scene categories need to be updated regularly. When environmental characteristics change significantly, the spectral clustering algorithm is re-run to adjust the scene category definitions. The parameters of the cross-scene invariant feature generation network can be continuously optimized through online learning to adapt to slow changes in signal propagation characteristics. During system maintenance, the stability of the feature distribution needs to be monitored, and a model retraining process is triggered when a significant shift is detected.

[0089] Example 2: See Figure 3 The technical solution involves constructing a dual-channel residual transfer learning model and implementing a dynamic compensation mechanism for the feature alignment module. This part of the solution primarily addresses the feature distribution shift problem of UAV radio signals under different detection scenarios. Through the designed network architecture and adaptive compensation method, stable feature representation across scenarios is achieved.

[0090] In constructing the dual-channel residual transfer learning model, the source and target domain channels employ differentiated network architectures. The source domain channel is based on a deep residual network architecture, consisting of multiple stacked residual blocks. Each residual block comprises convolutional layers, normalization layers, and activation functions, using a skip connection structure to mitigate the gradient vanishing problem in deep networks. The source domain channel receives preprocessed spectral feature data as input and extracts high-level semantic features through layer-by-layer nonlinear transformations. The target domain channel uses a relatively lightweight network structure with fewer residual blocks but retains core feature extraction capabilities. A parameter-sharing mechanism exists between the two channels; some weights of the underlying convolutional kernels are set to the same value during initialization to maintain consistency in basic feature extraction.

[0091] The design of the feature transfer path is a crucial component of the model. This path connects the deep feature space of the source domain channels with the intermediate feature space of the target domain channels, achieving feature mapping through a learnable transfer weight matrix. The dimension of the transfer weight matrix is ​​automatically determined based on the feature map sizes of the two channels, and is initially set as an identity matrix to maintain the fidelity of feature transfer. During training, the transfer weight matrix, along with other network parameters, undergoes end-to-end optimization, automatically adjusting the feature mapping relationship through backpropagation. A gating mechanism is implemented on the transfer path to dynamically control the information flow based on feature relevance, preventing interference from irrelevant features.

[0092] The feature alignment module employs a combined approach of distribution difference measurement and dynamic compensation. Internally, it comprises two main components: a covariance calculation unit and a compensation coefficient generator. The covariance calculation unit performs statistical analysis on the feature distributions of the source and target domains, calculating the difference in the second moment of their respective feature spaces. This process uses a sliding window strategy, updating the covariance estimate in real-time within training batches to reflect the latest state of the feature distribution. The compensation coefficient generator produces corresponding adjustment parameters based on the covariance difference values; these parameters are converted into a transformation matrix of the feature space through a nonlinear mapping function.

[0093] The dynamic compensation process achieves domain adaptation at the feature level. Compensation coefficients are applied to the output features of the target domain channels, adjusting the center position and dispersion of the feature distribution through linear transformation. The transformed target domain features are aligned with the source domain features in the latent space, allowing subsequent classifiers to share the same decision boundary. The update frequency of the compensation parameters is synchronized with the model training cycle; the latest compensation coefficients are recalculated after each batch of forward propagation. The compensation intensity is controlled by adjustable hyperparameters to avoid overcompensation leading to feature distortion.

[0094] The domain difference loss function is constructed based on a distance metric in the kernel space. This function calculates the maximum mean difference between the feature distributions of the source and target domains in the reproducing kernel Hilbert space, serving as a constraint for model optimization. The choice of kernel function considers the nonlinear characteristics of the feature space, employing a hybrid kernel method that combines similarity metrics at different scales. Mini-batch statistics are used to calculate the loss value, balancing computational accuracy and efficiency. The domain difference loss and classification loss together constitute the multi-task learning objective, with weighting coefficients adjusting the relative importance of the two losses.

[0095] The model training process employs a phased optimization strategy. The initial phase primarily optimizes the classification performance of the source domain channels, using a large amount of labeled source domain data for supervised learning. The intermediate phase introduces target domain data, jointly optimizing the parameters of the feature transfer path and feature alignment modules. The final phase fixes the weights of the source domain channels and focuses on adjusting the adaptation capability of the target domain channels. A warm-up and decay strategy is used for learning rate settings; a smaller learning rate is used initially for stable training, and the learning rate is gradually increased later to accelerate convergence. The optimization algorithm employs an adaptive moment estimation method, automatically adjusting the learning step size of each parameter.

[0096] The detailed implementation of the network architecture considers a balance between computational efficiency and memory usage. The residual blocks employ a bottleneck structure, using 1x1 convolutions to first reduce dimensionality and then increase it, thus reducing the number of parameters. Symmetrical padding is chosen for the convolutional layers to maintain the feature map space size. The activation function uses linear units with leakage to avoid neuron death issues. Batch normalization is employed in the normalization layers to accelerate training convergence and improve the model's generalization ability. The network depth is dynamically adjusted based on the input feature dimension to ensure the receptive field covers sufficient signal context information.

[0097] The real-time processing capability of the feature alignment module is optimized through parallel computing. The covariance matrix calculation employs a block-based strategy, decomposing the large matrix into multiple sub-matrices for parallel processing. The generation process of compensation coefficients is designed as a pipeline structure, overlapping with the feature extraction process to reduce latency. The dynamic compensation operation is implemented as matrix multiplication, utilizing hardware accelerators to improve computational throughput. The memory access mode has been optimized to reuse loaded data blocks as much as possible, reducing memory bandwidth pressure.

[0098] The model deployment employs a combination of parameter freezing and dynamic updates. The parameters of the source domain channels remain fixed after initial training to ensure the stability of the basic feature extractor. The target domain channels and feature alignment module support online learning and can periodically update parameters to adapt to environmental changes. An automatic saving mechanism for feature transfer path weight settings retains historically optimal configurations for rapid rollback. The entire model's inference process is designed as a multi-threaded pipeline, with feature extraction, transfer alignment, and classification decision stages executed in parallel.

[0099] In practical applications, this implementation needs to handle changes in signal characteristics under different scenarios. Multipath effects in urban environments can broaden the signal feature distribution, requiring the compensation mechanism to enhance the sharpness of feature selection. Low-noise environments in suburban settings require the compensation algorithm to preserve subtle feature differences. Doppler shifts caused by mountainous terrain necessitate dynamic compensation to maintain the stability of frequency characteristics. The model's adaptability to these scenario differences is ensured through a large amount of diverse training data, with data augmentation techniques used to simulate various possible signal distortion scenarios.

[0100] The model maintenance and updates are automated. Trends in feature distribution are checked periodically, and model fine-tuning is triggered when significant shifts are detected. New scene data is labeled using semi-supervised learning methods, and high-confidence samples are selected from model predictions to expand the training set. Model version management employs an incremental update strategy, retaining historical versions for performance comparison and rapid rollback. The overall system performance is evaluated using multi-dimensional monitoring metrics, including feature distribution similarity, classification consistency, and computational latency.

[0101] Example 3: See Figure 4 The design and implementation of the incremental training process and decision tree integration mechanism focus on addressing the model's adaptability to new scene categories and the reliability of cross-scene detection results. During incremental training, when the system detects a new signal detection scene category, a parameter freezing strategy is employed to maintain the stability of the source domain channels, with only limited adjustments made to components related to the target domain. All convolutional and fully connected layer parameters in the source domain channels are set to a non-trainable state, preserving the basic feature extraction capabilities learned in the original scene. The parameters of the last two residual blocks in the target domain channels remain trainable, allowing the model to fine-tune high-level feature representations for new scenes. The parameters of the feature alignment module also participate in incremental training, ensuring that the distribution compensation mechanism can adapt to new domain difference patterns.

[0102] The learning rate adjustment uses the momentum optimization algorithm, whose parameter update process can be expressed as:

[0103]

[0104] in Indicates the first Model parameters at the next iteration Based on the learning rate, This is a first-moment estimate of the gradient. For the second moment estimation of the gradient, This is a small constant added to maintain numerical stability. This adaptive learning rate mechanism can automatically adjust the update step size according to the importance of the parameters, achieving differentiated adjustments for network parameters at different levels. The model convergence condition is set to ensure that the value of the validation set loss function does not decrease beyond a preset threshold for multiple consecutive training epochs, avoiding premature stopping or overtraining.

[0105] The decision tree ensemble mechanism employs a random forest architecture, comprising multiple decision trees trained in parallel. Training data for each decision tree is obtained through bootstrapping, randomly sampling with replacement from the original training set to construct a differentiated data subset. The Gini impurity criterion is used for node splitting, measuring the degree of class confounding in the current node's samples. The split point selection process considers feature randomness, evaluating only feature candidates within a random subset at each node, enhancing the diversity of the ensemble model. A maximum depth limit for the decision trees prevents overfitting, while a minimum sample size requirement for leaf nodes ensures statistical significance.

[0106] In terms of interference signal suppression, the system monitors the confidence distribution of the detection results in real time. The classification certainty is assessed by calculating the entropy value of the confidence vector; when the entropy value exceeds a dynamically adjusted threshold, a potential interference signal is identified. The interference suppression filter employs an adaptive design, with its stopband range dynamically determined based on the signal's spectral characteristics. The filtered signal undergoes re-feature extraction and classification, forming an iterative optimization loop. This process has a maximum iteration limit to ensure real-time performance while improving signal quality.

[0107] Data management during incremental training employs a hierarchical storage strategy. Original scene data is compressed and stored as feature statistics, while new scene data retains its original samples for model fine-tuning. A weighted sampling strategy for training samples balances the data distribution between old and new scenes, preventing the model from becoming overly biased towards the new category. Parameter initialization for the target domain channels utilizes a transfer learning approach from existing models, accelerating convergence on new scenes. The compensation coefficient generation network in the feature alignment module employs elastic weight fixation technology, retaining its compensation capability for the original scene while adapting to the new scene.

[0108] The probability aggregation method in the decision tree ensemble mechanism employs a soft voting strategy. Each decision tree outputs the probability distribution of a sample belonging to each category. The ensemble module then performs a weighted average of these probabilities, with the weights related to the validation set accuracy of each individual tree. The final classification result selects the category with the highest probability, while simultaneously recording the probability values ​​of each category as confidence indicators. For samples close to the decision boundary, the system triggers a verification mechanism, enhancing the reliability of the judgment by analyzing the distribution of nearest neighbor samples in the feature space.

[0109] The dynamic threshold adjustment of the interference suppression mechanism is based on historical data analysis. The system maintains a sliding window to record the entropy statistics of recent signals, and the threshold is automatically adjusted according to the mean and standard deviation within the window. The filter design parameters include stopband width, transition band slope, and stopband attenuation, which are configured according to the spectral characteristics of the interference signal. The filtering process adopts a frequency domain processing method, converting the signal to the frequency domain through a fast Fourier transform, and then applying the filter transfer function before inversely transforming it back to the time domain.

[0110] The model update mechanism employs a version control approach. Each incremental training iteration generates a new model version, while historical versions are retained for performance comparison and fault recovery. Compatibility checks between versions ensure that the new model does not break the detection capabilities of existing scenes. Model deployment follows a gradual update strategy, first testing on select nodes to verify stability before gradually rolling it out to the entire system. Runtime resource usage monitoring prevents system overload caused by model updates.

[0111] In terms of implementation details, decision tree training is accelerated using parallel methods. The feature selection process utilizes hardware accelerators to simultaneously evaluate multiple candidate split points. Tree growth employs a breadth-first strategy, managing nodes to be split through a queue. Memory usage optimization includes discretized representation of feature values ​​and a sparse storage format. The parallel execution of multiple trees in the prediction phase fully leverages the computational power of multi-core processors.

[0112] The data preprocessing stage of incremental training includes feature standardization and updating. As new scene data is added, the mean and variance of the global features are recalculated to maintain consistency across different batches of data. Data augmentation techniques are applied to small sample scenarios, expanding the training samples through appropriate signal transformations. An outlier detection mechanism filters out noisy data that may interfere with training.

[0113] The spectral analysis in the interference suppression process employs a multi-resolution approach. The window function type and window length are adaptively selected based on signal characteristics to balance the trade-off between frequency and time resolution. The filter design considers both stopband suppression and passband flatness to avoid excessive distortion of the useful signal. Real-time performance is ensured through algorithm optimization, keeping computational complexity within the hardware's processing capabilities.

[0114] The system's monitoring and maintenance functions include model performance tracking and anomaly alerts. It periodically evaluates the model's classification metrics across various scenarios and detects performance degradation trends. Resource usage is monitored in real-time, including memory consumption, computational load, and data throughput. An automatic alert mechanism notifies operations and maintenance personnel when abnormal patterns are detected, such as a sudden drop in classification accuracy or excessive processing latency. The system logs detailed operational status and decision-making processes, supporting post-event analysis and problem diagnosis.

[0115] Example 4: Specific Implementation of Interference Signal Suppression Mechanism and UAV Identity and Behavior Recognition System. This scheme achieves reliable UAV target recognition in complex electromagnetic environments through multi-level filtering and hierarchical matching strategies. The following describes its workflow and technical details with examples of typical scenarios.

[0116] In urban environmental monitoring scenarios, the system is deployed on the rooftop of high-rise buildings to monitor drone signals within a radius of approximately 1.5 kilometers. When a suspected drone radio signal is received, the interference feature filtering layer first initiates the analysis process. Taking a detected 2.4GHz band signal as an example, the system records its spectral characteristics and extracts key parameters as shown in the table below:

[0117]

[0118] The interference signal judgment logic first checks the frequency domain parity check results. The system divides the spectrum into 24 sub-bands (each sub-band 100MHz), calculates the energy correlation coefficient between odd and even sub-bands, which is 0.22, lower than the preset threshold of 0.35, triggering a preliminary interference flag. The time domain correlation unit then analyzes the mutual information value between the signal envelope and the standard UAV template, measuring a result of 0.43, which does not reach the passing threshold of 0.6. The two-stage detection results trigger the activation of the interference suppression filter, and the system automatically configures the band-stop filter parameters: center frequency 2.405GHz, stopband width 15MHz, and transition band width 5MHz.

[0119] The identification module employs a two-level matching strategy to process the filtered signal. The first level, carrier feature matching, detects a frequency deviation of +225Hz, indicating QPSK modulation, and selects three candidate device models. The second level, higher-order spectral feature matching, calculates the Euclidean distance of the bispectral diagonal slices, with the nearest neighbor matching result being a DJI Movic Air 2 series device, achieving a similarity of 0.82. The registration information parsing module decodes the device encoding field "4A-33C9-B2," where the manufacturer code "4A" corresponds to DJI Innovations, and the serial number "33C9-B2" is confirmed as valid through a verification algorithm.

[0120] The behavior recognition system synchronously analyzes temporal activity patterns. Signal strength monitoring shows a 28dB increase within 3 seconds, consistent with the abrupt strength changes characteristic of the takeoff phase. Frequency stability analysis detects rapid frequency hopping behavior in the initial phase, with five frequency switching intervals of 48ms, 52ms, 45ms, and 51ms, averaging 49ms, falling within the typical frequency hopping interval range for takeoff. After entering the cruise phase, the system records periodic power fluctuations with a main period of 1.2 seconds and a fluctuation amplitude of ±3dB. A stable +85Hz Doppler frequency shift is also observed, matching the equipment's radial movement speed.

[0121] When the system detects a potential interference signal, the multimodal verification mechanism initiates an auxiliary decision-making process. Environmental background spectrum monitoring reveals periodic impulse noise in the 2.4 GHz band, with a period length of 8 ms and a pulse width of 20 μs. Signal-environment coupling analysis calculates the ratio of the eigenvalues ​​of the covariance matrix of the current signal to the background noise to be 2.1, which is lower than the interference alarm threshold of 3.0. The system automatically retrieves monitoring data from two adjacent nodes within a 500-meter range for spatiotemporal verification: Node A reports a similar signal arrival time difference of 0.8 ms, and Node B detects a frequency offset of +79 Hz. Consistency analysis of the three-node data confirms the physical authenticity of the target signal, ultimately maintaining the original classification decision.

[0122] The dynamic parameter adjustment of the interference suppression system is based on real-time spectrum analysis. When broadband noise interference is detected, the system automatically extends the stopband to 25MHz and increases the stopband attenuation to 50dB. For narrowband single-frequency interference, a bank of notch filters is used for precise suppression, with the bandwidth of each notch filter controlled within 2MHz. The quality assessment of the filtered signal includes in-band distortion measurement and useful signal retention rate calculation to ensure that it does not affect the accuracy of subsequent feature extraction.

[0123] The device identity database is maintained using a distributed architecture. Each monitoring node stores locally common device feature templates, and the central server periodically synchronizes and updates the global database. When an unregistered device is detected, the system initiates a semi-automatic labeling process, records its radio frequency fingerprint characteristics, and submits it for manual review. Registration information includes fields such as device model, radio frequency parameters, and legal usage scope, and blockchain technology is used to ensure data immutability.

[0124] The temporal resolution of the behavior analysis is configurable, employing a 100ms analysis window for high-speed maneuvering targets and extending to a 1-second window for static behaviors such as hovering. The motion state classifier integrates multiple features: frequency change rate is used to identify acceleration / deceleration behavior, signal multipath features assist in determining altitude changes, and polarization characteristic differences detect attitude adjustments. The system establishes a reference template library of typical flight patterns, containing time-frequency feature combinations of 12 standard maneuvering patterns.

[0125] The system employs a priority scheduling strategy for resource management during runtime. Signal processing threads are divided into two priority levels: real-time processing (signal acquisition, interference detection) and background processing (feature extraction, identity matching). When computing resources are scarce, the system dynamically reduces the sampling rate of non-critical tasks to ensure the real-time performance of core recognition functions. Memory management uses a circular buffer structure, where new data overwrites the oldest expired data, maintaining a fixed memory footprint.

[0126] The anomaly handling mechanism employs a multi-level response strategy. For transient interference, the system logs the event and continues to track the target signal; for persistent interference, it initiates a spectrum scan to locate the interference source; for system-level faults, it automatically switches to degraded mode, retaining only basic detection functions. All abnormal events generate structured reports, including fields such as timestamp, event type, impact assessment, and handling recommendations.

[0127] The data visualization interface displays key node information of the signal processing flow in real time. The spectrum chart marks the identified signal components and interference locations, the signal parameter panel updates the current detection values, and the classification results display the device model and confidence level indicator. The historical data playback function allows viewing the signal evolution process along a timeline, assisting in the analysis of signal interactions in complex scenarios.

[0128] Example 5: See Figure 5 This study focuses on a multimodal interference verification mechanism, addressing the misjudgment problem of UAV signal recognition in complex electromagnetic environments through environmental background spectrum analysis and multi-source data collaborative verification. The mechanism establishes a coupling relationship model between signal and environmental characteristics, adding an environmental consistency verification dimension to traditional RF fingerprinting, thus forming a three-dimensional signal reliability assessment system.

[0129] Background spectrum acquisition is achieved using an independent receiving channel, maintaining synchronous sampling with the main signal receiving link hardware. The background spectrum monitoring range is set to 150% of the UAV's operating frequency band, with the time window aligned with the main signal analysis period, achieving microsecond-level synchronization accuracy. The acquisition system uses a high dynamic range ADC converter to preserve subtle features of the background signal, while automatic gain control is implemented to prevent saturation of strong signals. After preprocessing, the background spectrum data generates an environmental feature vector, containing key parameters such as steady-state noise floor, transient pulse statistics, and periodic interference patterns.

[0130] Signal-environment coupling analysis constructs a joint feature space, modeling the correlation between UAV signal features and environmental background features. The feature space dimensions are dynamically adjusted according to signal complexity, with typical configurations including three main categories: time-frequency joint features, modulation domain features, and environmental interference features. The analysis process employs a sliding time window strategy, with the window length automatically adjusted based on signal stability. Longer analysis windows are used for steady-state signals to improve statistical reliability, while shorter windows are used for transient signals to maintain temporal resolution. The coupling degree calculation considers the nonlinear relationships between features, estimating the joint distribution similarity using a kernel method.

[0131] The multi-source verification process employs a tiered triggering mechanism. The initial verification compares the signal arrival time difference between the current node and its neighboring nodes, calculates the theoretical propagation delay range based on the geometric positions between nodes, and verifies whether the measured values ​​conform to the laws of physical propagation. The intermediate verification analyzes the frequency observation results of multiple nodes, checks the consistency of Doppler frequency shift, and eliminates frequency measurement deviations caused by local interference. The advanced verification integrates multi-dimensional features such as power attenuation gradient and polarization characteristic changes to construct a propagation path loss model, verifying whether signal strength changes conform to the laws of spatial propagation.

[0132] Neighbor node data retrieval employs a distributed query protocol. Each monitoring node maintains a local signal database, recording historical spectral characteristics and timestamp information. The central coordinator sends data requests to relevant nodes based on the temporal and spatial characteristics of the target signal. Response data, after time alignment and frequency standard calibration, enters the consistency analysis process. The spatiotemporal verification algorithm considers clock synchronization errors and position measurement errors between nodes, setting a reasonable tolerance range to avoid misjudgments due to measurement accuracy limitations.

[0133] Dynamic interference threshold adjustment is based on historical statistical analysis of environmental characteristics. The system continuously records the variation range of various environmental parameters and establishes a statistical distribution model of background characteristics. The threshold generator dynamically calculates the critical value for interference judgment based on the current environmental type and the dispersion of historical data. For environments with frequent sudden interference, the threshold is appropriately increased to reduce false alarms; for areas with stable electromagnetic environments, the threshold is decreased to improve detection sensitivity. The threshold update cycle is positively correlated with the rate of environmental change, shortening the adjustment cycle in scenarios with drastic fluctuations in the electromagnetic environment.

[0134] The calculation of the covariance eigenvector matrix employs a block-based processing method. The UAV signal eigenvector and environmental eigenvector are divided into multiple sub-segments, and the covariance relationships between these sub-segments are calculated separately before being integrated into a complete eigenvector matrix. This approach reduces the memory requirements of large matrix operations while allowing for parallel computing to accelerate the processing flow. Matrix eigenvalue analysis is used to assess the coupling strength between the signal and the environment, with the dominant eigenvector indicating the primary interference coupling modes.

[0135] The verification results are fused using a weighted voting mechanism. Each verification dimension is assigned a weight coefficient based on its physical reliability, with time difference verification having a higher weight, followed by frequency consistency, and power attenuation characteristics serving as a supplementary reference. Voting results are categorized into three levels: confirmation, challenge, and rejection, with different handling strategies triggered for each level. For challenged results, the system initiates a signal reacquisition process, re-evaluating the signal using more refined analysis parameters; for rejection results, the current signal is directly excluded and marked as interference.

[0136] Time alignment of multi-source data employs interpolation synchronization technology. Due to slight deviations in the sampling clocks of each node, the system achieves sub-sampling precision time alignment through cubic spline interpolation based on the relative positions of signal feature points. Frequency scaling is uniformly calibrated using a reference clock to eliminate measurement errors caused by hardware local oscillator offset. Spatial location compensation considers antenna directivity and terrain obstruction effects, correcting the propagation path of the received signal strength.

[0137] The calculation of the confidence level for signal-environment coupling interference integrates multiple metrics, including the results of various methods such as feature space projection distance, distribution overlap area, and canonical correlation analysis. These results are then nonlinearly combined to generate the final confidence score. The score range is divided into three regions: high confidence, medium confidence, and low confidence, each corresponding to a different depth of verification process. High-confidence signals directly enter the recognition process, medium-confidence signals require basic verification, and low-confidence signals trigger a complete multi-source verification process.

[0138] Long-term monitoring of environmental background characteristics establishes an electromagnetic environment fingerprint database. The system periodically collects background spectral characteristics under different time periods and climatic conditions to form a spatiotemporal distribution map of environmental characteristics. When a significant change in background characteristics is detected, the system automatically matches the closest historical pattern in the environmental fingerprint database and loads the corresponding preset interference judgment parameters. This mechanism enables the system to adapt to periodic environmental fluctuations such as day-night cycles and seasonal changes.

[0139] The real-time processing workflow employs dynamic priority scheduling for resource allocation. The signal acquisition phase is given the highest priority to ensure the integrity of sampling within critical time windows; environmental analysis is a secondary priority task; and the multi-source verification process allocates computing resources based on confidence levels. Memory management utilizes a feature caching mechanism, storing frequently accessed feature data in a fast storage area and historical reference data in a secondary memory area. This hierarchical storage design balances real-time requirements with resource consumption.

[0140] A tiered response strategy is established for handling abnormal situations. For transient interference events, interference characteristics are recorded and filtering rules are updated; for persistent interference, the characteristics of the interference source are analyzed and attempts are made to locate it; for system-level anomalies, a self-check program is initiated and the system switches to safe mode. All abnormal events generate structured logs containing complete information such as environment snapshots, system status, and handling measures, supporting post-event analysis and rule optimization.

[0141] The system maintenance interface provides visualization analysis tools for environmental features. A 3D spectrogram displays the time-frequency distribution of signals and environmental features; coupling analysis results are represented by a heatmap showing correlation strength; and multi-source verification data demonstrates spatial consistency through a topology diagram. Administrators can interactively adjust analysis parameters, observe the impact of parameter changes on identification results in real time, and optimize system performance under specific environments.

[0142] The version update mechanism supports incremental learning of the environment model. When the deployment environment undergoes long-term changes, the system collects new environment feature samples and updates the environment feature library and interference judgment rules through offline training. Update packages use a differential transmission method, sending only the changed parts to reduce network load. The version rollback function retains the historical environment model, quickly restoring to a stable state when the new version encounters adaptation issues.

[0143] 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.

[0144] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for cross-scene detection of UAV radio signals based on transfer learning, characterized in that, The method includes the following steps: Collect radio signals from drones within the target area and generate raw signal spectrum data; Based on the spectral distribution characteristics reflected by the original signal spectral data, multiple signal detection scenario categories are divided; The original signal spectrum data is processed by a multi-scale feature extraction network to generate a signal feature space with cross-scene invariance. Construct a dual-channel residual transfer learning model by inputting the signal feature space into the source domain channel and the target domain channel of the dual-channel residual transfer learning model; The feature alignment module in the dual-channel residual transfer learning model is used to dynamically compensate for the feature distribution differences between the source domain scene and the target domain scene. Based on the output after feature alignment, a decision tree ensemble mechanism is used to generate cross-scene detection results; Based on the degree of matching between the cross-scene detection results and the preset signal feature library, the drone's identity and behavior type are output; The step of generating a signal feature space with cross-scene invariance includes: The original signal spectrum data is spatially compressed using a group of convolutional kernels to generate a primary feature map. Temporal dependencies in the primary feature map are extracted using a long short-term memory network. By integrating the spatial dimension compression result with the temporal dependency, a dynamic signal feature fusion module is constructed. In the signal feature dynamic fusion module, scene category weight coefficients are introduced to generate cross-scene invariant feature vectors; The steps for generating cross-scene detection results using the decision tree ensemble mechanism include: The feature-aligned output is then fed into a random forest classifier. Construct multiple decision trees using the Gini coefficient splitting criterion; Aggregate the output probability distributions of each decision tree to generate a cross-scene detection confidence vector; The final detection result is determined based on the index of the maximum value in the confidence vector.

2. The method according to claim 1, characterized in that, The steps for classifying multiple signal detection scene categories include: Extract the frequency domain energy distribution parameters, time domain pulse density parameters, and modulation type parameters from the original signal spectrum data; The frequency domain energy distribution parameters, time domain pulse density parameters, and modulation type parameters are jointly analyzed using a spectral clustering algorithm. Based on the intra-class similarity threshold and inter-class difference threshold in the clustering results, a set of signal detection scene categories including densely populated urban areas, open suburban areas, and complex mountainous areas is generated.

3. The method according to claim 1, characterized in that, The steps for constructing the dual-channel residual transfer learning model include: Set the number of deep residual network layers for the source domain channel and the target domain channel; A feature transfer path is established between the source domain channel and the target domain channel, and the feature transfer path contains a learnable transfer weight matrix; Configure the domain difference loss function in the feature alignment module, which is constructed based on the maximum mean difference criterion.

4. The method according to claim 1, characterized in that, The steps for dynamically compensating for the feature distribution differences include: Calculate the difference in covariance matrices between the feature distributions of the source domain scene and the feature distributions of the target domain scene; Generate feature distribution compensation coefficients based on the covariance matrix difference values; An adaptive weighting mechanism is used to apply the feature distribution compensation coefficients to the output features of the target domain channel; The compensation parameters of the feature alignment module are updated through gradient backpropagation.

5. The method according to claim 4, characterized in that, The method also includes an incremental training process: When a new signal detection scenario category is added, the source domain channel parameters of the dual-channel residual transfer learning model are frozen. Only update the deep residual network parameters and feature alignment module parameters of the target domain channel; The learning rate of the target domain channel is adjusted using a momentum optimization algorithm until the model convergence error is lower than a preset threshold.

6. The method according to claim 1, characterized in that, The method also includes an interference signal suppression mechanism: An interference feature filtering layer is added to the decision tree integration mechanism; Extract the entropy features of the cross-scene detection confidence vector; When the entropy value exceeds the dynamic interference threshold, the interference suppression filter is activated. The interference suppression filter performs band-stop filtering on the input signal and then regenerates the detection result.

7. The method according to claim 1, characterized in that, The steps for outputting the drone's identity and behavior type include: The cross-scene detection results are matched with a preset signal feature library using cosine similarity. When the matching similarity reaches the authentication threshold, the registration information stored in the preset signal feature library is extracted; The drone's identity is parsed based on the device code field in the registration information; Identify takeoff, cruise, and hovering behavior types by combining signal temporal activity patterns.

8. The method for cross-scene detection of UAV radio signals based on transfer learning according to claim 1, characterized in that, The method also includes a multimodal interference verification mechanism: Simultaneously acquire the environmental background spectrum characteristics of the time window corresponding to the original signal spectrum data; Construct the covariance feature matrix of UAV signal characteristics and environmental background spectrum characteristics; The confidence level of signal-environment coupling interference is calculated using the covariance feature matrix. When the confidence level of the signal-environment coupling interference reaches the interference verification threshold, the multi-source verification process is triggered. The multi-source verification process calls historical spectrum data from neighboring detection nodes to perform spatiotemporal consistency verification. The confidence weight coefficient of the cross-scene detection results is adjusted based on the verification results.

Images