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

Abstract

The invention relates to the technical field of unmanned aerial vehicle signal detection, and discloses an unmanned aerial vehicle radio signal cross-scene detection method based on transfer learning. According to the method, unmanned aerial vehicle radio signals in a target area are collected to generate original signal spectrum data, and then a plurality of signal detection scene categories are divided according to spectrum distribution characteristics. The method comprises the following steps: processing original data through a multi-scale feature extraction network, generating a signal feature space with cross-scene invariance, and inputting the signal feature space into a source domain channel and a target domain channel of a dual-channel residual transfer learning model; a feature alignment module in the model is used for dynamically compensating feature distribution difference between a source domain scene and a target domain scene, a cross-scene detection result is generated through a decision tree integration mechanism, and finally, based on the matching degree of the result and a preset signal feature library, an unmanned aerial vehicle identity label and a behavior type are output. The method can effectively cope with signal feature differences in different scenes, and improves the adaptability and accuracy of unmanned aerial vehicle detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicle signal detection, in particular to a method for cross-scene detection of unmanned aerial vehicle radio signals based on transfer learning. BACKGROUND

[0002] With the rapid development of unmanned aerial vehicle technology, its application in aerial photography, logistics, surveying and mapping, etc. is becoming more and more widespread, but at the same time it has also brought safety hazards such as privacy leakage, so effective detection of unmanned aerial vehicles has become a problem to be solved. At present, the detection of unmanned aerial vehicles mainly relies on radio signal analysis, through capturing the communication signals, navigation signals, etc. between the unmanned aerial vehicle and the ground station, to realize the identification and tracking of the unmanned aerial vehicle. Unmanned aerial vehicle radio signals will exhibit significant differences in different scenarios. For example, there are a large number of electromagnetic interferences in urban environments, and the shielding of buildings will cause signal reflection, diffraction, resulting in distortion of the signal spectrum; while in open suburban or mountainous areas, the signal propagation path is relatively simple, but it may be affected by factors such as terrain undulations and vegetation shielding, the signal strength and spectral characteristics are quite different from those in urban scenarios. In addition, different weather conditions also affect the propagation of radio signals, further exacerbating the instability of signal characteristics. Most of the existing signal detection methods are models trained in specific scenarios, when applied to new scenarios, the change in signal feature distribution caused by scenario differences will greatly reduce the detection accuracy. Traditional machine learning methods need to re-collect a large amount of data for model training for each new scenario, which not only consumes a lot of time and human cost, but also is difficult to implement in some scenarios where data collection is difficult. Although some methods try to enhance the adaptability of the model through feature engineering, manually designed features are often difficult to cover all changes in scenarios, and have limited generalization ability. Unmanned aerial vehicles are of various types, and the radio signal frequency bands and modulation methods used by unmanned aerial vehicles of different brands and models differ, which also increases the difficulty of cross-scene detection. Existing methods are prone to confusion when dealing with signals of different types of unmanned aerial vehicles, especially in complex scenarios, it is more difficult to accurately identify the identity and behavior of the unmanned aerial vehicle after the signal characteristics are disturbed. SUMMARY

[0003] The purpose of the present application is to provide a method for cross-scene detection of unmanned aerial vehicle radio signals based on transfer learning, to solve the problems raised in the background.

[0004] To achieve the above purpose, the present application provides a method for cross-scene detection of unmanned aerial vehicle radio signals based on transfer learning, which comprises: collecting unmanned aerial vehicle radio signals in a target area to generate original signal spectrum data; Dividing a plurality of signal detection scene categories according to the spectrum distribution characteristics reflected by the original signal spectrum data; Processing the original signal spectrum data through a multi-scale feature extraction network to generate a signal feature space with cross-scene invariance; Constructing a dual-channel residual transfer learning model, and inputting the signal feature space into a source domain channel and a target domain channel of the dual-channel residual transfer learning model; Using the feature alignment module in the dual-channel residual transfer learning model, the feature distribution differences between the source domain scene and the target domain scene are dynamically compensated; Based on the output results after feature alignment, a decision tree integration mechanism is used to generate cross-scene detection results; Based on the matching degree between the cross-scene detection results and the preset signal feature library, the drone identity and behavior type are output.

[0005] Preferably, the step of dividing the signal detection scene categories includes: Extracting frequency domain energy distribution parameters, time domain pulse density parameters and modulation type parameters of the original signal spectrum data; A spectral clustering algorithm is used to jointly analyze the frequency domain energy distribution parameters, time domain pulse density parameters and modulation type parameters; According to the intra-class similarity threshold and inter-class difference threshold in the clustering results, a set of signal detection scene categories including urban dense areas, suburban open areas, and mountainous complex areas is generated.

[0006] Preferably, the step of generating a signal feature space having cross-scene invariance comprises: Using a convolution kernel group to compress the spatial dimension of the original signal spectrum data to generate a primary feature map; Extracting temporal dependencies in the primary feature map through a long short-term memory network; Fusing the spatial dimension compression result with the temporal dependency to construct a signal feature dynamic fusion module; A scene category weight coefficient is introduced into the signal feature dynamic fusion module to generate a cross-scene invariant feature vector.

[0007] Preferably, the step of constructing a dual-channel residual transfer learning model includes: Set the number of deep residual network layers for the source domain channel and the target domain channel; Establishing a feature migration path between the source domain channel and the target domain channel, wherein the feature migration path includes a learnable migration weight matrix; A domain difference loss function in the feature alignment module is configured, where the domain difference loss function is constructed based on a maximum mean difference criterion.

[0008] Preferably, the step of dynamically compensating the feature distribution difference comprises: calculating a covariance matrix difference value between the source domain scene feature distribution and the target domain scene feature distribution; generating a feature distribution compensation coefficient according to the covariance matrix difference value; applying the feature distribution compensation coefficient to the output features of the target domain channel using an adaptive weighting mechanism; updating the compensation parameters of the feature alignment module through gradient backpropagation.

[0009] Preferably, the method further comprises an incremental training process: when a new signal detection scene category is added, freezing the source domain channel parameters of the dual-channel residual transfer learning model; only updating the deep residual network parameters and feature alignment module parameters of the target domain channel; adjusting the learning rate of the target domain channel using a momentum optimization algorithm until the model convergence error is below a preset threshold.

[0010] Preferably, the step of generating a cross-scene detection result using a decision tree ensemble mechanism comprises: inputting the output results after feature alignment into a random forest classifier; constructing multiple decision trees using a Gini coefficient splitting criterion; aggregating the output probability distributions of each decision tree to generate a cross-scene detection confidence vector; determining the final detection result according to the maximum value index in the confidence vector.

[0011] Preferably, the method further comprises an interference signal suppression mechanism: adding an interference feature filtering layer to the decision tree ensemble mechanism; extracting the entropy value feature of the cross-scene detection confidence vector; when the entropy value feature exceeds a dynamic interference threshold, activating an interference suppression filter; the interference suppression filter re-generates the detection result after band-stop filtering the input signal.

[0012] Preferably, the step of outputting the UAV identity and behavior type comprises: performing cosine similarity matching between the cross-scene detection result and a preset signal feature library; when the matching similarity reaches an authentication threshold, extracting the registration information stored in the preset signal feature library; analyzing the UAV identity based on the device code field in the registration information; The signal time domain activity pattern recognition identifies the take-off, cruising and hovering behavior types.

[0013] Preferably, the method further comprises a multi-modal interference verification mechanism: Synchronously collecting the environmental background spectrum features of the time window corresponding to the original signal spectrum data; Constructing a covariance feature matrix of the unmanned aerial vehicle signal features and the environmental background spectrum features; Calculating a signal environmental coupling interference confidence degree through the covariance feature matrix; When the signal environmental coupling interference confidence degree reaches an interference verification threshold, triggering a multi-source verification process; The multi-source verification process calls the historical spectrum data of the adjacent detection nodes for spatio-temporal consistency verification; According to the verification result, the confidence weight coefficient of the cross-scene detection result is corrected.

[0014] Compared with the prior art, the present application has the following advantages: By collecting the unmanned aerial vehicle radio signals in the target area and generating original signal spectrum data, a basis material is provided for subsequent signal processing. According to the spectrum distribution features, multiple signal detection scene categories are divided, which can more clearly grasp the characteristics of signals in different scenes, making the subsequent processing more targeted.

[0015] The multi-scale feature extraction network processes the original signal spectrum data, which can mine the features of the signal at different scales, generate a signal feature space with cross-scene invariance, reduce the influence of scene changes on the core features of the signal, and make the signal features more stable, which is conducive to maintaining effective identification of the signal in different scenes. The construction of the dual-channel residual transfer learning model inputs the signal feature space into the source domain channel and the target domain channel respectively, and utilizes the advantages of the residual structure to effectively retain the key features of the signal. Meanwhile, through the idea of transfer learning, the source domain knowledge is transferred to the target domain, so that the model can play a role in the new scene without a large amount of retraining. The feature alignment module dynamically compensates for the feature distribution difference between the source domain scene and the target domain scene, which can adjust the feature difference in different scenes in real time, so that the features of the source domain and the target domain are better matched in the same space, thereby reducing the detection deviation caused by different scenes. The decision tree integration mechanism generates a cross-scene detection result according to the output results after feature alignment, integrates the advantages of multiple decision trees, can comprehensively judge from different angles, reduces the limitations of a single model, makes the detection result more reliable, and can better cope with complex and variable scenes. Based on the matching of the cross-scene detection result and the preset signal feature library, the unmanned aerial vehicle identity and behavior type are output, the detection result can be associated with known information, precise identification of the unmanned aerial vehicle is realized, and the unmanned aerial vehicle can be well adapted to scene changes in cities, suburbs or other complex environments, and the effective control ability of the unmanned aerial vehicle is improved. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The working principle diagram of the unmanned aerial vehicle radio signal cross-scene detection method based on transfer learning is described in the application. Figure 2 The flowchart for signal detection scene category division is described. Figure 3 The flowchart for dynamic compensation of the feature alignment module is described. Figure 4 The flowchart for generating detection results by the decision tree integration mechanism is described. Figure 5 The flowchart for the multi-modal interference verification mechanism is described. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only a part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0018] Please refer to Figure 1 The application provides an unmanned aerial vehicle radio signal cross-scene detection method based on transfer learning, which comprises the following steps: A wideband radio frequency receiving device is used to collect unmanned aerial vehicle radio signals in a target area, the signal sampling rate is not less than 100MHz, and original signal spectrum data containing time-frequency information is generated. After preprocessing, the original data form a standardized spectrum graph, the time resolution is set to 10ms, and the frequency resolution is adjusted to 1kHz.

[0019] The original signal spectrum data is subjected to feature analysis, the frequency energy gradient, time domain pulse interval and modulation cyclic spectrum features are extracted, and a spectrum clustering algorithm is used to divide the signal detection scene categories. During the clustering process, a dynamic similarity threshold is set, and three scene labels of city dense area, rural open area and mountain complex area are automatically generated.

[0020] A multi-scale feature extraction network is constructed, a 7*7 convolution kernel is used for spatial down-sampling in the first layer, and a bidirectional LSTM is used to extract time sequence features in the second layer. A scene category attention mechanism is introduced in the network output layer, a cross-scene invariant feature vector is generated through a trainable weight coefficient matrix, and the feature dimension is fixed to 256 dimensions.

[0021] The dual-channel residual transfer learning model is designed, the source domain channel contains 12-layer ResNet structure, and the target domain channel adopts 8-layer ResNet structure. A cross-domain feature transfer path is established between the two channels, and the transfer weight matrix is initialized as a unit matrix and dynamically updated through back propagation.

[0022] In the feature alignment module, dynamic distribution compensation is realized, the Jensen-Shannon divergence of the source domain and the target domain features is calculated as the difference measure, and the adaptive compensation coefficient is generated to act on the target domain feature mapping. The compensation coefficient update frequency is synchronized with the model training cycle.

[0023] Random forest is used as the decision tree ensemble mechanism, 50 decision trees are set as base classifiers, and the maximum depth of each tree is limited to 10 layers. The classifier input is the aligned 256-dimensional feature vector, and the output dimension is consistent with the number of UAV behavior types.

[0024] A preset signal feature library is established to store the radio frequency fingerprint features of known UAV 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 the matching is successful, the UAV ID and the current flight mode in the registration information are output.

[0025] Embodiment 1: refer to Figure 2 , signal detection scene classification and cross-scene invariance feature generation are two key links. In the signal detection scene classification stage, a multi-parameter joint analysis method is used to process the original signal spectrum data. The extraction process of frequency energy distribution parameters uses windowed short-time Fourier transform technology, and the energy of each frequency band in each time window is normalized to calculate the statistical characteristics of the energy proportion. The time domain pulse density analysis is realized by an adaptive threshold detection algorithm. First, the signal envelope is smoothed, then the rising and falling edges of the pulse are detected, and the number of pulses per unit time and the coefficient of variation of the time interval are counted. The modulation type recognition uses high-order cyclic spectrum analysis method, which extracts the cyclic frequency characteristics specific to the modulation mode by calculating the spectral correlation density function of the signal. These features can effectively distinguish different modulation types of UAV signals.

[0026] The implementation process of the spectral clustering algorithm adopts a multi-stage optimization strategy. In the feature preprocessing stage, the three types of features, including frequency domain energy, time domain pulse and modulation type, are standardized to eliminate the dimensional differences. The adaptive kernel function is used to construct the similarity matrix, and the kernel width parameter is automatically adjusted according to the density of the feature space. During the clustering process, a constraint condition is introduced, which requires the similarity of samples in the same class to exceed a preset threshold, and the difference between samples in different classes must reach the minimum separation degree. The clustering center is dynamically adjusted during the algorithm iteration process until the ratio of intra-class dispersion to inter-class distance reaches a stable state. The final output of the scene category set contains three typical environmental types, each with unique signal propagation characteristics. The characteristics of the urban dense area scene are obvious multipath effect and large signal delay spread; the characteristics of the rural open area scene are low background noise and dominant direct path; the characteristics of the mountain complex area scene are significant Doppler shift and severe signal strength fluctuation.

[0027] The generation of cross-scene invariant features is implemented using a deep neural network architecture. The spatial dimension compression module is composed of multiple convolution layers, and different sizes of convolution kernels capture local details and global structure features of the signal spectrum. Batch normalization processing is introduced after each convolution operation to accelerate network convergence and improve feature stability. The time sequence feature extraction module uses a bidirectional long short-term memory network structure to learn the time-varying dependence of the signal spectrum through a gating mechanism. The gradient clipping technique is used in the update process of the network hidden state to prevent gradient explosion during training. In the feature fusion stage, a cross-attention mechanism is designed to dynamically weight and combine spatial features and time sequence features. The generation process of attention weights considers the scene category information, and converts the scene label into a feature weight vector through a trainable mapping function.

[0028] The generation of scene category weight coefficients adopts an end-to-end training method. The weight generation network is composed of fully connected layers and nonlinear activation functions, and the input is the embedding representation of the scene category, and the output is the adjustment coefficient of each feature channel. During training, a multi-task learning strategy is used to optimize the parameters of the feature extraction network and the weight generation network simultaneously. The normalization of the feature vector uses the layer normalization method to maintain the scale consistency of each dimension feature. The finally generated cross-scene invariant feature vector has a clear physical meaning, where the low-frequency component mainly reflects the carrier characteristics of the unmanned aerial vehicle, the high-frequency component mainly represents the modulation details of the signal, and the time-varying component reflects the change trend of the flight state.

[0029] In terms of implementation details, the extraction of the frequency-domain energy distribution parameter employs an overlapping frame technique with a frame shift set to one quarter of the window length to improve the time resolution. The time-domain pulse detection algorithm employs a double-threshold strategy to effectively distinguish real signal pulses from noise fluctuations. In the modulation type identification process, the resolution of the cyclic spectrum analysis is set by considering the computational complexity and the discriminative ability of the features. The spectral clustering algorithm is implemented using an approximate solution method, which reduces the computational burden under large-scale data through Nystrom extension. The first layer of the convolutional neural network employs a large-size convolution kernel to capture the wide-band features of the signal spectrum. The number of hidden layer units of the long short-term memory network is dynamically adjusted according to the dimension of the input features to ensure the balance between the time series modeling ability and the computational efficiency. The query vector in the attention mechanism is obtained by converting the scene category label through an embedding layer, and the key vector and value vector come from the projection of spatial features and time series features, respectively.

[0030] The design of the feature fusion module considers the trade-off between computational efficiency and feature expression ability. The cross-connection of spatial features and time series features is implemented using grouped convolution to reduce the number of parameters. The calculation of attention weights uses a scaled dot product method to prevent the gradient vanishing problem of the softmax function. The introduction of scene category information uses the conditional batch normalization method, which takes the scene label as an additional input to the normalization layer. The dimension reduction of the feature vector uses the principal component analysis method to retain most of the variance information. The computational complexity of the entire feature generation process is controlled within the range allowed by real-time processing, meeting the requirements of processing speed in practical applications.

[0031] The network training process adopts a phased optimization strategy. First, the spatial feature extraction module and the time series feature extraction module are trained separately, and then the entire feature generation network is optimized jointly. The design of the loss function considers the feature discriminative ability and the cross-scene stability, balancing the two optimization objectives through a weight parameter. The optimization algorithm uses the adaptive moment estimation method, and the learning rate is dynamically adjusted according to the training progress. Regularization techniques include weight decay and dropout to prevent model overfitting. The organization of training data uses a scene-balanced sampling strategy to ensure that the number of samples for each scene is basically balanced.

[0032] The input and output interface design of the feature generation network considers the system integration requirements. The input data format uses a standardized spectrogram representation, including amplitude and phase channels. The dimension of the output feature vector is fixed, facilitating the unified interface design of subsequent processing modules. The forward propagation process of the network uses pipeline parallel computing to improve processing throughput. Memory management uses a block caching strategy to reduce data movement overhead. The entire implementation process uses modular design, and each functional component can be updated and maintained independently.

[0033] In practical deployment, the classification results of signal detection scenarios need to be updated periodically. When the environment characteristics change significantly, re-run the spectral clustering algorithm to adjust the scene category definition. The parameters of the cross-scene invariance feature generation network can be continuously optimized through online learning to adapt to the slow changes in signal propagation characteristics. During system maintenance, the stability of the feature distribution needs to be monitored, and when significant deviation is detected, the model retraining process is triggered.

[0034] Embodiment 2: see Figure 3 , the construction of a dual-channel residual transfer learning model and the dynamic compensation mechanism of the feature alignment module. This part of the technical solution mainly solves the feature distribution deviation problem of unmanned aerial vehicle radio signals in different detection scenarios. Through the designed network architecture and adaptive compensation method, stable feature expression across scenes is realized.

[0035] In the construction process of the dual-channel residual transfer learning model, the source domain channel and the target domain channel adopt differentiated network structure design. The source domain channel is based on the deep residual network architecture and contains multiple residual block stacks. Each residual block is composed of a convolution layer, a normalization layer, and an activation function, and uses a jump connection structure to alleviate the gradient vanishing problem of deep networks. The input of the source domain channel receives preprocessed spectral feature data, and extracts high-level semantic features through layer-by-layer nonlinear transformation. The target domain channel adopts a relatively lightweight network structure, with fewer residual blocks but retaining the core feature extraction capability. There is a parameter sharing mechanism between the two channels, and part of the bottom layer convolution kernel weights are set to the same value at initialization to maintain the consistency of basic feature extraction.

[0036] The design of the feature transfer path is a key component of the model. This path connects the deep feature space of the source domain channel and the intermediate feature space of the target domain channel, and realizes feature mapping through a learnable transfer weight matrix. The dimension of the transfer weight matrix is automatically determined according to the feature map size of the two channels, and the initial state is set to the identity matrix to maintain the fidelity of feature transmission. During training, the transfer weight matrix is optimized together with other network parameters in an end-to-end manner, and the feature mapping relationship is automatically adjusted through the backpropagation algorithm. A gating mechanism is set on the transfer path to dynamically control the information flow according to the feature correlation, avoiding interference from irrelevant features.

[0037] The feature alignment module adopts a combination of distribution difference measurement and dynamic compensation. The module contains two main components: a covariance calculation unit and a compensation coefficient generator. The covariance calculation unit performs statistical analysis on the feature distribution of the source domain and the target domain, and calculates the second moment difference of their respective feature spaces. This process uses a sliding window strategy to update the covariance estimate in real time within the training batch to reflect the latest state of the feature distribution. The compensation coefficient generator generates corresponding adjustment parameters based on the covariance difference value, and these parameters are converted into a transformation matrix of the feature space through a nonlinear mapping function.

[0038] The dynamic compensation process achieves domain adaptation at the feature level. The compensation coefficients act on 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 the subsequent classifier to share the same decision boundary. The update frequency of the compensation parameters is synchronized with the model training cycle, and the latest compensation coefficients are recalculated after each batch forward propagation. The compensation strength is controlled by an adjustable hyperparameter to avoid feature distortion caused by excessive compensation.

[0039] The construction of the domain discrepancy loss function is based on distance measurement in kernel space. The function calculates the maximum mean discrepancy between the source and target domain feature distributions in the reproducing kernel Hilbert space as a constraint condition for model optimization. The choice of kernel function considers the nonlinear characteristics of the feature space, and a hybrid kernel method is used to combine similarity measures of different scales. Small batch statistics are used in loss value calculation to balance the requirements of calculation accuracy and efficiency. The domain discrepancy loss and the classification loss together constitute the multi-task learning goal, and the relative importance of the two losses is adjusted by the weight coefficient.

[0040] The model training process adopts a phased optimization strategy. In the initial stage, the classification performance of the source domain channel is mainly optimized, and a large amount of labeled source domain data is used for supervised learning. In the intermediate stage, target domain data is introduced, and the parameters of the feature transfer path and the feature alignment module are jointly optimized. In the final stage, the weights of the source domain channel are fixed, and the adaptation ability of the target domain channel is concentrated. The learning rate setting uses a preheating and decay strategy, using a small learning rate in the initial stage to stabilize training, and gradually increasing the learning rate in the later stage to accelerate convergence. The optimization algorithm chooses the adaptive moment estimation method to automatically adjust the learning step of each parameter.

[0041] The details of the network architecture consider the balance between computational efficiency and memory occupancy. The bottleneck structure is used inside the residual block, which reduces the dimensionality first and then increases it through 1x1 convolution, reducing the number of parameters. The padding method of the convolution layer chooses symmetric padding to keep the spatial size of the feature map unchanged. The activation function uses the leaky linear unit to avoid the problem of neuron death. The normalization layer uses the batch normalization technique to speed up training convergence and improve model generalization. The network depth is dynamically adjusted according to the input feature dimension to ensure that the receptive field covers enough signal context information.

[0042] The real-time processing capability of the feature alignment module is optimized through parallel computing. The covariance matrix calculation adopts a block strategy, which decomposes the large matrix into multiple sub-matrices for parallel processing. The generation process of the compensation coefficient is designed as a pipeline structure, which overlaps with the feature extraction process to reduce the delay. The dynamic compensation operation is implemented as a matrix multiplication form, which uses hardware accelerators to improve the computing throughput. The memory access mode is optimized to reuse the loaded data blocks as much as possible to reduce the memory bandwidth pressure.

[0043] The model deploys using a combination of parameter freezing and dynamic updates. The parameters of the source domain channel remain fixed after initial training, ensuring the stability of the underlying feature extractor. The target domain channel and feature alignment module support online learning, enabling regular parameter updates to adapt to environmental changes. An automatic weight saving mechanism for feature migration paths preserves historically optimized configurations for rapid rollback. The entire model's inference process is designed as a multi-threaded pipeline, executing feature extraction, migration alignment, and classification decisions in parallel.

[0044] In practical applications, this implementation needs to handle variations in signal characteristics across different scenarios. Multipath effects in urban environments can cause signal feature distribution to broaden, requiring compensation mechanisms to enhance the sharpness of feature selection. Low-noise environments in suburban scenarios require compensation algorithms to preserve subtle feature differences. Doppler shifts caused by mountainous terrain require dynamic compensation to maintain frequency stability. The model's adaptability to these scenario variations is ensured by a large amount of diverse training data, using data augmentation techniques to simulate a variety of possible signal distortion scenarios.

[0045] An automated process is established for model maintenance and updates. Trends in feature distributions are regularly checked, and model fine-tuning is triggered when significant shifts are detected. New scenario data is annotated using a semi-supervised learning approach, with model predictions used to select high-confidence samples to expand the training set. Model version management utilizes an incremental update strategy, retaining historical versions for performance comparison and rapid rollback. The overall system's operational status is evaluated using multi-dimensional monitoring metrics, including feature distribution similarity, classification consistency, and computational latency.

[0046] Example 3: See Figure 4 , the design and implementation of the incremental training process and decision tree integration mechanism focus on solving the model adaptability problems of newly added scene categories and the reliability problems of cross-scene detection results. During the incremental training process, when the system detects a new signal detection scene category, a parameter freezing strategy is adopted to maintain the stability of the source domain channel, and only limited adjustments are made to the target domain related components. All convolutional layer and fully connected layer parameters of the source domain channel are set to a non-trainable state. This design retains the basic feature extraction capabilities learned in the original scene. The parameters of the last two residual blocks of the target domain channel 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 to ensure that the distribution compensation mechanism can adapt to the new domain difference pattern.

[0047] The learning rate is adjusted using the momentum optimization algorithm. The parameter update process of this algorithm can be expressed as:

[0048] in Indicates the The model parameters at the iteration, is the basic learning rate, is the first moment estimate of the gradient, is the second moment estimate of the gradient, is a small constant added to maintain numerical stability. This adaptive learning rate mechanism automatically adjusts the update step size based on parameter importance, enabling differentiated adjustments for network parameters at different layers. The model convergence condition is set to ensure that the validation set loss function value does not drop below a preset threshold over multiple consecutive training cycles, avoiding premature stopping or overtraining.

[0049] The decision tree ensemble mechanism is implemented using a random forest architecture, consisting of multiple decision trees trained in parallel. Training data for each decision tree is obtained through bootstrapping, where samples are randomly drawn with replacement from the original training set to construct differentiated data subsets. The node splitting criterion uses the Gini impurity metric, which measures the degree of class confusion within the sample at the current node. The split point selection process considers feature randomness, evaluating only feature candidates from a random subset at each node to enhance the diversity of the ensemble model. A maximum depth limit is set for the decision tree to prevent overfitting, and a minimum number of samples per leaf node is set to ensure statistical significance.

[0050] To mitigate interference signals, the system monitors the confidence distribution of detection results in real time. Classification certainty is assessed by calculating the entropy of the confidence vector. When the entropy exceeds a dynamically adjusted threshold, a potential interference signal is identified. The interference suppression filter utilizes an adaptive design, with its stopband range dynamically determined based on the signal's spectral characteristics. After filtering, the signal undergoes feature extraction and classification again, forming an iterative optimization loop. This process is capped at a maximum number of iterations to ensure real-time performance while improving signal quality.

[0051] Data management during incremental training utilizes a hierarchical storage strategy. Original scene data is compressed and stored as feature statistics, while new scene data retains the original samples for model fine-tuning. A weighted sampling strategy for training samples balances the data distribution between new and old scenes, preventing the model from being overly biased towards newly added categories. Parameter initialization for the target domain channel utilizes a migration approach from existing models to accelerate convergence on new scenes. The compensation coefficient generation network in the feature alignment module utilizes elastic weight solidification technology, adapting to new scenes while retaining the ability to compensate for existing ones.

[0052] The probability aggregation method in the decision tree ensemble mechanism adopts a soft voting strategy. Each decision tree outputs a probability distribution of the sample belonging to each class, and the ensemble module performs a weighted average of these probabilities. The weights are related to the validation set accuracy of the individual tree. The final classification result selects the class with the highest probability, and the probability values of each class are recorded as a confidence indicator. For samples close to the decision boundary, the system triggers a review mechanism to enhance the reliability of the judgment by analyzing the distribution of neighboring samples in the feature space.

[0053] 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 based on the mean and standard deviation within the window. The design parameters of the filter include the stopband width, the transition band slope, and the stopband attenuation, which are configured according to the spectral characteristics of the interference signal. The filtering process uses a frequency domain processing method, which converts the signal to the frequency domain through a fast Fourier transform, applies the filter transfer function, and then converts it back to the time domain.

[0054] The model update mechanism uses a version control method. Each incremental training produces a new model version, and historical versions are preserved for performance comparison and fault recovery. Compatibility checks between versions ensure that the new model does not compromise the detection capabilities of existing scenarios. Model deployment uses a gradual update strategy, first running in a small number of nodes to verify stability, and then gradually expanding to the entire system. Runtime resource usage monitoring prevents system overload caused by model updates.

[0055] In terms of implementation details, the training of decision trees uses a parallelization method for acceleration. The feature selection process uses a hardware accelerator to evaluate multiple candidate split points simultaneously. The tree growth uses a breadth-first strategy, managing the nodes to be split through a queue. Memory usage optimization includes discrete representation of feature values and sparse storage format. Multi-tree parallel execution in the prediction phase fully utilizes the computing power of multi-core processors.

[0056] The data preprocessing step of incremental training includes feature normalization update. As new scene data is added, the global mean and variance of features are recalculated to maintain consistency across different batches of data. Data augmentation techniques are applied to small sample scenarios to expand the training samples through reasonable signal transformations. The abnormal sample detection mechanism filters noise data that may interfere with training.

[0057] The spectral analysis in the interference suppression process uses a multi-resolution method. The window function type and window length are adaptively selected based on signal characteristics to balance the contradiction between frequency resolution and time resolution. The design of the filter takes into account both stopband suppression and passband flatness to avoid excessive distortion of useful signals. Real-time performance is achieved through optimization algorithms, keeping the computational complexity within the range of hardware processing capabilities.

[0058] The monitoring and maintenance functions of the system include model performance tracking and anomaly alerting. The classification metrics of the model under various scenarios are regularly evaluated to detect performance degradation trends. Resource usage is monitored in real-time, including memory occupancy, computational load, and data throughput. An automatic alerting mechanism notifies the operations personnel when abnormal patterns are detected, such as a sudden drop in classification accuracy or excessive processing delays. A logging system saves detailed records of the running state and decision-making process, supporting post-mortem analysis and problem diagnosis.

[0059] Embodiment 4: Implementation of interference signal suppression mechanism and UAV identity behavior recognition system. This scheme realizes reliable UAV target recognition in complex electromagnetic environments through multi-level filtering and hierarchical matching strategies. The following illustrates its workflow and technical details in combination with typical scenario examples.

[0060] In the urban environment monitoring scenario, the system is deployed on the rooftop of a high-rise building to monitor UAV signals within a radius of about 1.5 kilometers. When a suspected UAV radio signal is received, the interference feature filtering layer first starts the analysis process. Taking the detected 2.4 GHz frequency band signal as an example, the system records its spectral features and extracts the key parameters as shown in the following table:

[0061] The interference signal judgment logic first checks the frequency domain parity check result. The system divides the spectrum into 24 subbands (100 MHz per subband), and calculates the energy correlation coefficient of odd and even subbands as 0.22, which is lower than the preset threshold of 0.35, triggering the preliminary interference flag. The time domain correlation unit then analyzes the mutual information value of the signal envelope and the standard UAV template, measuring the result as 0.43, which does not reach the passing threshold of 0.6. The two-level detection results trigger the interference suppression filter to activate, and the system automatically configures the band-stop filter parameters: center frequency 2.405 GHz, stopband width 15 MHz, and transition band width 5 MHz.

[0062] The identity recognition module uses a two-level matching strategy to process the filtered signal. The first layer of carrier feature matching detects a frequency deviation of +225 Hz and a modulation type of QPSK, screening out 3 candidate device models. The second layer of high-order spectrum feature matching calculates the Euclidean distance of the bispectrum diagonal slice, and the nearest neighbor matching result is DJI MavicAir2 series devices with a matching similarity of 0.82. The registration information parsing module decodes the device code field "4A-33C9-B2", where the manufacturer code "4A" corresponds to DJI Innovation, and the serial number "33C9-B2" is confirmed valid through the verification algorithm.

[0063] The behavior recognition system synchronously analyzes temporal activity patterns. Signal strength monitoring shows a 28 dB rise within 3 seconds, consistent with the intensity mutation characteristics of the take-off phase. Frequency stability analysis detects rapid frequency hopping behavior in the initial stage, with 5 frequency switching intervals of 48 ms, 52 ms, 45 ms, 51 ms, an average of 49 ms, falling within the typical take-off phase frequency hopping interval range. After entering the cruise phase, the system records periodic power fluctuations, with a main period of 1.2 seconds and a fluctuation amplitude of ±3 dB, while observing a stable +85 Hz Doppler shift, matching the radial motion speed of the device.

[0064] When the system detects a potential interference signal, the multi-modal verification mechanism initiates an auxiliary decision-making process. Environmental background spectrum monitoring shows that there is periodic pulse noise in the 2.4 GHz frequency band, with a cycle length of 8 ms and a pulse width of 20 μs. Signal environment coupling analysis calculates the eigenvalue ratio of the covariance matrix of the current signal and background noise as 2.1, which is lower than the interference alarm threshold of 3.0. The system automatically calls the monitoring data of two adjacent nodes within a range of 500 meters for spatio-temporal 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 three-node data confirms the physical authenticity of the target signal, and finally maintains the original classification decision.

[0065] The dynamic parameter adjustment of the interference suppression system is based on real-time spectrum analysis. When detecting wideband noise interference, the system automatically expands the stopband width to 25 MHz, and increases the stopband attenuation to 50 dB. For narrowband single-frequency interference, a notch filter bank is used for precise suppression, with each notch filter bandwidth controlled within 2 MHz. Signal quality evaluation after filtering includes in-band distortion measurement and useful signal retention rate calculation, ensuring that the accuracy of subsequent feature extraction is not affected.

[0066] The maintenance of the device identity library adopts a distributed architecture. Each monitoring node stores local common device feature templates, and the central server regularly synchronizes and updates the global database. When detecting an unregistered device, the system initiates a semi-automatic labeling process, records its radio frequency fingerprint features, and submits them for manual review. The registration information includes device model, radio frequency parameters, legal use range, etc., and uses blockchain technology to ensure data tamper resistance.

[0067] The time resolution of behavior analysis is configurable, with a 100 ms analysis window for high-speed maneuvering targets and a 1 second window for hovering and other static behaviors. The motion state classifier integrates multiple features: frequency change rate for identifying acceleration / deceleration behavior, signal multipath feature for assisting height change judgment, and polarization characteristic difference for detecting attitude adjustment. The system establishes a reference template library of typical flight modes, including time-frequency feature combinations of 12 standard maneuvering modes.

[0068] The resource management of the system during operation adopts a priority scheduling strategy. The signal processing threads are divided into real-time processing (signal acquisition, interference detection) and background processing (feature extraction, identity matching) two priorities. When the computing resources are tight, the system dynamically reduces the sampling rate of non-critical tasks to ensure the real-time performance of the core recognition function. The memory management adopts a ring buffer structure, and the newest data covers the oldest expired data, maintaining a fixed memory occupation.

[0069] The abnormal processing mechanism is designed with a multi-level response strategy. For transient interference, the system records event logs and continues to track the target signal; for persistent interference, it starts spectrum scanning to locate the interference source position; for system-level faults, it automatically switches to a degraded mode of operation, retaining only basic detection functions. All abnormal events generate structured reports, including time stamp, event type, impact assessment, and disposal recommendations.

[0070] The data visualization interface displays the key node information of the signal processing flow in real time. The spectrum is labeled with the identified signal components and interference positions, the signal parameter panel updates the current detection values, and the classification results display the device model and confidence indicator. The historical data playback function supports viewing the signal evolution process on the time axis, assisting in analyzing the signal interaction in complex scenarios.

[0071] Example 5: see Figure 5 , around the multi-modal interference verification mechanism, through environmental background spectrum analysis and multi-source data collaborative verification, solves the misjudgment problem of unmanned aerial vehicle signal recognition in complex electromagnetic environment. This mechanism establishes a coupling relationship model between signal and environmental characteristics, adds an environmental consistency verification dimension based on traditional radio frequency fingerprint recognition, forming a three-dimensional signal credibility evaluation system.

[0072] The environmental background spectrum acquisition uses an independent receiving channel to realize synchronization sampling with the main signal receiving link hardware. The background spectrum monitoring range is set to 150% of the bandwidth of the unmanned aerial vehicle working frequency band, and the time window is aligned with the main signal analysis period, with microsecond-level synchronization accuracy. The acquisition system uses a high dynamic range ADC converter to preserve the subtle features of the background signal, while setting up an automatic gain control to prevent strong signal saturation. After preprocessing, the background spectrum data generates an environmental feature vector, including key parameters such as steady-state noise base, transient pulse statistics, and periodic interference patterns.

[0073] Signal environment coupling analysis constructs a joint feature space to correlate the UAV signal features with the environmental background features. The feature space dimension is dynamically adjusted according to the signal complexity, and a typical configuration includes three categories: time-frequency joint features, modulation domain features, and environmental interference features. The analysis process uses a sliding time window strategy, and the window length is automatically adjusted according to the signal stability. For steady-state signals, a longer analysis window is used to improve statistical reliability, and for transient signals, the window length is shortened to maintain time resolution. The coupling degree calculation considers the nonlinear relationship between features, and estimates the joint distribution similarity through kernel methods.

[0074] The multi-source verification process designs a hierarchical triggering mechanism. The primary verification compares the time difference of signal arrival between the current node and the adjacent nodes, calculates the theoretical propagation delay range based on the geometric position between nodes, and checks whether the measured value conforms to the physical propagation law. The intermediate verification analyzes the frequency observation results of multiple nodes, checks the consistency of Doppler frequency shift, and excludes the frequency measurement deviation caused by local interference. The advanced verification integrates multiple-dimensional features such as power attenuation gradient and polarization characteristic change to construct a propagation path loss model, and verifies whether the signal strength change conforms to the spatial propagation law.

[0075] The adjacent node data call uses a distributed query protocol. Each monitoring node maintains a local signal database, recording historical spectral features and timestamp information. The central coordinator sends data requests to related nodes based on the time and spatial features of the target signal. The response data is time-aligned and frequency-calibrated before entering the consistency analysis process. The spatio-temporal verification algorithm considers the clock synchronization error and position measurement error between nodes, sets a reasonable tolerance range, and avoids false positives due to measurement accuracy limitations.

[0076] The dynamic interference threshold adjustment is based on historical statistical analysis of environmental features. The system continuously records the variation range of various environmental parameters to establish a statistical distribution model of background features. The threshold generator dynamically calculates the critical value of 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 positives; for areas with stable electromagnetic environment, the threshold is reduced to improve detection sensitivity. The threshold update period is positively related to the environmental change rate, and is shortened in scenes with severe electromagnetic environment fluctuations.

[0077] The covariance feature matrix calculation uses a block processing method. The UAV signal feature vector and the environmental feature vector are divided into multiple segments, and the covariance relationship between the segments is calculated, and then integrated into a complete feature matrix. This processing method reduces the memory requirements of large matrix operations and allows parallel computing to speed up the processing process. Matrix eigenvalue analysis is used to evaluate the coupling strength between the signal and the environment, and the dominant eigenvector indicates the main interference coupling mode.

[0078] The verification result fusion adopts a weighted voting mechanism. Each verification dimension is assigned a weight coefficient based on its physical reliability. The time difference verification has a higher weight, followed by frequency consistency, and power decay characteristics as a supplementary reference. The voting results are divided into three levels: confirmation, questioning, and denial. Different levels trigger different handling strategies. For questioning results, the system starts a signal recapture process and reevaluates using more refined analysis parameters. For denial results, the current signal is directly excluded and marked as interference.

[0079] The time alignment of multi-source data uses interpolation synchronization technology. Due to the slight deviation of the sampling clock of each node, the system achieves sub-sampling accuracy time alignment through cubic spline interpolation based on the relative position of signal feature points. The frequency scale is unified using reference clock calibration to eliminate measurement errors caused by hardware local oscillator offset. Spatial position compensation considers antenna directivity and terrain shielding effects to correct the propagation path of received signal strength.

[0080] The calculation of signal environment coupling interference confidence integrates various measurement indicators. The results of various methods such as feature space projection distance, distribution overlap area, and canonical correlation analysis are combined nonlinearly to generate the final confidence score. The score interval is divided into high confidence, medium confidence, and low confidence, corresponding to different verification process depths. High-confidence signals directly enter the identification process, medium-confidence signals require basic verification, and low-confidence signals trigger a complete multi-source verification process.

[0081] Long-term monitoring of environmental background features establishes an electromagnetic environment fingerprint library. The system regularly collects background spectral features under different time periods and different climate conditions to form a spatiotemporal distribution map of environmental features. When significant changes in background features are detected, the closest historical pattern in the environment fingerprint library is automatically matched, and the corresponding interference judgment parameter preset is loaded. This mechanism enables the system to adapt to periodic environmental fluctuations such as day-night alternation and seasonal changes.

[0082] The resource allocation of the real-time processing flow uses dynamic priority scheduling. The signal capture stage is given the highest priority to ensure the integrity of the sampling in the critical time window. Environmental analysis is a secondary priority task. The multi-source verification process allocates computing resources based on confidence levels. Memory management uses a feature caching mechanism. Frequently accessed feature data is retained in the fast storage area, and historical reference data is stored in the secondary memory area. This hierarchical storage design balances real-time requirements and resource consumption.

[0083] The abnormal situation handling establishes a hierarchical response strategy. For transient disturbance events, the disturbance characteristics are recorded and the filtering rules are updated; for persistent disturbances, the disturbance source characteristics are analyzed and localization is attempted; for system-level anomalies, self-check procedures are initiated and the system is switched to safe mode operation. All abnormal events generate structured logs containing complete information such as environmental snapshots, system states and handling measures, supporting post-mortem analysis and rule optimization.

[0084] The system maintenance interface provides visual analysis tools for environmental characteristics. The three-dimensional spectrum diagram displays the time-frequency distribution of signals and environmental characteristics, the coupling analysis results are represented by a heat map indicating the correlation strength, and the multi-source verification data are displayed through a topology diagram to show spatial consistency. The administrator can interactively adjust the analysis parameters, observe the impact of parameter changes on the identification results in real time, and optimize the system's performance in a specific environment.

[0085] The version update mechanism supports incremental learning of environmental models. When long-term changes occur in the deployment environment, the system collects new environmental characteristic samples and updates the environmental characteristic library and disturbance judgment rules through offline training. The update package uses a differential transmission method, sending only the changed part to reduce network load. The version rollback function preserves historical environmental models, allowing quick recovery to a stable state when the new version has adaptability problems.

[0086] It should be noted that, in the present document, relational terms such as first and second and the like can be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0087] While embodiments of the present application have been shown and described with reference to particular embodiments thereof, it will be understood by those skilled in the art that various changes in form and details can be made therein without departing from the spirit and scope of the application as defined by the appended claims and their equivalents.

Claims

1. A cross-scene detection method for UAV radio signals based on transfer learning, characterized in that: The method comprises the following steps: Collect UAV radio signals in the target area and generate raw signal spectrum data; Dividing a plurality of signal detection scene categories according to the spectrum distribution characteristics reflected by the original signal spectrum data; Processing the original signal spectrum data through a multi-scale feature extraction network to generate a signal feature space with cross-scene invariance; Constructing a dual-channel residual transfer learning model, and inputting the signal feature space into a source domain channel and a target domain channel of the dual-channel residual transfer learning model; Using the feature alignment module in the dual-channel residual transfer learning model, the feature distribution differences between the source domain scene and the target domain scene are dynamically compensated; Based on the output results after feature alignment, a decision tree integration mechanism is used to generate cross-scene detection results; Based on the matching degree between the cross-scene detection results and the preset signal feature library, the drone identity and behavior type are output.

2. The method according to claim 1, characterized in that The step of dividing the signal detection scene into multiple categories includes: Extracting frequency domain energy distribution parameters, time domain pulse density parameters and modulation type parameters of the original signal spectrum data; A spectral clustering algorithm is used to jointly analyze the frequency domain energy distribution parameters, time domain pulse density parameters and modulation type parameters; According to the intra-class similarity threshold and inter-class difference threshold in the clustering results, a set of signal detection scene categories including urban dense areas, suburban open areas, and mountainous complex areas is generated.

3. The method according to claim 1, characterized in that The step of generating a signal feature space with cross-scene invariance comprises: Using a convolution kernel group to compress the spatial dimension of the original signal spectrum data to generate a primary feature map; Extracting temporal dependencies in the primary feature map through a long short-term memory network; Fusing the spatial dimension compression result with the temporal dependency to construct a signal feature dynamic fusion module; A scene category weight coefficient is introduced into the signal feature dynamic fusion module to generate a cross-scene invariant feature vector.

4. The method according to claim 1, wherein The steps of constructing a 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; Establishing a feature migration path between the source domain channel and the target domain channel, wherein the feature migration path includes a learnable migration weight matrix; A domain difference loss function in the feature alignment module is configured, where the domain difference loss function is constructed based on a maximum mean difference criterion.

5. The method according to claim 1, wherein The step of dynamically compensating the characteristic distribution difference comprises: Calculate the difference value of the covariance matrix between the source domain scene feature distribution and the target domain scene feature distribution; generating a characteristic distribution compensation coefficient according to the covariance matrix difference value; Adopting an adaptive weighting mechanism to apply the feature distribution compensation coefficient to the output features of the target domain channel; The compensation parameters of the feature alignment module are updated through gradient back propagation.

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

7. The method according to claim 1, characterized in that The step of using the decision tree integration mechanism to generate cross-scenario detection results includes: The output of feature alignment is input into the random forest classifier; Construct multiple decision trees using the Gini coefficient splitting criterion; Aggregate the output probability distribution of each decision tree to generate a cross-scene detection confidence vector; The final detection result is determined according to the maximum value index in the confidence vector.

8. The method according to claim 7, characterized in that The method also includes an interference signal suppression mechanism: Adding an interference feature filtering layer to the decision tree integration mechanism; Extracting entropy features of the cross-scene detection confidence vector; When the entropy value characteristic exceeds a dynamic interference threshold, activating an interference suppression filter; The interference suppression filter performs band-stop filtering on the input signal and then regenerates the detection result.

9. The method according to claim 1, characterized in that The step of outputting the drone identity and behavior type includes: Perform cosine similarity matching on the cross-scene detection results and the preset signal feature library; When the matching similarity reaches the authentication threshold, extracting the registration information stored in the preset signal feature library; Parsing the drone identity based on the device code field in the registration information; Combine the signal time domain activity pattern to identify the takeoff, cruise and hovering behavior types.

10. The cross-scene detection method of UAV radio signals based on transfer learning according to claim 1 is characterized in that: The method also includes a multimodal interference verification mechanism: Synchronously collecting the environmental background spectrum characteristics of the time window corresponding to the original signal spectrum data; Construct the covariance feature matrix of the UAV signal characteristics and the environmental background spectrum characteristics; Calculating the signal environment coupling interference confidence by using the covariance characteristic matrix; When the signal environment coupled interference confidence reaches the interference verification threshold, triggering the multi-source verification process; The multi-source verification process calls the historical spectrum data of the neighboring detection nodes to perform spatiotemporal consistency verification; The confidence weight coefficient of the cross-scene detection result is modified according to the verification result.

Images