Machine learning-based unmanned aerial vehicle measurement and control link signal detection method, identification method and device

Abstract

This invention discloses a machine learning-based method, identification method, and apparatus for detecting and identifying UAV telemetry and control link signals, relating to the field of signal detection, and aimed at improving applicability to complex electromagnetic backgrounds such as non-Gaussian noise environments. The invention includes: acquiring radio frequency (RF) signals; preprocessing the RF signals using wavelet transform; extracting and fusing the dimensional transformation state statistical features and signal envelope features of the preprocessed RF signals to obtain fused features, and performing UAV presence identification based on the dimensional transformation state statistical features; calculating the classification results of the fused features using a basis vector learning model and a spatial metric search model, respectively; and performing weighted fusion of the calculated classification results, using the category corresponding to the maximum fused classification probability as the identification result. This invention improves the accuracy of UAV telemetry and control link signal detection in complex non-Gaussian channel environments, exhibits better universality in non-Gaussian noise environments, and demonstrates strong robustness.

Description

Technical Field

[0001] This invention relates to the field of signal detection technology, and in particular to a method, identification method and device for detecting signals in UAV telemetry and control links based on machine learning. Background Technology

[0002] In recent years, the drone industry and its applications have developed rapidly due to their irreplaceable role in many sectors. However, incidents of unauthorized drone flights and drones carrying dangerous items have occurred frequently, posing serious threats to social security, infrastructure security, and citizens' privacy.

[0003] Unauthorized or indiscriminate flights of drones can lead to collisions with civil aircraft or interference with flight paths, seriously threatening the lives of crew members and passengers. Therefore, the detection and identification of various drone telemetry and control link signals have become particularly urgent and necessary.

[0004] Chinese patent document CN117874489A discloses a UAV signal recognition method based on the quadratic phase difference method. This method extracts features through differential analysis and clusters them, then performs a second differential analysis and clustering to identify UAV telemetry and control link signals. This method is highly sensitive to noise, which can lead to significant errors in the clustering results. Furthermore, it can only identify signals in ideal environments; its performance is affected in complex electromagnetic channels, such as Rayleigh fading channels, requiring a high signal-to-noise ratio. Additionally, due to the characteristics of low-altitude, slow-moving, and small UAVs—low flight altitude, sudden takeoff, small size, and complex non-Gaussian noise channel environments—traditional methods are prone to missed detections and false positives. Moreover, with the continuous updating of UAV types, traditional identification methods are no longer suitable for current needs.

[0005] The rapid development of deep learning technology has provided an efficient and accurate solution for signal recognition of low-altitude, slow-moving, and small unmanned aerial vehicles (UAVs). Deep learning models possess the ability to autonomously learn features, extracting high-level features from large-scale data and performing exceptionally well in UAV signal detection and recognition tasks. Using deep learning technology for UAV signal recognition has become a key research focus. Chinese patent document CN113095175A discloses a method for identifying low-altitude UAVs based on radio frequency characteristics of a data transmission station. This method extracts transient signal features and uses the KNN algorithm to identify these features. While it reduces the requirements for environmental noise compared to traditional methods, its classification weighting based on feature contribution artificially removes relatively smaller feature components, making it unable to identify more refined signal features in complex environments. Furthermore, this method is primarily based on PCA classification, requiring data standardization. In non-Gaussian distribution environments, the results obtained by the PCA method are not optimal, thus making it unsuitable for non-Gaussian channel environments. Summary of the Invention

[0006] The purpose of this invention is to provide a machine learning-based method, identification method, and apparatus for detecting and identifying UAV telemetry and control link signals, addressing all or part of the aforementioned problems, in order to improve the accuracy and robustness of UAV telemetry and control link signal detection and identification under complex electromagnetic backgrounds such as non-Gaussian noise environments.

[0007] The technical solution adopted in this invention is as follows:

[0008] A machine learning-based method for identifying UAV telemetry and control link signals includes:

[0009] Acquire radio frequency signals;

[0010] The radio frequency signal is preprocessed;

[0011] Extract and fuse the dimensional transformation state statistical features and signal envelope features of the preprocessed radio frequency signal to obtain the fused features;

[0012] The classification results of the fused features are calculated using both the basis vector learning model and the spatial domain metric search model.

[0013] The calculated classification results are weighted and fused, and the category corresponding to the maximum fused classification probability is taken as the recognition result.

[0014] In addition, this application also provides a machine learning-based UAV telemetry and control link signal identification device, which includes:

[0015] The first module is used to acquire radio frequency signals;

[0016] The second module is used for preprocessing the radio frequency signal;

[0017] The third module is used to extract and fuse the dimensional transformation state statistical features and signal envelope features of the preprocessed radio frequency signal to obtain the fused features;

[0018] The fourth module is used to calculate the classification results of the fused features using the basis vector learning model and the spatial domain metric search model, respectively.

[0019] The fifth module is used to perform weighted fusion of the calculated classification results, and the category corresponding to the maximum fused classification probability is used as the recognition result.

[0020] On the other hand, this application also provides a machine learning-based method for detecting UAV telemetry and control link signals, which includes:

[0021] Acquire radio frequency signals;

[0022] The radio frequency signal is preprocessed;

[0023] Extract the dimensional transformation state statistical features of the preprocessed radio frequency signal. If the dimensional transformation state statistical features do not reach the detection threshold, it means that a drone has been detected; otherwise, it means that no drone has been detected.

[0024] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0025] The UAV telemetry and control link signal identification and detection solution proposed in this application addresses the current limitations of methods in extracting nonlinear features, which hinders their application in complex non-Gaussian noise channel environments. It extracts a novel dimensional transformation state statistical feature to effectively reflect the nonlinear characteristics of the signal. This application also achieves efficient fusion mapping and dimensionality reduction of the signal feature space, enabling not only precise extraction of UAV telemetry and control link signal features and improved accuracy in detecting UAV telemetry and control link signals in complex non-Gaussian channel environments, but also reducing signal processing resource overhead during detection. Furthermore, since this application uses transformed state statistical features for signal detection and identification, it does not require prior knowledge of the noise distribution, exhibiting better universality and robustness in non-Gaussian noise environments. Attached Figure Description

[0026] The present invention will be described by way of example and with reference to the accompanying drawings, wherein:

[0027] Figure 1 This is a flowchart of a machine learning-based UAV telemetry and control link signal detection method provided in an embodiment of this application.

[0028] Figure 2This is a flowchart of a machine learning-based UAV telemetry and control link signal recognition method provided in an embodiment of this application.

[0029] Figure 3 This is a diagram illustrating the recognition effect of the method in one embodiment of this application.

[0030] Figure 4 This is a performance comparison chart between the method in this application and the baseline method. Detailed Implementation

[0031] All features disclosed in this specification, or all steps in all disclosed methods or processes, may be combined in any way, except for mutually exclusive features and / or steps.

[0032] Any feature disclosed in this specification (including any appended claims and abstract) may be replaced by other equivalent or similar features, unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is merely one example of a series of equivalent or similar features.

[0033] Due to the low flight altitude of UAVs and the relatively complex electromagnetic environment they operate in, the target telemetry and control (TT&C) links exhibit a wide variety of signals with characteristics such as high bandwidth and variability. For example, signal transmission systems include OFDM and FM signals, and modulation types include FSK, BPSK, QPSK, and 16QAM. Therefore, accurately extracting the key features of various UAV TT&C link signals is a significant technical challenge. To address the limitation of current UAV TT&C link signal identification methods in complex non-Gaussian channel environments, this application proposes a machine learning-based UAV TT&C link signal detection and identification method, aiming to effectively improve the accuracy and robustness of UAV TT&C link signal identification in complex non-Gaussian channel environments.

[0034] like Figure 1 As shown, the machine learning-based UAV telemetry and control link signal detection method provided in this application is used to detect the existence of UAV telemetry and control link signals. The method includes the following steps:

[0035] S1. Acquire radio frequency signals.

[0036] Since the existence of the drone is currently unknown, the collected radio frequency signals are electromagnetic signals within the target frequency band obtained by scanning the target area. The drone's telemetry and control link signal is located within the scanned frequency band. If the radio frequency signal detects target features, it indicates the presence of a drone.

[0037] S2. Preprocess the radio frequency signal.

[0038] Preprocessing of radio frequency (RF) signals primarily involves removing background noise, i.e., denoising the RF signal. This not only helps improve the signal-to-noise ratio but also reduces detection complexity and enhances recognition and detection performance. Common preprocessing methods such as filtering can be employed, and this application does not limit the specific denoising method used.

[0039] The enhanced signal data obtained after preprocessing is further used to extract features for UAV telemetry and control link signal detection and identification.

[0040] S3. Extract the dimensional transformation state statistical features of the preprocessed radio frequency signal.

[0041] Based on the preprocessing of radio frequency signals, and addressing the weakness of traditional extraction methods in extracting nonlinear features, this application calculates and extracts novel dimensional transformation state statistical features of radio frequency signals, providing feature inputs for the detection, identification, and classification of UAV telemetry and control signals.

[0042] To obtain the dimensional transformation state statistical characteristics, the radio frequency signal characteristics are mapped to a high-dimensional space for processing. Based on the signal state statistical values ​​in the high-dimensional space, the dimensional transformation state statistical characteristics effectively reflect the nonlinear characteristics of the signal. When using the dimensional transformation state statistical characteristics for UAV telemetry and control link signal detection, it is not necessary to know the noise distribution in advance. The existence of the signal can be determined only by parameters such as noise state statistics and noise standard deviation, which has better universality in non-Gaussian noise environments.

[0043] As an optional implementation method, the method for extracting the dimensional transformation state statistical features of the preprocessed radio frequency signal is as follows:

[0044] ,

[0045] by This represents the acquired radio frequency signal. In the formula, For preprocessed radio frequency signals, For dimensional transformation function, for The probability density function, The order of the state statistics for dimensional transformation. To calculate the desired function, These are the coefficient factor and adjustment factor for the state statistics of dimensional transformation, respectively. Indicates to Extracted dimensional transformation state statistical features.

[0046] Based on this, by setting detection thresholds If the statistical features of the dimensional transformation state do not reach the detection threshold, it indicates that a UAV telemetry and control link signal has been detected; otherwise, it indicates that no UAV telemetry and control link signal has been detected, thus realizing the existence detection of the UAV telemetry and control link signal. Specifically, the existence detection of the UAV is expressed as follows:

[0047] ,

[0048] In the formula, These represent the detection results for the presence of UAV telemetry and control link signals, where... This indicates that a drone telemetry and control link signal has been detected. This indicates that no UAV telemetry and control link signal exists. Detection threshold. Calculated using the following method:

[0049] ,

[0050] In the formula, Indicates the probability of a false alarm. This represents the statistical value of the noise status. Indicates the standard deviation of noise. Indicates radio frequency signal The covariance matrix.

[0051] Based on the detection of the UAV telemetry and control link signal, i.e., the detection result is Furthermore, the types of UAV telemetry and control link signals can be further identified. Based on this, this application also proposes a machine learning-based method for identifying UAV telemetry and control link signals, which includes the following steps:

[0052] S1. Acquire radio frequency signals.

[0053] S2. Preprocess the radio frequency signal.

[0054] The steps S1 and S2 described above are exactly the same as those in the previous detection method embodiment.

[0055] S3. Extract and fuse the dimensional transformation state statistical features and signal envelope features of the preprocessed radio frequency signal to obtain the fused features.

[0056] Dimensional transformation state statistical features only reflect the probabilistic characteristics of radio frequency signals, and have little reflection on the overall attribute characteristics of the signal. In contrast, signal envelope features can reflect the overall fluctuation of radio frequency signals and have good complementarity with dimensional transformation state statistical features, which can ensure the accuracy of UAV telemetry and control link signal identification.

[0057] The method for extracting the dimensional transformation state statistical features of the preprocessed radio frequency signal can also be found in the detection method embodiment above, and will not be repeated here. The following describes in detail how to extract the signal envelope features of the preprocessed radio frequency signal.

[0058] In some alternative implementations, the signal envelope features include the variance and kurtosis features of the radio frequency signal. Methods for extracting both include:

[0059] S31. Perform time-domain sampling on the preprocessed radio frequency signal.

[0060] Discrete sampling points are obtained by analyzing and sampling the time-domain waveform of the radio frequency signal at predetermined sampling intervals over a certain period of time.

[0061] S32. Calculate the mean amplitude of each sampling point as the mean amplitude of the radio frequency signal. Calculate the square of the difference between the amplitude of each sampling point and the mean amplitude, that is, calculate the difference between the amplitude of each sampling point and the mean amplitude, and calculate the square of each difference; use the mean of each square as the variance characteristic.

[0062] S33. Calculate the fourth-order central matrix of each sampling point and divide it by the square of the variance feature to obtain the kurtosis feature.

[0063] Variance and kurtosis characteristics can reflect the overall fluctuation of the radio frequency signal within the sampling period, and thus reflect the fluctuation characteristics of the radio frequency signal.

[0064] Based on the obtained dimensional transformation state statistical features and signal envelope features, the features are fused to obtain fused features.

[0065] In some alternative implementations, the method for fusing features includes:

[0066] S34. Connect the dimensional transformation state statistical features and signal envelope features in parallel and standardize them to obtain the feature matrix.

[0067] Taking the previous example of signal envelope features including variance and kurtosis features, we will discuss the statistical features of dimensional transformation states. Variance characteristics and kurtosis characteristics The feature vectors are concatenated in parallel directions to construct a feature matrix. The features in the feature matrix are then standardized to eliminate the influence of different dimensions and improve the accuracy of feature fusion.

[0068] S35. Select the feature with the largest average mutual information in the feature matrix.

[0069] During the initial feature screening, features with excessively close similarity are deduplicated to reduce feature redundancy. Specifically, for each feature in the feature matrix, the mutual information between that feature and all other features is calculated, and the average is obtained. Then, the average mutual information of each feature is sorted by value, and the K features with the highest average mutual information are selected and retained according to a predetermined number or proportion. This reduces duplicate features and improves fusion speed and accuracy.

[0070] S36. Calculate the mean distance between each selected feature and other features. Assign a first weight to the feature based on the mean distance. Use the first weight to perform a weighted summation on each selected feature to obtain a re-estimated value for the feature.

[0071] After completing the initial feature screening, for the K retained features, calculate the similarity distance between each feature and other features, and take the average value.

[0072] Suppose that at time t, the i-th feature is represented as , N represents the total number of sampling times, K represents the total number of features, and so on. The mean distance of a feature can then be calculated using the following method:

[0073] ,

[0074] in, Let represent the mean distance of the i-th feature at time t.

[0075] by This represents the re-estimated value of the j-th feature, which is obtained by weighting and summing the selected features using the first weight dynamically assigned to feature j.

[0076] Based on the calculated mean distance, the exponent of the negative mean distance is used as the first weight, giving features smaller distances to other features a higher weight. At different sampling times, the first weight assigned to each feature adaptively changes to update the feature vector, achieving dynamic estimation of the current feature while preserving the inherent non-linear structure of the data. The first weight is represented as... ,in, h This represents the state response factor.

[0077] S37. Replace the original values ​​with the re-estimated values ​​of each feature and map them to the regenerating kernel Hilbert space.

[0078] The re-estimated features are mapped to the regenerated kernel Hilbert space. A Gaussian probability density function is constructed by setting zero mean and empirical kernel width parameters. The feature values ​​are input for projection calculation, and the distance of the original features in the high-dimensional projection space is quantified based on the projection function.

[0079] S38. Calculate the difference between the re-estimated value and the mapped value of each feature, perform nonlinear operation on the difference, and complete the fusion of the state statistical features and signal envelope features of the dimension transformation.

[0080] Steps S37 and S38 are expressed in arithmetic form as follows:

[0081] ;

[0082] in, Indicates fusion features, These are the feature values ​​before fusion. This represents the feature estimates before fusion. This indicates an estimate. This represents the calculation of statistical characteristics of the state. Indicates the first j The state response function of each feature.

[0083] S4. Calculate the classification results of the fused features using both the basis vector learning model and the spatial metric search model. The classification results may include a classification vector indicating the classification type and a classification probability.

[0084] As an optional implementation, the basis vector learning model includes a preprocessing module, a model classification calculation module, and an activation output module. The preprocessing module preprocesses the input signal features, primarily normalizing them to transform them into numerical vectors within the range [0,1], thus eliminating the influence of different feature dimensions on classification. The model classification calculation module includes convolutional units, residual calculation units, and pooling units. The convolutional units perform convolutional compression on the normalized features; the residual calculation unit uses cross-layer connections to directly pass the input features to the output, adding them to the results of the convolutional units to alleviate gradient decay; the pooling unit is used to downsample the data to reduce feature size and lower the computational complexity of subsequent layers. After processing at each layer, the activation output module in the basis vector learning model calculates the classification result, which includes the classification vector output by activation and the classification probability after normalization of the classification vector.

[0085] Based on the design concept of the basis vector learning model described above, methods for calculating the classification results of fused features using the basis vector learning model include:

[0086] The fused features are input into the basis vector learning model so that the basis vector model processes the fused features as follows:

[0087] Normalization fusion characteristics;

[0088] The normalized fusion features are compressed by convolution and then connected across layers with the input fusion features;

[0089] Pooling is applied to the fusion features of the connections;

[0090] The pooled fusion features are activated to output a classification vector, which is then normalized to a classification probability.

[0091] Basis vector learning models are trained using machine learning network models. In some feasible implementations, the training methods for basis vector learning models include a pre-training phase and an advanced optimization training phase.

[0092] The pre-training phase includes:

[0093] S41. Pre-train the basis vector learning model using a subset of the radio frequency signal sample set.

[0094] In the pre-training phase, the radio frequency signal sample set (including radio frequency signal samples and their corresponding categories) is randomly sampled, and the resulting subset becomes the signal dataset for this training. Based on the method described in the previous embodiment, sample signal features are extracted and fused to obtain a sample feature dataset. This sample signal feature dataset is then input into the basis vector learning model for pre-training, and the model parameters (network parameters, etc.) are updated progressively. After training is complete, the network parameters of the machine learning network model are frozen. Following training, the classification function of the basis vector learning model is corrected by initializing scaling and offset parameters, transferring it to the basis vector learning process for a small number of sample recognition tasks.

[0095] The advanced optimization training phase includes:

[0096] S42. Using the network parameters of the pre-trained basis vector learning model as the initial network parameters, initialize the scaling and offset parameters of the basis vector learning model.

[0097] In the advanced optimization training phase, the deep learning network model adopts a two-layer loop training strategy of model-independent basis vector learning algorithm to optimize the scaling parameters, offset parameters and the initial network parameters of the new classifier, so that the deep learning network model can better adapt to the small sample signal classification task and thus obtain a higher recognition and classification accuracy.

[0098] For example, using the initial network parameters from pre-training Initialize the basis vector learning model by setting the scaling and offset parameters. The initial values ​​are set to 1 and 0 respectively.

[0099] S43. Optimize the initial network parameters, scaling parameters, and offset parameters using the remaining samples of the radio frequency signal sample set to obtain the final basis vector learning model.

[0100] In the inner loop of the few-shot recognition task, for each task, the activation function is calculated using the corresponding sample set, and the scaling and offset parameters are updated. The update method is as follows:

[0101] ;

[0102] In the formula, These are the scaling parameter and the offset parameter, respectively. Initial values ​​are set in advance before training to update the initial network parameters of the model. To update the step size; These are the initial input parameters for the network; This indicates the scaling (or offset) parameter used in the previous step of the task. This represents the scaling (or offset) parameter obtained during the update in the task. i=1 corresponds to the scaling parameter, and i=2 corresponds to the offset parameter. This indicates the task of identifying and classifying the currently learned UAV radio frequency signals (i.e., UAV telemetry and control link signals). Indicates in The losses incurred in carrying out this mission; Expressing the request The gradient; This represents the representation function of the initial network parameters and sample feature data. After updating the inner parameters in a loop, the outer loop is trained to update the model's network parameters. During the outer loop, the initial network parameters are learned based on the current basis vectors. The corresponding classification loss is calculated based on the sample feature data, and the initial network parameters are updated by gradient descent with the goal of minimizing the loss. Iteratively update parameters An optimized basis vector learning model is obtained and used for the classification and identification of radio frequency signals.

[0103] As an optional implementation, a method for calculating the classification results of fused features using a spatial domain metric search model includes:

[0104] S44. Calculate the similarity between the fusion feature and the fusion feature of each sample in the RF signal sample set, and select the sample with the highest similarity in each type of sample as the similarity node.

[0105] S45. The type of the sample with the highest confidence among all similarity nodes is used as the classification result of the fused features.

[0106] During the classification phase, the fused features are input into the spatial similarity measurement model. This model performs K-means clustering on samples in the UAV RF signal sample library (hereinafter referred to as the sample library) and allows dynamic adjustment of hyperparameters such as the neighborhood search radius and the number of nearest neighbors (k value) based on the size of the UAV RF signal sample library to balance classification accuracy and generalization ability. By calculating the Euclidean distance between the fused features and all samples in the sample library, the k closest samples are selected (the closest sample is selected from each cluster), and the type with the highest confidence score is used as the final classification result for the fused signal, with its similarity serving as the classification probability.

[0107] S5. Perform weighted fusion on the calculated classification results, and use the category corresponding to the maximum fusion classification probability as the recognition result.

[0108] After classifying the two types of models, this embodiment optimizes the classification weights of the two models based on the optimal association information criterion to reduce classification error. The classification weights are calculated by optimizing the classification weights of the two models with the goal of minimizing association information. Then, the classification features of the two models are weighted and summed using the optimized classification weights to obtain the final classification vector. The classification vector is then normalized to obtain the final classification probability.

[0109] The optimization method for classification weights is as follows:

[0110] ;

[0111] In the formula, These are the classification weights for UAV telemetry and control link signals, representing the basis vector learning model and the spatial metric search model, respectively. These are the classification probabilities of the basis vector learning model and the spatial domain metric search model, respectively. During calculation, constraints can be applied. I is the function for calculating the association information. Based on this, we utilize... Classification vectors of the basis vector learning model Classification vectors of spatial domain metric search model The weighted fusion is performed to obtain the final classification vector C, which is represented as:

[0112] .

[0113] in, It is a one-hot encoded vector, and its value is taken as follows: it is 1 only when the category position corresponding to the similar sample in the corresponding model is 1, and 0 otherwise.

[0114] After obtaining the final classification vector, the final classification vector is normalized to calculate the final classification probability. The category corresponding to the highest probability value is taken as the identification category of the radio frequency signal.

[0115] like Figure 3 The diagram shows the classification and mapping results for five types of radio frequency signals: (a), (b), (c), (d), and (e). According to... Figure 3 It can be seen that the proposed solution has a very high accuracy rate in classifying radio frequency signals.

[0116] Figure 4 The comparison shown is conducted under the same conditions to demonstrate the recognition performance of the proposed solution for radio frequency signals. In this comparison, the recognition method of this application is compared with the "recognition method based on higher-order cumulants" and the "recognition method based on cyclic spectrum". The results show that, under the same signal-to-noise ratio, especially in low signal-to-noise ratio environments, the method of this application has a significantly higher accuracy in recognition, which is sufficient to prove the applicability of the proposed method to complex nonlinear noise channel environments.

[0117] Based on the ideas of this application, this application also proposes a machine learning-based UAV telemetry and control link signal identification device, which includes:

[0118] The first module is used to acquire radio frequency signals;

[0119] The second module preprocesses the radio frequency signal;

[0120] The third module is used to extract and fuse the dimensional transformation state statistical features and signal envelope features of the preprocessed radio frequency signal to obtain the fused features;

[0121] The fourth module is used to calculate the classification results of the fused features using the basis vector learning model and the spatial domain metric search model, respectively.

[0122] The fifth module is used to perform weighted fusion of the calculated classification results, and the category corresponding to the maximum fused classification probability is used as the recognition result.

[0123] Each module in the above-mentioned identification device corresponds to each step in the identification method embodiment above. The data configured for each module can be referenced to the features designed in the corresponding steps of the identification method.

[0124] This invention is not limited to the specific embodiments described above. The invention extends to any new feature or combination disclosed in this specification, as well as any new method or process step or combination disclosed herein.

Claims

1. A machine learning-based method for identifying signals in a UAV telemetry and control link, characterized in that, include: Acquire radio frequency signals; The radio frequency signal is preprocessed; Extract and fuse the dimensional transformation state statistical features and signal envelope features of the preprocessed radio frequency signal to obtain the fused features; The method for extracting the dimensional transformation state statistical features of the preprocessed radio frequency signal is as follows: ， In the formula, For preprocessed radio frequency signals, For dimensional transformation function, Let be the probability density function. The order of the state statistics for dimensional transformation. To calculate the desired function, These are the coefficient factor and adjustment factor for the state statistics of dimensional transformation, respectively. Indicates to Extracted dimensional transformation state statistical features; A method for fusing dimensional transformation state statistical features and signal envelope features of preprocessed radio frequency signals includes: concatenating the dimensional transformation state statistical features and the signal envelope features in parallel and standardizing them to obtain a feature matrix; selecting features with the largest average mutual information in the feature matrix; calculating the mean distance between each selected feature and other features, assigning a first weight to the feature based on the mean distance, and using the first weight to perform a weighted summation on each selected feature to obtain a re-estimated value of the feature; replacing the feature matrix with the re-estimated values ​​of each feature and mapping it to a regenerating kernel Hilbert space; calculating the difference between the re-estimated value and the mapped value of each feature, and performing a nonlinear operation on the difference; The classification results of the fused features are calculated using a basis vector learning model and a spatial metric search model, respectively. The basis vector learning model includes a preprocessing module, a model classification calculation module, and an activation output module. The model classification calculation module includes a convolution unit, a residual calculation unit, and a pooling unit. The calculated classification results are weighted and fused, and the category corresponding to the maximum fused classification probability is taken as the recognition result.

2. The machine learning-based UAV telemetry and control link signal identification method as described in claim 1, characterized in that, Preprocessing the radio frequency signal includes: The radio frequency signal is then denoised.

3. The machine learning-based UAV telemetry and control link signal identification method as described in claim 1, characterized in that, Methods for extracting signal envelope features from preprocessed radio frequency signals include: Time-domain sampling is performed on the preprocessed radio frequency signal; Calculate the mean amplitude of each sampling point; calculate the square of the difference between the amplitude of each sampling point and the mean amplitude, and use the mean of the squares as the variance feature; Calculate the fourth-order centrality matrix of each sampling point and divide it by the square of the variance feature to obtain the kurtosis feature; The signal envelope features include the variance features and the kurtosis features.

4. The machine learning-based UAV telemetry and control link signal identification method as described in claim 1, characterized in that, The method for calculating the classification result of the fused features using a basis vector learning model includes: The fused features are input into the basis vector learning model, so that the basis vector model processes the fused features as follows: Normalize the fusion features; The normalized fusion features are compressed by convolution and then connected across layers with the input fusion features; Pooling is applied to the fusion features of the connections; The pooled fusion features are activated to output a classification vector, which is then normalized to a classification probability.

5. The machine learning-based UAV telemetry and control link signal identification method as described in claim 4, characterized in that, The training method for the basis vector learning model includes: The basis vector learning model is pre-trained using a subset of the radio frequency signal sample set; The scaling and offset parameters of the basis vector learning model are initialized using the network parameters of the pre-trained basis vector learning model as the initial network parameters. The initial network parameters, scaling parameters, and offset parameters are optimized using the remaining samples of the radio frequency signal sample set to obtain the final basis vector learning model.

6. The machine learning-based UAV telemetry and control link signal identification method as described in claim 1, characterized in that, The method for calculating the classification result of the fused features using a spatial domain metric search model includes: Calculate the similarity between the fusion feature and the fusion feature of each sample in the radio frequency signal sample set, and select the sample with the highest similarity in each type of sample as the similarity node; The type of the sample with the highest confidence among all similarity nodes is used as the classification result of the fused feature.

7. A machine learning-based UAV telemetry and control link signal identification device, characterized in that, include: The first module is used to acquire radio frequency signals; The second module is used for preprocessing the radio frequency signal; The third module is used to extract and fuse the dimensional transformation state statistical features and signal envelope features of the preprocessed radio frequency signal to obtain the fused features; The method for extracting the dimensional transformation state statistical features of the preprocessed radio frequency signal is as follows: ， In the formula, For preprocessed radio frequency signals, For dimensional transformation function, Let be the probability density function. The order of the state statistics for dimensional transformation. To calculate the desired function, These are the coefficient factor and adjustment factor for the state statistics of dimensional transformation, respectively. Indicates to Extracted dimensional transformation state statistical features; A method for fusing dimensional transformation state statistical features and signal envelope features of preprocessed radio frequency signals includes: concatenating the dimensional transformation state statistical features and the signal envelope features in parallel and standardizing them to obtain a feature matrix; selecting features with the largest average mutual information in the feature matrix; calculating the mean distance between each selected feature and other features, assigning a first weight to the feature based on the mean distance, and using the first weight to perform a weighted summation on each selected feature to obtain a re-estimated value of the feature; replacing the feature matrix with the re-estimated values ​​of each feature and mapping it to a regenerating kernel Hilbert space; calculating the difference between the re-estimated value and the mapped value of each feature, and performing a nonlinear operation on the difference; The fourth module is used to calculate the classification probability of the fused features using a basis vector learning model and a spatial metric search model, respectively. The basis vector learning model includes a preprocessing module, a model classification calculation module, and an activation output module. The model classification calculation module includes a convolution unit, a residual calculation unit, and a pooling unit. The fifth module is used to perform weighted fusion of the calculated classification probabilities, and the category corresponding to the maximum fusion probability is used as the recognition result.

Images