Unmanned aerial vehicle signal intelligent identification method based on time-frequency multichannel characteristics

Abstract

The invention discloses an unmanned aerial vehicle signal intelligent identification method based on time-frequency multichannel characteristics, and belongs to the technical field of unmanned aerial vehicle signal identification and signal processing. Aiming at the problems of weak anti-noise interference capability, low identification precision and poor adaptability of unmanned aerial vehicle identification in a complex electromagnetic environment, the method comprises the following steps: receiving radio-frequency signals in different frequency bands, eliminating direct-current components and low-frequency drift through continuous detrending denoising, obtaining time-frequency information by using discrete Fourier transform, and splicing frequency spectrums; and performing multi-feature channel processing along a frequency domain to construct energy spectrum data, tensing the data into a three-dimensional feature matrix, and inputting the three-dimensional feature matrix into a multi-channel two-dimensional convolutional neural network model to perform hierarchical feature extraction and classification identification. According to the method, rapid and accurate identification of the type and the flight mode of the unmanned aerial vehicle is realized, the robustness and generalization in an interference environment are improved, and the method is suitable for low-altitude airspace safety and electromagnetic spectrum management.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of unmanned aerial vehicle signal recognition and signal processing, and particularly relates to an unmanned aerial vehicle signal intelligent recognition method based on time-frequency multi-channel features. BACKGROUND

[0002] Unmanned aerial vehicle signal recognition is a method for realizing efficient recognition of airspace violation unmanned aerial vehicle targets by intelligently processing received signals. Compared with traditional radar detection, this method has the advantages of low environmental dependence and low cost, and has important application value in maintaining electromagnetic spectrum order and ensuring low-altitude airspace safety. The main technical means for unmanned aerial vehicle recognition currently includes detection methods based on vision, acoustics, radar and radio frequency. Vision detection technology captures unmanned aerial vehicle images through cameras or infrared imaging devices, and identifies the shape features or motion trajectories of unmanned aerial vehicles using image processing algorithms, but the performance is significantly reduced under high-speed moving targets and complex weather interference. Acoustic detection uses a high-sensitivity microphone array to collect the sound signals generated by the propellers or motors of unmanned aerial vehicles, and identifies unmanned aerial vehicles through audio features, but environmental noise such as wind noise can seriously affect the signal-to-noise ratio, and effective signals are easily filtered out by active noise reduction technology. Radar detects unmanned aerial vehicles by emitting electromagnetic waves and receiving reflected signals, and analyzes the time-frequency characteristics of the echoes, but has problems such as large background clutter interference, small reflection cross section and high cost, which limit its application range.

[0003] With the development of deep learning methods, radio frequency recognition fusion intelligent algorithms provide a new research direction for unmanned aerial vehicle signal intelligent recognition. This method listens to the communication link signals between unmanned aerial vehicles and remote controllers, analyzes the spectrum patterns or data flow characteristics to achieve recognition. However, under complex electromagnetic interference and noise background, existing radio frequency recognition methods still face technical bottlenecks. Traditional signal processing methods have deficiencies in feature extraction, making it difficult to fully represent the subtle features of unmanned aerial vehicle signals. Existing recognition models have poor adaptability to frequency domain space heterogeneity features, resulting in insufficient generalization ability when encountering new interference patterns or unmanned aerial vehicle models. At the same time, there is a lack of systematic solutions for actual electromagnetic interference scenarios. In complex environments with comb-shaped spectrum, blocking and other human interference, the robustness of the recognition model is difficult to guarantee. These defects seriously affect the actual deployment effect of unmanned aerial vehicle recognition technology in military and civilian applications. SUMMARY

[0004] To solve the above technical problems, the present application provides an unmanned aerial vehicle signal intelligent recognition method based on time-frequency multi-channel features, which realizes fast and accurate recognition of unmanned aerial vehicle types and flight modes, improves robustness and generalization in interference environments, and is suitable for low-altitude airspace safety and electromagnetic spectrum management.

[0005] To achieve the above purpose, the present application provides an unmanned aerial vehicle signal intelligent recognition method based on time-frequency multi-channel features, which comprises: acquire a radio frequency signal of the unmanned aerial vehicle; pre-process the radio frequency signal to obtain pre-processed spectrum data, the pre-processing including detrended denoising, discrete Fourier transform spectrum analysis, and multi-feature channel processing; tensorize the pre-processed spectrum data to construct a three-dimensional feature matrix; input the three-dimensional feature matrix into a multi-channel two-dimensional convolutional neural network model to perform feature extraction and classification recognition, and obtain an identification result of the type and flight mode of the unmanned aerial vehicle.

[0006] Optionally, acquiring the radio frequency signal of the unmanned aerial vehicle includes: using a software radio device to perform dual-channel reception to collect high and low frequency band radio frequency signals; collecting a wide frequency range and recording a noise floor signal; collecting signals for multiple standard operating modes of the unmanned aerial vehicle, including static, autonomous hovering, video-free flight, and video recording flight.

[0007] Optionally, the process of detrended denoising includes: converting the original radio frequency signal sequence into a random walk process; dividing the integrated sequence into equal-length subintervals and segmenting from the beginning and end of the sequence; in each subinterval, calculating the local trend by local quadratic polynomial fitting; subtracting the local trend from the original integrated sequence to obtain the fluctuation function; calculating the root mean square value of the signal sequence after detrending in each subinterval as the signal amplitude fluctuation intensity.

[0008] Optionally, the process of discrete Fourier transform spectrum analysis includes: dividing the low frequency and high frequency signal amplitude intensity after detrending into equal-length subintervals; performing discrete Fourier transform on the signals in each interval to obtain single-sided spectrum data for the low frequency band and the high frequency band; splicing the spectrum data of the low frequency band and the high frequency band into a complete spectrum; calculating the energy ratio of the last sample of the low frequency band to the first sample of the high frequency band, and normalizing the amplitude spectrum to energy spectrum.

[0009] Optionally, the process of multi-feature channel processing includes: uniformly dividing the complete bandwidth spectrum vector into multiple continuous subchannel frequency bands along the frequency domain dimension; each subchannel contains an equal number of sampling points; combining the multi-channel feature spectrum data into a matrix, where each row of the matrix corresponds to one channel data; Each feature channel has a specific frequency range, which is defined by a channel start value and an end value.

[0010] Optionally, the tensorization process includes: Reshaping the multi-channel feature spectrum data matrix into a three-dimensional feature matrix; The three dimensions of the three-dimensional feature matrix respectively represent energy, frequency, and channel; The frequency dimension captures the neighborhood aggregation mechanism of the frequency point sequence, the channel dimension maps the difference distribution of the physical characteristics of different frequency band signals, and the energy dimension is fixed as a single value to adapt to the two-dimensional convolution kernel sliding.

[0011] Optionally, the multi-channel two-dimensional convolutional neural network model includes a primary convolutional layer, a maximum pooling layer, a deep convolutional layer, a second maximum pooling layer, a fully connected layer, and a result output layer; The primary convolutional layer is configured to slide the convolution kernel along the frequency dimension to realize feature fusion across the channel dimension; The maximum pooling layer performs down-sampling on the feature map to compress the frequency dimension; The deep convolutional layer uses more convolution kernels to extract high-order frequency domain features; The fully connected layer flattens the feature map and constructs a non-linear mapping through a hidden layer, and the output layer uses a Softmax classifier to output class probability.

[0012] Optionally, the training process of the multi-channel two-dimensional convolutional neural network model includes: Adopting an adaptive matrix estimation optimizer to update the model parameters; Using a cross-entropy loss function as the optimization objective, the loss function is calculated based on the true label and the predicted class probability of the sample; The training process sets multiple gradient noise intensities for scenes with electromagnetic space noise and human interference, and performs performance evaluation through multi-model comparison and multi-interference scene comparison.

[0013] The technical effect of the present application: the present application discloses a method for intelligent recognition of unmanned aerial vehicle signals based on time-frequency multi-channel features, which eliminates the direct current component and low-frequency drift in the radio frequency signal by using a continuous detrended denoising method, obtains time-frequency information by using discrete Fourier transform, and constructs energy spectrum data by spectrum splicing and amplitude normalization, processes multiple feature channels along the frequency domain for spatial heterogeneity, realizes spatial decoupling and structured reorganization of frequency domain features by combining a multi-channel two-dimensional convolutional neural network model, realizes hierarchical extraction and differential modeling of cross-band features while ensuring the feature topology adjacency relationship, thereby improving the accuracy and robustness of unmanned aerial vehicle signal recognition under complex electromagnetic interference and noise background, realizing rapid and accurate identification of unmanned aerial vehicle types and flight modes, enhancing the generalization ability of the model under electromagnetic space noise and human interference such as comb spectrum and blocking scene, and being suitable for low-altitude airspace safety and electromagnetic spectrum management. BRIEF DESCRIPTION OF DRAWINGS

[0014] The drawings constituting a part of the present application are used to provide a further understanding of the present application, the illustrative embodiments of the present application and the description thereof are used to explain the present application, and do not constitute an improper limitation on the present application. In the drawings: Figure 1 The flowchart of the method for intelligent recognition of unmanned aerial vehicle signals based on time-frequency multi-channel features of the embodiments of the present application. DETAILED DESCRIPTION

[0015] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the drawings and in combination with the embodiments.

[0016] It should be noted that the steps shown in the flowchart of the drawings can be executed in a computer system such as a group of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in an order different from that shown herein.

[0017] As Figure 1 shown, the method for intelligent recognition of unmanned aerial vehicle signals based on time-frequency multi-channel features in the present embodiment includes: obtaining a radio frequency signal of an unmanned aerial vehicle; preprocessing the radio frequency signal to obtain preprocessed spectrum data, the preprocessing including detrended denoising, discrete Fourier transform spectrum analysis and multi-feature channel processing; tensorizing the preprocessed spectrum data to construct a three-dimensional feature matrix; inputting the three-dimensional feature matrix into a multi-channel two-dimensional convolutional neural network model to perform feature extraction and classification recognition, and obtaining the recognition result of the unmanned aerial vehicle type and flight mode.

[0018] Further, the radio frequency signal of the unmanned aerial vehicle is acquired, comprising: Adopting a software radio device for double-channel reception, high and low frequency band radio frequency signals are collected; Collecting a wide frequency range and recording the background noise signal; Signal collection is performed on multiple standard operating modes of the unmanned aerial vehicle, including static, autonomous hovering, video-free flight, and video recording flight.

[0019] Specifically, the implementation process of the embodiment includes: Unmanned aerial vehicle radio frequency signal wide band coverage reception: through the NI-USRP-2943R software defined radio device, double-channel reception mode is adopted to collect radio frequency signals (high and low frequency band), the maximum instantaneous bandwidth is 40MHz, the frequency range covered is 1.2MHz-6GHz, the frequency step precision is 1kHz, the device configuration dynamic gain range is (-10dB to 50dB), the IQ sampling rate is 200MS / s, and the ADC resolution is 16 bits. The background noise signal of the radio frequency is recorded during the collection process, and then four standard operating modes of the unmanned aerial vehicle (including static, autonomous hovering, video-free flight, and video recording flight) are collected. Data collection is realized through a real-time radio frequency signal capture program developed based on LabVIEW.

[0020] Further, the detrended denoising process includes: Convert the original radio frequency signal sequence into a random walk process; Divide the integrated sequence into equal-length subintervals, and segment from the beginning and end of the sequence respectively; In each subinterval, the local trend is calculated by local quadratic polynomial fitting; Subtract the local trend from the original integrated sequence to obtain the fluctuation function; Calculate the root mean square value of the signal sequence after detrending in each subinterval as the signal amplitude fluctuation intensity.

[0021] Specifically, the implementation process of the embodiment includes: For the direct current component and low frequency drift of the radio frequency signal of the unmanned aerial vehicle and other frequency domain interference, a detrended frequency domain signal denoising method, a frequency spectrum analysis method based on discrete Fourier transform, and an unmanned aerial vehicle frequency spectrum data multi-feature channel processing method are proposed. The specific description is as follows: For the non-stationary and trend-containing characteristics of the radio frequency signal collected by the unmanned aerial vehicle, in order to eliminate the direct current component and low frequency drift in the signal and avoid interference with subsequent frequency domain analysis, continuous detrending is introduced to extract the fluctuation characteristics of the unmanned aerial vehicle signal by segment detrending. By quantifying the fluctuation intensity of the signal after removing the local trend, the intrinsic characteristics of the signal are better presented. First, the original signal sequence is converted into a random walk process , eliminating the mean and reducing the effect of non-stationarity, which is expressed as: (1); The integrated sequence is divided into equal-length sub-intervals. In order to make full use of data, segmentation is performed at the beginning and end of the sequence respectively. In each signal sub-interval, the local trend is calculated by local quadratic polynomial fitting. Then the local trend is subtracted from the original integrated sequence to obtain the corresponding fluctuation function : (2); In equation (2), is the sub-interval window size, is the window index of signal segmentation. Further, the root mean square value of each signal sub-interval after detrending is calculated by the fluctuation function, which is the signal amplitude fluctuation intensity of the interval and is expressed as: (3); In equation (3), represents the signal amplitude intensity, represents the total number of signal segment samples.

[0022] Further, the process of discrete Fourier transform spectrum analysis includes: The low-frequency and high-frequency signal amplitude intensity after detrending is divided into equal-length sub-intervals; Discrete Fourier transform is performed on the signals in each interval to obtain single-sided spectrum data of the low-frequency and high-frequency bands; The spectrum data of the low-frequency and high-frequency bands are spliced into a complete spectrum; The energy ratio of the last sample of the low-frequency band to the first sample of the high-frequency band is calculated, and the amplitude spectrum is converted into an energy spectrum by normalization.

[0023] Specifically, the implementation process of the embodiment includes: The low-frequency and high-frequency signal amplitude intensity of the UAV after continuous detrending is divided into equal-length sub-intervals, and discrete Fourier transform is performed on the signals in each interval according to equations (4) and (5). The single-sided spectrum data of the low-frequency band and the high-frequency band are spliced into a complete spectrum. Among them, and represent the low-frequency and high-frequency signal amplitude intensity after DFA processing, and represent the results of discrete Fourier transform. is expressed as the signal interval total number of samples in the spectrum, denoted as frequency index, denoted as time index, denotes complex exponential function.

[0024] (4); (5); In order to eliminate the gain difference between radio devices and ensure the smoothness of the spliced spectrum, the energy ratio of the last sample of the low frequency band to the first sample of the high frequency band is calculated according to formula (6) , and the amplitude spectrum is converted into energy spectrum by normalization.

[0025] (6).

[0026] c represents the energy ratio of the last sample of the low frequency band to the first sample of the high frequency band, q represents the sample point, and Q represents the sample point of one sampling period of the high and low frequency band signal.

[0027] Further, the process of multi-feature channel processing includes: The complete bandwidth spectrum vector is evenly divided into multiple continuous sub-channel frequency bands along the frequency domain dimension; Each sub-channel contains an equal number of sample points; The multi-channel feature spectrum data is combined into a matrix, where each row of the matrix corresponds to one channel data; Each feature channel has a specific frequency range, which is defined by the starting value and the end value of the channel.

[0028] Specifically, the implementation process of the embodiment includes: The UAV signal spectrum data covers the entire UAV link bandwidth, and if the UAV signal sampling frequency is , the spectral resolution is represented as , is the number of sample points. Let the complete bandwidth spectrum be a vector , and its element is represented as . In view of the spatial heterogeneity characteristics of the UAV signal in the energy spectrum data, the spectrum signal is evenly divided into 8 continuous sub-channel frequency bands along the frequency domain dimension, each channel contains n / 8 sample points, and the spectrum data containing multi-feature channels is represented as: (7); In formula (7), denotes the amplitude value of the th sub-channel frequency band frequency point, represents the channel index, . To realize the feature extraction of the neural network on the multi-channel feature spectrum data, the multi-channel feature spectrum data is combined into a matrix , the matrix Each row corresponds to sub-channel data, and is expressed as: (8); The bandwidth of each channel is , , which represents the spectrum resolution of each feature channel. The frequency range of each feature channel is expressed as: (9); , wherein and represent the start and end values of the channel spectrum feature, , represents the minimum value of the channel spectrum feature, , represents the number of sub-channels. The above process obtains the unmanned aerial vehicle signal spectrum data of the multi-feature channel, which facilitates the accurate extraction of the multi-feature channel by the subsequent multi-channel two-dimensional convolutional neural network model.

[0029] Further, the tensorization process includes: The multi-channel feature spectrum data matrix is reshaped into a three-dimensional feature matrix; The three dimensions of the three-dimensional feature matrix respectively represent energy, frequency and channel; The frequency dimension captures the neighborhood aggregation mechanism of the frequency point sequence, the channel dimension maps the difference distribution of the physical characteristics of different frequency band signals, and the energy dimension is fixed as a single value to adapt to the two-dimensional convolution kernel sliding.

[0030] Specifically, the implementation process of the embodiment includes: The application designs a signal recognition model based on a multi-channel two-dimensional convolutional neural network (MC-2DCNN). The self-adaptive feature extraction capability effectively overcomes the defects of insufficient feature representation of traditional design methods in a complex noise environment, and the multi-feature channel extraction strategy is introduced to effectively retain the heterogeneity features of the signal sub-band, and accurately adapt to the multi-channel convolution characteristics of the MC-2DCNN. The specific description is as follows: The unmanned aerial vehicle energy spectrum signal is divided into independent channels, combined with data dimension reorganization to form a three-dimensional tensor, when = 2048, the signal dimension is expressed as 1x256x8, which respectively represents energy, frequency and channel. Through tensor reshaping, the matrix is constructed into a three-dimensional feature matrix This process fully considers the spatial heterogeneity of spectral signals, providing a data structure for subsequent feature extraction and signal recognition tasks. Specifically, the frequency dimension effectively captures the short-time frequency domain correlation of micro-features such as harmonic components and transient noise in UAV signals through a neighborhood aggregation mechanism of local frequency point sequences, maintaining their local continuity and topological correlation. The channel dimension maps the spatial diversity of multi-sensor observation data and the differential distribution of physical characteristics of signals in different frequency bands, preserving the heterogeneous information of each sub-band and establishing coupling relationships for cross-channel feature interactions. The energy dimension, as a convolutional kernel unit, is fixed at 1, precisely adapting to the sliding characteristics of two-dimensional convolutional kernels and enabling rapid feature map formation. This quantization processing strategy, through spatial decoupling and structured recombination of frequency domain features, ensures the topological adjacency of frequency domain features while achieving differentiated modeling of cross-frequency band features through a channel separation mechanism, laying a geometrically interpretable data foundation for subsequent feature extraction by convolutional neural networks.

[0031] Furthermore, the multi-channel two-dimensional convolutional neural network model includes a primary convolutional layer, a max pooling layer, a deep convolutional layer, a second max pooling layer, a fully connected layer, and a result output layer; The primary convolutional layer is configured with convolutional kernels that slide along the frequency dimension, enabling feature fusion across channel dimensions; The max pooling layer downsamples the feature map, compressing the frequency dimension; The deeper convolutional layers use more convolutional kernels to extract higher-order frequency domain features; The fully connected layer flattens the feature map and constructs a non-linear mapping through the hidden layer. The output layer uses a Softmax classifier to output the class probability.

[0032] Specifically, the implementation process of this embodiment includes: A two-layer 2D convolutional neural network structure is employed to achieve hierarchical feature extraction and classification of the input signal. The model input data is a 1×256×8 three-dimensional feature matrix, corresponding to energy, frequency, and channels, respectively. MC-2DCNN includes primary convolutional layers, deep convolutional layers, max-pooling layers, fully connected layers, and a recognition output layer. The parameter transfer and feature extraction process for each layer are as follows: ① Primary convolutional layer: The convolutional kernel is configured to slide along a 1×256×8 dimension at the frequency point. This layer exhibits local perceptual capabilities in the frequency dimension and achieves heterogeneous feature fusion across channel dimensions. It uses 16 1×3 convolutional kernels to slide along the frequency dimension and extract features, extracting local frequency domain fluctuation patterns, represented as: (10); in, This represents the output of the primary convolutional layer. For input data, With the energy dimension fixed at 1, and are frequency bin and channel index respectively, is the convolution kernel, represents the batch normalization operation, represents the activation function. There is one independent convolution kernel for each output channel connected to all input channels. On the input 8 channels, a 1x3 convolution kernel is used for weighted summation and bias addition . The convolution operation is followed by a ReLU activation function to introduce a non-linear decision boundary, and batch normalization (BN) to accelerate training convergence. The primary convolution layer has 8 input channels and 16 output feature maps, and the output is with a frequency dimension of 256.

[0033] ② Max-pooling layer (Pool1): After the primary convolution layer, spatial down-sampling and feature compression are continued. The frequency dimension is compressed to 128 using max-pooling with a 1x2 pooling window. The maximum value is taken for each 1x2 non-overlapping window in the frequency dimension, and the output is represented as: (11) ; where, represents the max-pooling layer output, and the variable represents the new frequency bin index (0-127) after pooling, is the offset of the pooling window. This design preserves significant features while suppressing high-frequency noise and expanding the effective receptive field of subsequent convolution layers. The output of this layer is represented as: .

[0034] ③ Deep convolution layer: This layer uses 32 1x3 convolution kernels to further extract high-order frequency domain features and enhance the model's perception of cross-frequency feature combinations, represented as: (12) ; represents the output of the deep convolution layer, and the number of feature channels of the convolution kernel is multiplied, enhancing the model's ability to extract cross-frequency feature combinations, represents the output of the max-pooling layer, represents the weights of the deep convolution layer. Similarly, the activation function ReLU and batch normalization BN are used for non-linear transformation and stable training, and the output feature map is represented as: .

[0035] ④ Max-pooling layer (Pool2): The feature map is again down-sampled, and the frequency dimension is compressed to 64, further expanding the receptive field and extracting more discriminative features. The output of this layer is represented as: .

[0036] (5) Full connection layer and result output layer: after feature extraction, the model realizes classification and recognition decision through the full connection layer. First, the 1x64x32 feature map is flattened into a 2048-dimensional vector, then a hidden layer containing 64 ReLU neurons is used to construct a nonlinear mapping, and the output layer uses a Softmax classifier to output the probability of 4 categories (background noise / Bebop / AR / Phantom). The classification probability calculation formula is as follows: (13); In formula (13), represents the classification probability, is the classification output result, is the hidden layer output feature vector, represents four categories, and represent the weights and biases for the category .

[0037] Further, the training process of the multi-channel two-dimensional convolutional neural network model includes: using an adaptive moment estimation optimizer to update the model parameters; using a cross-entropy loss function as the optimization objective, which is calculated based on the true label of the sample and the predicted category probability; The training process sets multiple gradient noise intensities for scenarios with electromagnetic space noise and artificial electromagnetic interference, and performs performance evaluation through multi-model comparison and multi-interference scene comparison.

[0038] Specifically, the implementation process of the embodiment includes: using an adaptive moment estimation optimizer (Adam) to update the parameters, and using a cross-entropy loss function as the optimization objective, which is represented as: (14); wherein, is the total number of samples, is the number of categories, is the true label of the th sample, is the probability that the th sample is predicted to be the th category.

[0039] Through the above unmanned aerial vehicle intelligent recognition model construction and combined with the training strategy, a rapid and accurate intelligent recognition scheme for multiple unmanned aerial vehicle types and multiple flight modes is formed.

[0040] The application discloses a method for intelligent recognition of unmanned aerial vehicle signals based on time-frequency multi-channel features, which eliminates the direct current component and low-frequency drift in the radio frequency signal by using a continuous detrended denoising method, acquires time-frequency information by using discrete Fourier transform, and constructs energy spectrum data by spectrum splicing and amplitude normalization, processes multiple feature channels along the frequency domain for spatial heterogeneity, realizes spatial decoupling and structured reorganization of the frequency domain features by combining a multi-channel two-dimensional convolutional neural network model, realizes hierarchical extraction and differential modeling of cross-band features by means of a channel separation mechanism while guaranteeing the feature topology adjacency relationship, thereby improving the accuracy and robustness of the unmanned aerial vehicle signal recognition under complex electromagnetic interference and noise background, realizing rapid and accurate recognition of the unmanned aerial vehicle type and flight mode, enhancing the generalization ability of the model under electromagnetic space noise and human interference such as comb spectrum and blocking scenes, and being suitable for low-altitude airspace safety and electromagnetic spectrum management.

[0041] The above is only a preferred specific embodiment of the application, but the protection scope of the application is not limited to this, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the application, which should be covered in the protection scope of the application. Therefore, the protection scope of the application should be subject to the protection scope of the claims.

Claims

1. A method for intelligent identification of UAV signals based on time-frequency multi-channel features, characterized in that, include: Acquire the radio frequency signals of the drone; The radio frequency signal is preprocessed to obtain preprocessed spectrum data. The preprocessing includes detrending denoising, discrete Fourier transform spectrum analysis, and multi-feature channel processing. The preprocessed spectral data is tensorized to construct a three-dimensional feature matrix. The three-dimensional feature matrix is ​​input into a multi-channel two-dimensional convolutional neural network model for feature extraction and classification to obtain the identification results of drone type and flight mode.

2. The intelligent identification method for UAV signals based on time-frequency multi-channel features as described in claim 1, characterized in that, Acquiring the radio frequency signals of the drone includes: A software-defined radio device is used for dual-channel reception to collect high and low frequency radio frequency signals. Acquire data covering a wide frequency range and record the background noise signal; Signals were collected from multiple standard operating modes of the drone, including static, autonomous hovering, flight without video acquisition, and flight with video recording.

3. The intelligent identification method for UAV signals based on time-frequency multi-channel features as described in claim 1, characterized in that, The de-trending and denoising process includes: The original radio frequency signal sequence is converted into a random walk process; Divide the integrated sequence into equal-length sub-intervals, and then segment the sequence from the beginning and end respectively; Within each subinterval, the local trend is calculated by fitting a local quadratic polynomial. Subtracting the local trend from the original integral sequence yields the fluctuation function; Calculate the root mean square value of the detrended signal sequence for each sub-interval, which is used as the intensity of signal amplitude fluctuation.

4. The intelligent identification method for UAV signals based on time-frequency multi-channel features as described in claim 1, characterized in that, The process of discrete Fourier transform spectrum analysis includes: The amplitude and intensity of the detrended low-frequency and high-frequency signals are divided into equal-length sub-intervals. Perform Discrete Fourier Transform on the signals in each interval to obtain the single-sided spectral data of the low-frequency and high-frequency bands; The spectrum data of low-frequency band and high-frequency band are stitched together to form a complete spectrum; Calculate the energy ratio of the last sample in the low-frequency band to the starting sample in the high-frequency band, and convert the amplitude spectrum into an energy spectrum by normalization.

5. The intelligent identification method for UAV signals based on time-frequency multi-channel features as described in claim 1, characterized in that, The process of multi-feature channel processing includes: The complete bandwidth spectrum vector is uniformly divided into multiple continuous sub-channel frequency bands along the frequency domain dimension; Each sub-channel contains an equal number of sampling points; Multi-channel feature spectrum data are combined into a matrix, where each row of the matrix corresponds to one channel of data; Each feature channel has a specific frequency range, which is defined by the channel's start and end values.

6. The intelligent identification method for UAV signals based on time-frequency multi-channel features as described in claim 1, characterized in that, The process of tensor quantization includes: Reshape the multi-channel feature spectrum data matrix into a three-dimensional feature matrix; The three dimensions of the three-dimensional feature matrix represent energy, frequency, and channel, respectively. The frequency dimension captures the neighborhood aggregation mechanism of the frequency point sequence, the channel dimension maps the difference distribution of the physical characteristics of signals in different frequency bands, and the energy dimension is fixed to a single value to adapt to the sliding of the two-dimensional convolution kernel.

7. The intelligent identification method for UAV signals based on time-frequency multi-channel features as described in claim 1, characterized in that, The multi-channel two-dimensional convolutional neural network model includes a primary convolutional layer, a max pooling layer, a deep convolutional layer, a second max pooling layer, a fully connected layer, and a result output layer; The primary convolutional layer is configured with convolutional kernels that slide along the frequency dimension, enabling feature fusion across channel dimensions; The max pooling layer downsamples the feature map, compressing the frequency dimension; The deeper convolutional layers use more convolutional kernels to extract higher-order frequency domain features; The fully connected layer flattens the feature map and constructs a non-linear mapping through the hidden layer. The output layer uses a Softmax classifier to output the class probability.

8. The intelligent identification method for UAV signals based on time-frequency multi-channel features as described in claim 1, characterized in that, The training process of a multi-channel two-dimensional convolutional neural network model includes: An adaptive moment estimator optimizer is used to update the model parameters; The cross-entropy loss function is used as the optimization objective, and the loss function is calculated based on the true label and predicted class probability of the sample. The training process sets multiple noise intensities for scenarios with electromagnetic spatial noise and human interference, and evaluates performance through multi-model comparison and multi-interference scenario comparison.

Images