A method for unmanned aerial vehicle data link electromagnetic interference classification and threat assessment

Abstract

The application relates to a method for classifying and evaluating electromagnetic interference of a UAV data link, and particularly relates to the technical field of evaluating electromagnetic interference of a UAV data link. The method comprises the following steps: acquiring state parameters of a to-be-predicted UAV data link and I / Q data of an electromagnetic interference signal of the to-be-predicted UAV data link; obtaining a data link performance parameter histogram of the to-be-predicted UAV data link and a graph set of the electromagnetic interference signal according to the state parameters of the to-be-predicted UAV data link and the I / Q data of the electromagnetic interference signal; inputting the graph set and the data link performance parameter histogram into a prediction model to obtain an electromagnetic interference type and a threat degree; and the prediction model is obtained by training an MIMT-CNN network comprising a multi-channel image feature extraction module, a first addition layer, a feature fusion processing module, a second addition layer and a multi-task output module. The application can improve the accuracy of classification and threat prediction results of the electromagnetic interference signal, and reduce the time and storage costs.

Description

Technical Field

[0001] This invention relates to the field of electromagnetic interference assessment technology for UAV data links, and in particular to a method for classifying and assessing electromagnetic interference in UAV data links. Background Technology

[0002] Unmanned aerial vehicles (UAVs), with their automation, low cost, and multifunctionality, are widely recognized as a technology with enormous potential, experiencing rapid growth in many fields across military, civilian, and commercial sectors. However, the electromagnetic environment of airborne wireless communication networks is becoming increasingly complex due to the presence of unauthorized users and malicious interference. As an airborne electronic system, UAV data links are susceptible to communication anomalies, interruptions, and even damage due to electromagnetic interference. Given the limitations of UAV platform missions, relying on ground stations for electromagnetic situational awareness of their data links is problematic. Firstly, it requires extensive domain expertise, placing a heavy cognitive burden on personnel. Secondly, it results in slow response times to electromagnetic interference, making it unable to cope with rapidly changing electromagnetic environments. Therefore, UAV data links need to autonomously identify electromagnetic signals and predict electromagnetic interference threats.

[0003] Many studies on modeling and predicting electromagnetic interference are based on platforms such as electronic components, circuits, and electronic devices. Methods include constructing equivalent circuit models and topological networks, statistical probability-based models, and machine learning.

[0004] Methods based on equivalent circuit models require extensive knowledge of technical principles and a detailed understanding of electronic device structures. Models using statistical probability require domain experts to extract features from the data samples of the test platform, thereby reducing data dimensionality and facilitating classification or prediction by the algorithm. Deep learning, a branch of machine learning, can autonomously learn more complex features from data through multiple nonlinear transformations involving billions of weight parameters, thus reducing reliance on specialized knowledge and feature extraction rules. Convolutional Neural Networks (CNNs), a typical application of deep learning in computer vision, are data-driven deep neural network structures. Due to their excellent feature extraction capabilities, they have been applied in the classification and threat assessment of electromagnetic interference signals. However, previous methods using CNNs for electromagnetic interference signal classification and threat prediction often required separate modeling for each task, significantly increasing the computational and storage costs of the model and ignoring the relationship between the two tasks, resulting in inaccurate results. Summary of the Invention

[0005] The purpose of this invention is to provide a method for classifying and assessing electromagnetic interference in UAV data links, which can improve the accuracy of electromagnetic interference signal classification and threat prediction results, and reduce time and storage costs.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A method for classifying and assessing electromagnetic interference in UAV data links, comprising:

[0008] Acquire the state parameters of the UAV data link to be predicted and the I / Q data of the electromagnetic interference signal of the UAV data link to be predicted; the state parameters include: automatic gain control voltage, signal-to-noise ratio and bit error rate;

[0009] Based on the state parameters of the unmanned aerial vehicle (UAV) data link to be predicted, a histogram of the data link performance parameters of the UAV data link to be predicted is obtained.

[0010] An atlas of electromagnetic interference signals of the UAV data link to be predicted is obtained based on the I / Q data of the electromagnetic interference signals of the UAV data link to be predicted; the atlas includes a short-time Fourier transform time spectrum diagram and a density constellation diagram.

[0011] The atlas of electromagnetic interference signals of the UAV data link to be predicted and the histogram of the data link performance parameters of the UAV data link to be predicted are input into the prediction model to obtain the electromagnetic interference type and electromagnetic interference threat level of the UAV data link to be predicted; the prediction model is obtained by training a MIMT-CNN network; the MIMT-CNN network specifically includes: a multi-channel image feature extraction module, a first summing layer, a feature fusion processing module, a second summing layer and a multi-task output module connected in sequence.

[0012] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: The present invention inputs the atlas of electromagnetic interference signals of the UAV data link to be predicted and the histogram of the data link performance parameters of the UAV data link to be predicted into the prediction model to obtain the electromagnetic interference type and electromagnetic interference threat level of the UAV data link to be predicted; the prediction model is obtained by training a MIMT-CNN network; the MIMT-CNN network specifically includes: a multi-channel image feature extraction module, a first summing layer, a feature fusion processing module, a second summing layer, and a multi-task output module connected in sequence. By applying a multi-input channel MIMT-CNN network to link electromagnetic interference classification and threat assessment, the accuracy of electromagnetic interference signal classification and threat prediction results can be improved, and time and storage costs can be reduced. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 A flowchart illustrating a method for classifying and assessing electromagnetic interference in UAV data links, provided as an embodiment of the present invention;

[0015] Figure 2 A flowchart for training a MIMT-CNN network provided in an embodiment of the present invention;

[0016] Figure 3 This is a structural diagram of the MIMT-CNN network provided in an embodiment of the present invention;

[0017] Figure 4 The graph shows the change of the loss function value with the number of iterations during the training process of the MIMT-CNN network provided in this embodiment of the invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] like Figure 1 As shown, a multi-task CNN with multi-input (MIMT-CNN) network is used for electromagnetic environment risk perception of UAV data links. Based on this, this embodiment of the invention provides a method for UAV data link electromagnetic interference classification and threat assessment, including:

[0021] Step 101: Obtain the state parameters of the UAV data link to be predicted and the I / Q data of the electromagnetic interference signal of the UAV data link to be predicted. The state parameters include: Automatic gain control (AGC) voltage, signal-to-noise ratio (SNR), and bit error rate (BER).

[0022] Step 102: Obtain the data link performance parameter histogram of the UAV data link to be predicted based on the state parameters of the UAV data link to be predicted.

[0023] Step 103: Obtain an atlas of electromagnetic interference signals of the UAV data link to be predicted based on the I / Q data of the electromagnetic interference signals of the UAV data link to be predicted; the atlas includes a short-time Fourier transform spectrum diagram and a density constellation diagram.

[0024] Step 104: Input the atlas of electromagnetic interference signals of the UAV data link to be predicted and the histogram of the data link performance parameters of the UAV data link to be predicted into the prediction model to obtain the electromagnetic interference type and electromagnetic interference threat level of the UAV data link to be predicted; the prediction model is obtained by training a multi-task CNN with multi-input (MIMT-CNN) network; the MIMT-CNN network specifically includes: a multi-channel image feature extraction module, a first summing layer, a feature fusion processing module, a second summing layer, and a multi-task output module connected in sequence.

[0025] As an optional implementation, the MIMT-CNN network first extracts features from visualized data link performance parameters (data link performance parameter histogram), the short-time Fourier transform spectrum of the electromagnetic interference signal, and the density constellation diagram as inputs to the multi-channel network. Then, based on multi-task learning, electromagnetic signal classification and threat assessment are simultaneously learned within a shared parallel CNN. Next, to further improve model performance, Bayesian optimization methods are used to optimize the structure and hyperparameters of the parallel part of the network. Figure 2 As shown, the method for determining the prediction model includes:

[0026] Electromagnetic interference injection experiments were conducted on the data link under different types of electromagnetic interference to obtain the state parameters of the UAV data link and the I / Q data of the electromagnetic interference signal under each type of electromagnetic interference.

[0027] Determine the electromagnetic interference threat level of the UAV data link corresponding to each type of electromagnetic interference.

[0028] Based on the state parameters of the UAV data link under each type of electromagnetic interference, a histogram of the data link performance parameters under each type of electromagnetic interference is obtained.

[0029] Based on the I / Q data of the electromagnetic interference signals of the UAV data link under each electromagnetic interference type, an atlas of electromagnetic interference signals for each electromagnetic interference type is obtained; the atlas includes a short-time Fourier transform time-spectrum diagram and a density constellation diagram.

[0030] Using histograms of data link performance parameters of UAV data links under various electromagnetic interference types and atlases of electromagnetic interference signals of UAV data links under various electromagnetic interference types as sample sets, the MIMT-CNN network is trained to obtain the prediction model. CNN is a deep neural network specifically designed for processing grid-like data such as images. By introducing concepts such as local receptive fields, neurons, activation, and sparsity, the network can perceive deep abstract features of data like the human brain, and is widely used in the field of computer vision. Let K be the number of samples in each class, and the data link normalization parameter histogram matrix be... The STFT spectrum matrix is ​​as follows The standardized density constellation diagram matrix is The input image matrix is The model's output targets are the type of electromagnetic interference and its threat level to the data link. Let the electromagnetic interference type be labeled ν. k Since electromagnetic interference categories are independent of each other, one-hot encoding is used to make the category values ​​more reasonable. The data link interference threat level is labeled as u k Then the label matrix is The mapping established by the network is The MIMT-CNN network is trained with the label matrix as output and the mapping established by the network as input to obtain the prediction model.

[0031] Electromagnetic interference (EMI) injection experiments were conducted to obtain the state parameters of the data link and the in-phase / quadrature (I / Q) data of the electromagnetic space. As an optional implementation method, EMI injection experiments were performed on the data link under different EMI types to obtain the state parameters of the UAV data link and the I / Q data of the EMI signal for each EMI type. Specifically, this included:

[0032] Data link electromagnetic interference injection experiments were conducted under different electromagnetic interference types, interference intensities, and interference frequency parameters. Data link status parameters (AGC / SNR / BER) were collected from the data link detection software, and I / Q data of the data link received signal under the corresponding electromagnetic interference conditions were collected using an electromagnetic spectrum detection receiver.

[0033] As an optional implementation method, the electromagnetic interference threat level (electromagnetic interference threat level) of the UAV data link corresponding to each type of electromagnetic interference is determined, specifically including:

[0034] A unified electromagnetic interference threat level classification method is adopted for different types of electromagnetic interference to assess the interference level of the data link. Based on the operational practice of UAV data links in application scenarios, a data link is defined as being threatened when the difference between the interference power and the loss-of-lock power is less than 6dB. Based on the difference between the interference power and the data link loss-of-lock threshold, interference is divided into four levels, corresponding to differences of 6dB, 3dB, 1dB, and 0dB, respectively.

[0035] As an optional implementation, a histogram of UAV datalink performance parameters under each electromagnetic interference type is obtained based on the state parameters of the UAV datalink under each electromagnetic interference type. The datalink state parameters are then transformed into a visualized parameter histogram, specifically including:

[0036] The data link state parameters are normalized, and then visualized to obtain a histogram for each electromagnetic interference state, which is used to characterize the state information of the data link when it is interfered with; the initial state parameters of the UAV when there is no electromagnetic interference. State parameter p after applying electromagnetic interference m ∈{p AGC ,p SNR ,p BER}, the maximum value of the state parameter Using formula The three state parameters (AGC, SNR, and BER) are normalized respectively, where, This is the processed data. Through normalization, the original data underwent a linear transformation, mapping the initial input to the range [0, 1], thereby transforming the performance parameters into a dimensionless expression. The normalized data is then visualized in a histogram.

[0037] As an optional implementation, an atlas of electromagnetic interference signals for the UAV data link under each electromagnetic interference type is obtained based on the I / Q data of the electromagnetic interference signals under each electromagnetic interference type. Specifically, this includes:

[0038] The I / Q data is converted into a time-spectrum diagram through a short-time Fourier transform (STFT) to show the change of interference signal power with time and frequency. After standardizing the I / Q data, a standardized density constellation diagram is drawn to show the phase information and noise information of the interference.

[0039] Time-frequency analysis is performed on the I / Q data to obtain the STFT spectrum, i.e., the short-time Fourier transform time-frequency spectrum, thus obtaining the time-frequency information of the interference signal. An analysis window of length M is slid across the received signal r(n), i.e., the I / Q data. The short-time Fourier transform of the signal is calculated by calculating the Discrete Fourier Transform (DFT) of the windowed sampled signal. Let L be a non-zero overlap length used to compensate for signal attenuation at the window edges. Assume R... k (f) is the DFT of the sampled signal at frequency f with time k(ML) as the window center. Where n represents the time sampling points, and g(n) is a window function of length M. N r It is the length of the sampled signal r(n).

[0040] The square of the DFT modulus for each window segment |R k (f)| 2 Combining the results, we obtain matrix |R STFT (f)| 2 This matrix contains the amplitude and phase at each time and frequency point. For |R STFT (f)| 2 Visualization yields the STFT spectrum.

[0041] The I / Q data is standardized, and a standardized density constellation diagram is plotted to obtain the phase information of the interference signal. The amplitude of the received signal (I / Q data) varies under different interferences, therefore, it is necessary to select an appropriate region to observe the points on the constellation diagram. If the selected region is too large, the signal samples are compressed into a small area, making it impossible to effectively observe the distribution of the sampling points; if the selected region is too small, some signal samples may be excluded from the image. Therefore, the amplitude of the sampling points is scaled and adjusted to [-1,1] to obtain a standardized constellation diagram of uniform size. Due to the superposition of the data link working signal, interference signal, and noise signal, the sampling points overlap on the constellation diagram, making it difficult to determine their distribution characteristics. However, the sample point distribution density varies in different regions, so the density of the points can be used to enhance the features of the constellation diagram image.

[0042] By calculating the ratio of the number of sampling points d(i,j) to the window area within a circular window centered at point coordinate (i,j) with radius r, the normalized point density ρ(i,j) of a given point can be obtained.

[0043]

[0044] Then, the constellation diagram is colored according to its density. After preprocessing, each point in the constellation diagram is no longer independent or has the same amount of information. This processing method accumulates the points in the constellation diagram in the time dimension, making the data feature dimension higher, incorporating more prior knowledge of the modulation signal, and realizing feature enhancement of the constellation diagram.

[0045] As an optional implementation, a MIMT-CNN is constructed and trained using randomly partitioned training data. Bayesian optimization is used to optimize the network's hyperparameters on the validation set. The trained network model is then tested on a test set, and its accuracy and generalization ability are verified based on the actual interference type classification results and interference threat assessment results. The step of training the MIMT-CNN network to obtain the prediction model using a sample set consisting of histograms of data link performance parameters of UAV data links under each electromagnetic interference type and atlases of electromagnetic interference signals of UAV data links under each electromagnetic interference type, specifically includes:

[0046] The sample set is divided into a training set, a validation set, and a test set according to a set ratio.

[0047] The MIMT-CNN network is initially trained using the training set to obtain a trained MIMT-CNN network.

[0048] The trained MIMT-CNN network is optimized using a validation set to obtain an optimized MIMT-CNN network.

[0049] The optimized MIMT-CNN network is tested using a test set to obtain the prediction model.

[0050] like Figure 3As shown, the network input is divided into multiple channels, which, according to the input type, input the histogram of data link performance parameters, the spectrum of short-time Fourier transform, and the density constellation diagram corresponding to the same interference state. Each channel first extracts features, then performs feature fusion, and finally obtains the multi-task output. The feature extraction layer structure of each channel is the same, consisting of two feature extraction layers superimposed with several convolutional layers and activation layers, used for image feature learning and dimensionality reduction. After feature extraction, the features are added together and enter two parallel feature fusion layers, each composed of a different number of fusion modules. The feature fusion part (feature fusion processing module) uses a parallel network structure to capture the correlation between different tasks, extracting feature information from different perspectives to obtain more comprehensive feature information. The network hyperparameters of the fusion layer are improved using a Bayesian optimization algorithm, selectively fused according to the characteristics of the network input feature parameters. The fused parameter matrix enters the multi-task output part (multi-task output module), outputting the electromagnetic interference classification result and the predicted value of the data link interference performance level, respectively. The model's input layer for each channel initially takes a 100×100×3 RGB (Red, Green, Blue) image as input. The three types of feature image data are divided by 255 to normalize each pixel value to the [0, 1] interval. Then, each channel's input layer extracts features. Because the input data volume is very large, it's unsuitable to load all data at once for gradient calculation and weight updates. Therefore, the input data is divided into mini-batches, which allows the network to achieve greater generalization ability. Each input channel passes through the same number and size of convolutional layers, activation layers, pooling layers, and batch normalization (BN) layers. Each channel contains two consecutive feature extraction layers, consisting of two 3×3 convolutional layers, two ReLU activation layers, and a 2×2 max-pooling layer, used for network feature extraction and dimensionality reduction. For inputs with a wide distribution of image features, larger convolutional layers can better achieve feature extraction, while for inputs with a concentrated distribution of image features, smaller convolutional layers can better perform learning. Therefore, mimicking the Inception structure, the feature fusion processing part enhances the network's adaptability to feature scales by using convolutional layers of different sizes in parallel. Specifically, to select better layer parameters and improve the network's adaptability to features, the size and number of convolutional layers are obtained through Bayesian optimization. The network's output layer has two branches: classification and regression. The regression output is a fully connected layer of size 1, outputting the data link interference threat level *u*. The classification layer consists of a fully connected layer of size 4 and a softmax layer, outputting the electromagnetic interference type *v*. Where w i Let v be the output value of the i-th node in the layer preceding Softmax, v be the classification result, and C be the number of output nodes, i.e., the number of classification categories. Softmax converts the output values ​​of multi-class classification into a probability distribution in the range [0,1] with a sum of 1, and uses the highest probability as the basis for class determination.

[0051] As an optional implementation, the multi-channel image feature extraction module includes three parallel image feature extraction sub-modules with identical structures. Each image feature extraction sub-module includes, in sequence, an input layer, a feature extraction unit, a first convolutional layer, a first ReLU activation layer, a second convolutional layer, a first Batchnorm layer, and a second ReLU activation layer.

[0052] As an optional implementation, the feature fusion processing module includes: a first feature extraction submodule and a second feature extraction submodule connected in parallel. The first feature extraction submodule includes a first feature fusion layer, a second feature fusion layer and a first fully connected layer connected in sequence. The second feature extraction submodule includes a third feature fusion layer and a second fully connected layer connected in sequence.

[0053] As an optional implementation, the multi-task output module includes: a classification unit and a third fully connected layer connected in parallel; the classification unit includes a fourth fully connected layer and a Softmax layer connected in sequence.

[0054] As an optional implementation, the feature extraction unit includes two sequentially connected feature extraction layers; the feature extraction layers include: a third convolutional layer, a third ReLU activation layer, a fourth convolutional layer, a fourth ReLU activation layer, and a max pooling layer, which are sequentially connected.

[0055] As an optional implementation, the first feature fusion layer, the second feature fusion layer and the third feature fusion layer each include: a fifth convolutional layer, a second batchnorm layer and a fifth ReLU activation layer connected in sequence.

[0056] As an optional implementation, the step of using the training set to perform initial training on the MIMT-CNN network to obtain a trained MIMT-CNN network specifically involves:

[0057] With the goal of minimizing the loss function, the MIMT-CNN network is initially trained using the training set to obtain a trained MIMT-CNN network. The loss function is used to evaluate the degree of difference between the actual value and the predicted value; the smaller the loss function, the better the model's performance. Since the MIMT-CNN network has multiple task outputs, the network's loss function consists of two parts: classification loss L... v and regression loss L u The classification loss L cUsing FocalLoss loss, compared to the commonly used cross-entropy loss, Focal loss doesn't change the loss for inaccurately classified samples, but it reduces the loss for accurately classified samples. Overall, it's equivalent to increasing the weight of inaccurately classified samples in the loss function, thus forcing the network to focus on difficult samples. v =-α(1-p v ) γ log(p v ), where γ is the focusing parameter, γ≥0. Increasing γ improves the network's sensitivity to misclassified data; in this paper, it is set to the default value γ=2. α is the balancing coefficient, α=0.25. The value of α can be set to reduce the weight of a large number of negative samples. p v To predict the probability of belonging to this category, (1-p) v ) γ The modulation coefficients reduce the weight of easily classified samples. Due to the change in the modulation coefficients, difficult-to-classify samples take center stage, causing the model to focus more on these samples during training and improving classification accuracy.

[0058] Regression loss L u Using the semi-mean squared error loss function: in It's a network response, u k It is the predicted target value, K bs Network response within a small batch range k The total number, K is u k The total number of observations.

[0059] As training progresses, the two loss functions L... v and L u The rate of decrease is inconsistent. To unify the loss to the same order of magnitude and avoid small gradients (L...),... v L with a large gradient u Take it away, therefore for L v and L u By using different weights ο1 and ο2, the final loss function becomes a weighted sum of the classification loss and the regression loss, thus enhancing the model's generalization ability.

[0060] L all =ο1L v +ο2L u To balance the impact of classification loss and regression loss on training convergence speed, after adjustment, ο1 = 1 and ο2 = 0.1 are set.

[0061] As an optional implementation, the step of testing the optimized MIMT-CNN network with a test set to obtain the prediction model specifically involves:

[0062] The optimized MIMT-CNN network was tested on a test set to obtain the prediction model, aiming to maximize classification accuracy and minimize RMSE and MAPE. Since the network can perform both classification and regression tasks simultaneously, each task needs to be evaluated separately. Classification performance is commonly evaluated using accuracy. Assume v T and v F Let represent the number of correctly classified and incorrectly classified categories, respectively. Then the classification accuracy Acc is...

[0063] Regression performance is commonly assessed using Root Mean Squared Error (RMSE) and Mean Absolute Percent Error (MAPE) to evaluate prediction accuracy. RMSE measures how well the predicted values ​​fit the actual curve. In contrast, MAPE is more sensitive to the error between the predicted and actual values.

[0064] Among them, u i It is the target result. It is the prediction result, K test This represents the number of samples in the test set. MAPE is one of the most commonly used metrics for evaluating prediction accuracy, ranging from [0, +∞), and is highly affected by outliers.

[0065] This invention provides an embodiment of dividing a sample set into a training set, a validation set, and a test set in a 6:2:2 ratio. The training set optimizes the weights by updating the model parameters. The validation set is used to optimize the network's hyperparameters to prevent overfitting and underfitting, with the goal of obtaining a better model. Finally, the network is evaluated on the test set to determine the final model.

[0066] The data link operating signals are transmitted from the ground station data link, using a binary phase shift keying (BPSK) communication method. Therefore, the interference signal types selected are continuous wave, broadband white Gaussian noise interference, and BPSK interference that are similar to the data link communication parameters, all common in actual interference scenarios. The network is trained using normalized performance parameters, STFT spectrum diagrams, and normalized density constellation diagrams obtained from electromagnetic susceptibility experiments.

[0067] When training using the Adaptive Moment Estimation Adam optimizer, the resulting loss function curve for the training process is as follows: Figure 4 As shown, the function converges relatively quickly, but the loss function value after convergence is low, approximately 0.2.

[0068] The MIMT-CNN network achieved a classification accuracy of 95.45%, with a prediction RMSE of 0.49 and a MAPE of 10.83%. The input-to-output time was 14.81 ms.

[0069] The present invention has the following technical effects:

[0070] 1. The convolutional neural network used is a deep learning intelligent method that can automatically perceive electromagnetic interference risks, increasing the autonomy and intelligence of the UAV data link and avoiding increasing the cognitive burden on operators.

[0071] 2. Using visualized data link performance parameters, electromagnetic interference signal STFT spectrum diagrams, and density constellation diagrams as model inputs improves model accuracy. This is because information obtained through electromagnetic susceptibility testing has different data formats and variation patterns. By visualizing this heterogeneous information and using it as model input, interpretability is enhanced, aligning with human visual perception. Furthermore, a single image contains a large amount of data, rich in effective information, and exhibits strong noise resistance, resulting in high model accuracy. MIMT-CNN can fuse different types of image information, making the model's understanding of the electromagnetic environment more accurate and further improving model precision.

[0072] 3. MIMT-CNN can simultaneously classify electromagnetic signals and assess threats, resulting in high prediction efficiency. The model fully utilizes the correlations among features in the output dataset, assessing the electromagnetic interference threat level while determining the type of electromagnetic interference, thus avoiding the time and computational costs of training multiple models.

[0073] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0074] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for classifying and assessing electromagnetic interference in UAV data links, characterized in that, include: Acquire the state parameters of the data link of the UAV to be predicted and the I / Q data of the electromagnetic interference signal of the data link of the UAV to be predicted; The status parameters include: automatic gain control voltage, signal-to-noise ratio, and bit error rate; Based on the state parameters of the unmanned aerial vehicle (UAV) data link to be predicted, a histogram of the data link performance parameters of the UAV data link to be predicted is obtained. An atlas of electromagnetic interference signals of the UAV data link to be predicted is obtained based on the I / Q data of the electromagnetic interference signals of the UAV data link to be predicted; the atlas includes a short-time Fourier transform time spectrum diagram and a density constellation diagram. The atlas of electromagnetic interference signals of the UAV data link to be predicted and the histogram of the data link performance parameters of the UAV data link to be predicted are input into the prediction model to obtain the electromagnetic interference type and electromagnetic interference threat level of the UAV data link to be predicted; the prediction model is obtained by training a MIMT-CNN network; the MIMT-CNN network specifically includes: a multi-channel image feature extraction module, a first summing layer, a feature fusion processing module, a second summing layer and a multi-task output module connected in sequence.

2. The method for classifying and assessing electromagnetic interference in UAV data links according to claim 1, characterized in that, The multi-channel image feature extraction module includes three parallel image feature extraction sub-modules with identical structures. Each image feature extraction sub-module includes, in sequence, an input layer, a feature extraction unit, a first convolutional layer, a first ReLU activation layer, a second convolutional layer, a first batchnorm layer, and a second ReLU activation layer.

3. The method for classifying and assessing electromagnetic interference in UAV data links according to claim 1, characterized in that, The feature fusion processing module includes: a first feature extraction submodule and a second feature extraction submodule connected in parallel. The first feature extraction submodule includes a first feature fusion layer, a second feature fusion layer and a first fully connected layer connected in sequence. The second feature extraction submodule includes a third feature fusion layer and a second fully connected layer connected in sequence.

4. The method for classifying and assessing electromagnetic interference in UAV data links according to claim 1, characterized in that, The multi-task output module includes: a classification unit and a third fully connected layer connected in parallel; the classification unit includes a fourth fully connected layer and a Softmax layer connected in sequence.

5. The method for classifying and assessing electromagnetic interference in UAV data links according to claim 1, characterized in that, The method for determining the prediction model includes: Electromagnetic interference injection experiments were conducted on the data link under different types of electromagnetic interference to obtain the state parameters of the UAV data link and the I / Q data of the electromagnetic interference signal under each type of electromagnetic interference. Determine the electromagnetic interference threat level of the UAV data link corresponding to each type of electromagnetic interference; Based on the state parameters of the UAV data link under each type of electromagnetic interference, a histogram of the data link performance parameters of the UAV data link under each type of electromagnetic interference is obtained. Based on the I / Q data of the electromagnetic interference signals of the UAV data link under each electromagnetic interference type, an atlas of electromagnetic interference signals of the UAV data link under each electromagnetic interference type is obtained; the atlas includes a short-time Fourier transform time spectrum diagram and a density constellation diagram. The MIMT-CNN network is trained using a sample set consisting of various electromagnetic interference types, the electromagnetic interference threat level of UAV data links corresponding to each electromagnetic interference type, histograms of data link performance parameters of UAV data links under each electromagnetic interference type, and a set of electromagnetic interference signals of UAV data links under each electromagnetic interference type.

6. The method for classifying and assessing electromagnetic interference in UAV data links according to claim 2, characterized in that, The feature extraction unit includes two sequentially connected feature extraction layers; the feature extraction layers include a third convolutional layer, a third ReLU activation layer, a fourth convolutional layer, a fourth ReLU activation layer, and a max pooling layer, which are sequentially connected.

7. The method for classifying and assessing electromagnetic interference in UAV data links according to claim 3, characterized in that, The first feature fusion layer, the second feature fusion layer, and the third feature fusion layer each include: a fifth convolutional layer, a second batchnorm layer, and a fifth ReLU activation layer connected in sequence.

8. The method for classifying and assessing electromagnetic interference in UAV data links according to claim 5, characterized in that, The method of using a set of data as a sample set, including the electromagnetic interference threat level of each type of electromagnetic interference, the data link performance parameter histogram of each type of electromagnetic interference, and the set of electromagnetic interference signals of each type of electromagnetic interference, to train the MIMT-CNN network to obtain the prediction model, specifically includes: dividing the sample set into a training set, a validation set, and a test set according to a set ratio. The MIMT-CNN network is initially trained using the training set to obtain a trained MIMT-CNN network; The trained MIMT-CNN network is optimized using a validation set to obtain an optimized MIMT-CNN network; The optimized MIMT-CNN network is tested using a test set to obtain the prediction model.

Images