Wireless communication composite interference identification method and system suitable for low interference-to-noise ratio condition

Abstract

The invention belongs to but not limited to the technical field of wireless communication, and discloses a wireless communication composite interference identification method and system suitable for a low interference-to-noise ratio condition, and the method comprises the steps: carrying out the data preprocessing of an input IQ sequence of the communication composite interference of an unmanned plane, and obtaining two modal data: time-frequency diagram and frequency domain data; constructing a lightweight neural network, and performing feature extraction on the two independent input data; jointly processing the input features of the two modals by using a connection operation, generating a multi-modal joint feature, and realizing feature fusion; using a sigmoid function to map the output of the full connection layer into the probability corresponding to the composition component of the unmanned aerial vehicle communication composite interference signal, applying a multi-threshold optimization algorithm to obtain the prediction category of the interference component contained in the composite signal, and calculating the final loss through a binary cross entropy loss function. According to the method, the unmanned aerial vehicle communication composite interference identification accuracy can be improved in a low interference-to-noise ratio and different training sample data sets.

Description

Technical Field

[0001] The present invention belongs to but is not limited to the field of wireless communication technology, and in particular relates to a method and system for identifying composite interference in wireless communication suitable for low interference-to-noise ratio conditions. Background Art

[0002] Drones are widely used in military and civilian fields due to their versatility. However, due to the openness of ground-to-air wireless channels and the dominant Loss (line-of-sight) propagation links, physical layer communications are vulnerable to interference attacks, resulting in drone communication failures. In particular, the emergence of composite interference signals further exacerbates the vulnerability of drone communication systems, which can reduce the quality of drone communications or completely interrupt drone communication links. Therefore, efficiently and accurately identifying the types of composite interference signals in drone communications is of great practical significance for improving the anti-interference capability of drone communication systems and ensuring their safe communications.

[0003] Although the data-driven deep learning UAV communication composite interference identification method that does not rely on any prior knowledge can improve the applicability of the model in complex communication scenarios. However, the composite interference identification algorithm has the problem of low recognition accuracy under low interference-to-noise ratio conditions. At the same time, the proposed composite interference signal recognition model usually requires a lot of storage and signal processing resources, and fails to effectively utilize the global characteristics of the interference signal. In addition, in the actual UAV communication composite interference signal identification work, researchers often focus on processing the image data of a single mode of the composite interference signal, ignoring the complementarity of information between different modes of the composite interference signal.

[0004] In view of the above analysis, the technical problems that urgently need to be solved in the existing technology are: the single interference identification method for UAV communications based on deep learning is difficult to cope with the situation of composite interference signal identification. At the same time, there are problems such as low accuracy in composite interference signal identification under low interference-to-noise ratio (INR) and large number of model parameters. Summary of the invention

[0005] In view of the problems existing in the prior art, the present invention provides a method and system for identifying composite interference in wireless communications suitable for low interference-to-noise ratio conditions.

[0006] The present invention is implemented as follows: a method for identifying composite interference in wireless communication suitable for low interference-to-noise ratio conditions, comprising the following steps:

[0007] Step 1: preprocess the IQ sequence of the input UAV communication composite interference to obtain two modal data: time-frequency diagram and frequency domain data;

[0008] Step 2: Build a lightweight neural network MIRNet and extract features from two independent input data respectively;

[0009] Step 3: Use the connection operation to jointly process the input features of the two modalities to generate multi-modal joint features and achieve feature fusion;

[0010] Step 4: Use the sigmoid function to map the output of the fully connected (FC) layer to the probability corresponding to the components of the composite interference signal of UAV communication, and apply the multi-threshold optimization algorithm to obtain the predicted category of the interference component contained in the composite interference signal, and calculate the final loss through the binary cross entropy loss function.

[0011] Furthermore, MIRNet includes an EConv (extraction and squeeze Convolution) module, a D-FAF (downsampling-frequency adaptive filter) module consisting of a downsampling module and a FAF module, and a feature fusion module;

[0012] Firstly, asymmetric convolution (AConv), EConv, CBAM and downsampling modules are cascaded to extract shallow features of the input data. Secondly, three D-FAF modules are stacked to extract deep features of the time-frequency graph, and one D-FAF module is used to extract deep features of the frequency domain data. Then, after a global pooling operation, the one-dimensional feature vectors of the time-frequency graph branch network TFINet and the frequency domain data branch network FDNet are obtained respectively. Finally, the features of the two modalities are concatenated and fused through the connection operation, and the sigmoid function is used for activation to obtain the final probability output.

[0013] Furthermore, the FAF module is composed of the LN (Layer Norm) layer, the Channel Conv module, and the adaptive filtering (AF) module in sequence; the LN layer is used to alleviate the gradient vanishing problem caused by the depth of the neural network; the Channel Conv is used to improve the model's ability to capture feature details; the AF module, as the core of the FAF module, first performs Fourier transform on the input features, then passes through the CBAM module, and finally performs inverse Fourier transform; that is, given a feature After being processed by the AF module, The processing is expressed as:

[0014]

[0015] in and denote Fourier transform and inverse Fourier transform, respectively, ⊙ denotes Hadamard product, and C(·) denotes CBAM;

[0016] Furthermore, the D-FAF module consists of a downsampling module and a FAF module; the downsampling module is used to extract local features of the input data; the FAF module is used to extract global features; by splicing local features and global features, a more comprehensive feature representation is obtained.

[0017] Furthermore, in step 3, the output feature vector F of FDNet is FD Indicated as M i =[m 1 ,m 2 ,…,m i ], the output feature vector F of TFINet TFI Indicated as N j =[n 1 ,n 2 ,…,n j ], the joint feature can be represented as a stack of two one-dimensional vectors F FD and F TFI , the feature fusion process can be expressed as follows:

[0018]

[0019] Where ⊕ represents the connection function.

[0020] Furthermore, in step 4, the input label vector Y and the output probability vector P of the training sample are used to optimize the multi-threshold vector V; the output vector P of the neural network model = [p 1 ,p 2 ,...p M ] According to the decision function F(·) is converted into a label decision vector Y = F(P) = [f(p 1 ),f(p 2 ),...,f(p M )]; the decision function f(·) is expressed as:

[0021]

[0022] where ε i represents the decision threshold of the i-th class label, ε i ∈(0,1).

[0023] Furthermore, a direct search algorithm is used to calculate the optimal threshold for each label; iterate with a certain step size within the given threshold range, and calculate the F1 score of each label under the current threshold; finally, the threshold that maximizes the F1 score is selected as the optimal threshold for the label; that is, during the model training process, the threshold of the i-th label obtained from L training samples is as follows:

[0024]

[0025] where yji , Represents the training samples x j The true labels and predicted labels.

[0026] Furthermore, the loss function in step 4 is a binary cross entropy loss function, and the L sample output labels are calculated independently:

[0027]

[0028] in, represents the predicted label after multi-threshold optimization, y (i) Indicates the actual label; Adam optimizer is used in the training process, and the weights and biases are adjusted iteratively using the gradient of Loss;

[0029]

[0030] in, represents the kth trainable weight in the l+1 layer, is the corresponding bias and η is the step size.

[0031] Another object of the present invention is to provide a method for identifying composite interference in wireless communication applicable to low interference-to-noise ratio conditions, comprising the following steps:

[0032] The data preprocessing module is used to preprocess the IQ sequence of the input UAV communication composite interference to obtain two modal data: time-frequency diagram and frequency domain data;

[0033] The feature extraction module builds a lightweight neural network MIRNet and extracts features from two independent input data respectively;

[0034] The feature fusion module is used to jointly process the input features of the two modalities using a connection operation to generate multi-modal joint features and realize feature fusion;

[0035] The signal recognition module is used to map the output of the fully connected (FC) layer to the probability corresponding to the components of the composite interference signal of UAV communication using the sigmoid function. The multi-threshold optimization algorithm is applied to obtain the predicted category of the interference component contained in the composite interference signal, and the final loss is calculated through the binary cross entropy loss function.

[0036] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0037] First, the present invention proposes a method for identifying composite interference in UAV communications based on multi-modal feature fusion of multi-label learning, and designs a lightweight MIRNet (multi-modal interference network) model to achieve more accurate and stable identification of composite interference signals in UAV communications under low interference-to-noise ratio conditions. A data set containing 31 types of composite interference signals was constructed, and simulation experiments were carried out under sufficient samples, insufficient samples, and different training methods. The results show that the proposed method can improve the accuracy of identifying composite interference in UAV communications in low interference-to-noise ratio and different training sample data sets. At the same time, the designed neural network model has relatively few parameters and is easy to deploy.

[0038] In order to improve the recognition accuracy of composite interference signals of UAV communications under low interference-to-noise ratio conditions, a composite interference signal recognition method for UAV communications based on multi-label learning is proposed in this paper. A frequency domain adaptive filtering (FAF) module is designed to capture the global dependency of features, and a dual-input MIRNet (multi-modal interference network) is constructed to mine and fuse the feature information of different modes of composite interference signals of UAV communications.

[0039] For drones with limited flight time and limited computing power, this paper designs an asymmetric deep convolution module, which uses asymmetric one-dimensional convolution kernels to enhance the performance of square convolution kernels. This design not only improves the model's ability to extract subtle features and finer resolution, but also effectively reduces the number of network parameters by using deep convolution, making it easier to deploy the model.

[0040] A dataset of composite interference signals for UAV communication was constructed and a large number of simulation experiments were conducted. The results show that under low interference-to-noise ratio conditions, the proposed MIRNet outperforms other baseline models in terms of accuracy in identifying composite interference in UAV communication, while maintaining a relatively small number of neural network parameters. The present invention also comprehensively analyzes the performance of composite interference identification in UAV communication under scenarios with insufficient sample data and different training methods, and conducts degradation experiments to further verify the effectiveness and generalization ability of the designed model.

[0041] Second, the present invention proposes a multi-label-based multimodal UAV communication composite interference recognition method, and designs a lightweight MIRNet for fusing time-frequency images and frequency domain data features to improve the recognition accuracy of UAV communication composite interference under low interference-to-noise ratio conditions. The present invention constructs a UAV composite communication interference dataset, and conducts simulation experiments under sufficient samples, insufficient samples, and different training methods. The results show that the multimodal UAV communication composite interference recognition method proposed in the present invention is superior to the single-modal signal method. Under the condition of low interference-to-noise ratio of INR≤2dB, the recognition accuracy of UAV communication composite interference of MIRNet is higher than that of the baseline model, and the model requires the least number of parameters. In the final ablation experiment, it is verified that the design of asymmetric deep convolution, frequency domain adaptive filtering module, and multi-threshold optimization algorithm can effectively improve the recognition accuracy of the algorithm. The method proposed in the present invention mainly focuses on the labeled UAV communication composite interference dataset. Future research can further explore how to improve the recognition accuracy of the model for UAV communication composite interference in an unlabeled small sample environment.

[0042] Third, the expected benefits and commercial value of the technical solution of the present invention after transformation are as follows: The multi-modal UAV communication composite interference identification method based on multi-label learning of the present invention can improve the recognition accuracy of composite interference signals under low interference-to-noise ratio conditions, and can better adapt to the needs of actual wireless communication composite interference signal identification tasks. Secondly, the present invention can reduce the dependence on data, and can extract the characteristic information of composite interference signals from limited training samples, thereby reducing the dependence on a large number of training samples, thereby obtaining better composite interference signal recognition effects even in scenarios with insufficient samples. At the same time, the present invention can improve the robustness of the system. By combining the two modes of time-frequency diagram and frequency domain data of the composite interference signal, it can better resist the influence of noise, improve the robustness of the system, and increase reliability in actual complex environments.

[0043] Wireless communication composite interference signal recognition has broad application prospects in the fields of military, intelligent communication, radar, etc. This invention can be applied to various types of UAV communication composite interference signals and can be used in multiple application scenarios. While improving the accuracy of composite interference signal recognition, it reduces the requirements for the amount of training set sample data, saving a lot of data collection and labeling costs. At the same time, the designed MIRNet composite interference recognition model is lightweight, reducing the requirements for hardware resources.

[0044] The technical solution of the present invention fills the technical gap in the industry at home and abroad: the current research methods are often limited by the strong interference of noise, making it difficult to accurately distinguish interference signals from background noise, and usually rely on a large amount of labeled data to obtain better performance. This technical solution improves the ability to identify composite interference signals in scenarios with low interference-to-noise ratios and insufficient samples by introducing multimodal learning. Compared with traditional recognition methods, this solution fills the gap in the recognition technology of composite interference signals in wireless communications in scenarios with low interference-to-noise ratios and insufficient samples in the domestic and foreign industries, improves the recognition accuracy of the model and the reliability of the system, and has great application value and promotion prospects.

[0045] The technical solution of the present invention solves a technical problem that people have been eager to solve but have never succeeded in solving: the problem of low signal recognition accuracy under low interference-to-noise ratio: under low interference-to-noise ratio conditions, traditional signal recognition methods are affected by noise interference and it is difficult to effectively separate interference signals from noise, resulting in low recognition accuracy. By designing an adaptive filtering module, the global feature extraction capability of the signal is enhanced, and the recognition accuracy of the system under low interference-to-noise ratio conditions is effectively improved, which solves this technical problem to a certain extent.

[0046] Small sample identification problem: In some practical applications, UAV communication composite interference signals belong to rare categories, and existing single-modal composite interference identification methods have poor recognition effects on these categories. The multi-modal UAV communication composite interference identification method based on multi-label learning can extract more feature information from the composite interference signal by combining the signal's time-frequency diagram and frequency domain data, which can improve the accuracy of composite interference identification in scenarios with insufficient samples.

[0047] The technical solution of the present invention overcomes technical bias: the multi-modal UAV communication composite interference identification method based on multi-label learning can overcome some technical biases to a certain extent. By combining the two modal information of the time-frequency diagram and frequency domain data of the interference signal, improving the model design, and reducing the dependence on a large amount of training sample data, the method can improve the recognition accuracy of UAV communication composite interference signals under low interference-to-noise ratio, reduce bias against specific categories, and enhance the robustness and versatility of the system.

[0048] Fourth, the present invention provides an effective method for identifying composite interference in wireless communications under low signal-to-noise ratio conditions, solving the technical problem that the existing technology is difficult to accurately identify interference sources in complex UAV communication environments. This method uses advanced deep learning technology, especially a lightweight neural network called MIRNet, to accurately extract and fuse the input IQ sequence data, significantly improving the accuracy and efficiency of interference identification.

[0049] First, the technical solution of this method includes four steps: preprocessing of composite interference signals, feature extraction, feature fusion, and probability mapping. In the preprocessing stage, the representation of input data is optimized through the conversion of time-frequency diagrams and frequency domain data, laying the foundation for subsequent in-depth analysis. Next, two independent network branch structures are used in MIRNet to conduct in-depth learning of data of different modalities, realizing a comprehensive analysis of interference signals.

[0050] The feature fusion stage is the core of the technical solution. Through the connection operation, the features of the time-frequency diagram and the frequency domain data are merged, and the powerful ability of the deep network structure is used to interact between the modalities. This innovative fusion method significantly improves the accuracy of interference identification. In the final stage, the network output is converted into a specific interference type probability through the sigmoid function and multi-threshold optimization algorithm, accurately indicating the possibility of each interference.

[0051] In addition, the FAF module in MIRNet enhances feature expression by introducing Fourier transform processing, which not only improves the model's sensitivity to frequency domain information, but also optimizes the global attention of features through the CBAM module, further enhancing the network's ability to capture key information in interference signals.

[0052] In summary, this recognition method has obvious technical progress. It effectively solves the challenge of identifying composite interference in UAV communication in complex low interference-to-noise ratio environments through efficient neural network architecture and innovative algorithm design. This method not only improves the accuracy and real-time performance of recognition, but also has strong practicality and broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a flow chart of a method for identifying composite interference in wireless communication applicable to low interference-to-noise ratio conditions provided by an embodiment of the present invention;

[0054] Figure 2 It is a framework diagram of a multi-modal UAV communication composite interference identification method based on multi-label learning provided by an embodiment of the present invention;

[0055] Figure 3 is a schematic diagram of MIRNet provided in an embodiment of the present invention; GAP represents global average pooling, and C represents the number of convolution kernels;

[0056] Figure 4 is a schematic diagram of asymmetric depth convolution provided by an embodiment of the present invention;

[0057] Figure 5 Schematic diagram of an EConv module and a downsampling module provided in an embodiment of the present invention; (a) EConv; (b) downsampling;

[0058] Figure 6 is a schematic diagram of a FAF module provided by an embodiment of the present invention; wherein E represents the ratio of the number of output channels to the number of input channels;

[0059] Figure 7 is a composite interference recognition accuracy diagram of different modal recognition methods provided by an embodiment of the present invention;

[0060] Figure 8 is a comparison diagram between an embodiment of the present invention and other composite interference identification methods;

[0061] Fig. 9 : This is a comparison chart of different modality recognition accuracy in the scenario of insufficient samples provided by the embodiment of the present invention; (a) 5% sample size, (b) 3% sample size, (c) 1% sample size;

[0062] Fig.10 : is a performance comparison diagram of the embodiment of the present invention in the scenario of insufficient samples and the comparison model; (a) 5% sample size; (b) 3% sample size; (c) 1% sample size;

[0063] Fig.11 1 is an evaluation index diagram of different models in the scenario of insufficient samples provided by the embodiment of the present invention; (a) 5% sample size; (b) 3% sample size; (c) 1% sample size;

[0064] Fig.12 It is a diagram of recognition accuracy of different models using two training strategies provided by an embodiment of the present invention;

[0065] Fig.13 This is a composite interference recognition accuracy graph of 1% training sample data provided by an embodiment of the present invention;

[0066] Fig.14 It is a time-frequency diagram of composite interference of UAV communication with INR=30dB provided by an embodiment of the present invention;

[0067] (a) BPSK; (b) ST+BPSK+SSM+LSM; (c) BPSK+LSM+NPM; (d) BPSK+LSM+SSM+NPM. DETAILED DESCRIPTION

[0068] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0069] like Figure 1 As shown, an embodiment of the present invention provides a method for identifying composite interference in wireless communication applicable to low interference-to-noise ratio conditions, comprising the following steps:

[0070] Step 1: preprocess the IQ sequence of the input UAV communication composite interference to obtain two modal data: time-frequency diagram and frequency domain data;

[0071] Step 2: Build a lightweight neural network MIRNet and extract features from two independent input data respectively;

[0072] Step 3: Use the connection operation to jointly process the input features of the two modalities to generate multi-modal joint features and achieve feature fusion;

[0073] Step 4: Use the sigmoid function to map the output of the FC (fully connected) layer to the probability corresponding to the components of the composite interference signal of drone communication, and apply the multi-threshold optimization algorithm to obtain the predicted category of the interference component contained in the composite signal, and calculate the final loss through the binary cross entropy loss function.

[0074] MIRNet's structural design

[0075] MIRNet is composed of two branch networks with similar structures, which aims to extract deep features from the two modes of composite interference signals of drone communications. The specific structural details are as follows: Figure 3 As shown in the figure. The main modules of MIRNet include EConv (extraction and squeeze Convolution) module, D-FAF (downsampling-frequency adaptive filter) module composed of downsampling module and frequency domain adaptive filter (FAF) module, and feature fusion module. First, asymmetric convolution (AConv), EConv, CBAM and downsampling modules are cascaded to extract shallow features of input data. Secondly, three D-FAF modules are stacked to extract deep features of time-frequency graph, and one D-FAF module is used to extract deep features of frequency domain data. Then, after global pooling operation, one-dimensional feature vectors of time-frequency graph branch network TFINet and frequency domain data branch network FDNet are obtained respectively. Finally, through the connection operation, the features of the two modes are spliced ​​and fused, and activated by sigmoid function to obtain the final probability output, which is convenient for the subsequent identification of composite interference in UAV communication.

[0076] 1) Asymmetric Deep Convolution (ADWC): This paper designs an asymmetric deep convolution module to capture subtle changes in input features and enhance robustness to rotational distortion. Figure 4As shown in the figure, during the training phase, the outputs of three parallel 3×3, 1×3 and 3×1 deep convolution kernels are added together, and more diverse features can be captured by convolving the input feature map in different directions. During the model inference phase, in order not to introduce additional computational complexity, the additivity of the convolution operation is used to superimpose the three convolution kernels and restore them to a square convolution kernel of the original size, and the fused deep convolution kernel parameters are used to reduce the computational overhead of the inference phase.

[0077] 2) EConv module and downsampling module: We improve the convolution calculation speed by making the convolution structure lightweight, thereby achieving the goal of lightweight network. We also use skip connections to alleviate the gradient vanishing or gradient exploding problems caused by the increase of network depth.

[0078] like Figure 5 As shown in (a), 1×1 convolution is first used to expand the number of channels, allowing more information to be captured in subsequent convolution processing. Secondly, asymmetric deep convolution is used to extract features in high-dimensional space. In the deep convolution, each channel only performs convolution operations with itself, which reduces the computational complexity and the number of model parameters. At the same time, the asymmetric design enables the convolution kernel to capture richer features in different directions. Finally, the high-dimensional feature space is reduced to the original number of channels through 1×1 convolution, ensuring the richness of feature representation while maintaining computational efficiency.

[0079] like Figure 5 (b) Setting the stride of the asymmetric depthwise convolution to 2 can further optimize the network computational efficiency while maintaining the richness of features.

[0080] 3)FAF module: Figure 6 As shown in the figure, the FAF module is composed of the LN (Layer Norm) layer, the Channel Conv module, and the adaptive filter (AF) module. As the core of the FAF module, the AF module first performs Fourier transform on the input features, then passes through the CBAM module, and finally performs inverse Fourier transform. That is, given the input features After being processed by the AF module, the output features are obtained The processing is expressed as:

[0081]

[0082] in and denote Fourier transform and inverse Fourier transform respectively, ⊙ denotes Hadamard product, and C(·) denotes CBAM. From formula (1), we can see that X FThe features of different spatial positions of correspond to different frequency components of X, which contain global information from X. The above processing can be regarded as the result of adaptive filtering of X.

[0083] The effectiveness of the adaptive filtering module is verified by deriving the following formula. Applying the convolution theorem to formula (3): time domain convolution is mathematically equal to its corresponding frequency domain Hadamard product, and the result is:

[0084]

[0085] According to (3) and (4), we can get the following relationship:

[0086]

[0087] in, The size of the tensor is consistent with the input feature X, and can be regarded as a dynamic convolution kernel that is as large as X in space. Mathematically, it is equivalent to using a large-size dynamic convolution kernel to mix the input features, and the convolution operation can be regarded as filtering the input features. In summary, the FAF module designed in this paper can perceive features globally and adaptively filter the input features by dynamically adjusting the parameters and shape of the convolution kernel, thereby effectively suppressing noise interference.

[0088] 4) D-FAF module: It consists of a downsampling module and a FAF module. The downsampling module is used to reduce the spatial dimension of the feature and effectively extract the local features of the input data. The FAF module adaptively utilizes the spatial context information and extracts the global features. By splicing local features and global features, the detail information and the overall context information can be effectively combined to obtain a more comprehensive feature representation.

[0089] 5) Multi-threshold optimization algorithm

[0090] Most existing supervised recognition algorithms treat different combinations of UAV communication composite interference signals as a new class, which leads to an exponential growth in the number of labels, and thus the number of model parameters. The application of multi-label learning can solve the above problems, and assigning an independent discrimination threshold to each component of the UAV communication composite interference signal can improve the recognition performance of the model.

[0091] Use the label vector Y of the training sample and the output probability vector P to optimize the multi-threshold vector V. According to the decision function, the output vector P of the neural network model is converted into [p 1 ,p 2 ,...p M ] is converted into a label decision vector Y = F(P) = [f(p 1 ),f(p2 ),...,f(p M )]. The decision function f(·) is expressed as

[0092]

[0093] where ε i represents the decision threshold of the i-th class label, ε i ∈(0,1). The threshold algorithm is regarded as an optimization problem. To find the best threshold, a direct search algorithm is used to calculate the optimal threshold for each label. Iterate with a certain step size within the given threshold range, and calculate the F1 score of each label under the current threshold. Finally, the threshold that maximizes the F1 score is selected as the optimal threshold for the label. That is, during the model training process, the threshold of the i-th class label obtained from L training samples is as follows:

[0094]

[0095] where y ji , Represents the training samples x j The true label and predicted label of , M represents the number of independent distractors.

[0096] 6) Feature fusion and loss function

[0097] The feature fusion operation occurs before the fully connected layer. This deep fusion allows the features extracted by the two networks to be more closely combined to form a richer feature representation. FD Indicated as M i =[m 1 ,m 2 ,…,m i ], the output feature vector F of TFINet TFI Indicated as N j =[n 1 ,n 2 ,…,n j ], the joint feature can be represented as a stack of two one-dimensional vectors F FD and F TFI , the feature fusion process is expressed as:

[0098]

[0099] Where ⊕ represents the connection function.

[0100] The goal of network training is to optimize the network weights and biases by minimizing the loss between the actual output and the training data labels. In the MIRNet model, after the last fully connected layer, a sigmoid function is used as the output function, which limits the range of the output probability to the (0,1) interval.

[0101] The loss function used in the present invention is the binary cross entropy loss function, which calculates the L sample output labels independently:

[0102]

[0103] in, represents the predicted label after multi-threshold optimization, y (i) Indicates the actual label. The Adam optimizer is used in the training process, and the weights and biases are adjusted iteratively using the gradient of Loss.

[0104]

[0105] in, represents the kth trainable weight in the l+1 layer, is the corresponding bias and η is the step size.

[0106] The MIRNet model training process is shown in Algorithm 1.

[0107]

[0108] like Figure 2 As shown, an embodiment of the present invention provides a method for identifying composite interference in wireless communication applicable to low interference-to-noise ratio conditions, comprising the following steps:

[0109] The data preprocessing module is used to preprocess the IQ sequence of the input UAV communication composite interference to obtain two modal data: time-frequency diagram and frequency domain data;

[0110] The feature extraction module is used to build a lightweight neural network MIRNet and extract features from two independent input data respectively;

[0111] The feature fusion module is used to jointly process the input features of the two modalities using a connection operation to generate multi-modal joint features and realize feature fusion;

[0112] The signal recognition module is used to map the output of the fully connected (FC) layer to the probability corresponding to the components of the composite interference signal of UAV communication using the sigmoid function, and apply the multi-threshold optimization algorithm to obtain the predicted category of the interference component contained in the composite signal, and calculate the final loss through the binary cross entropy loss function.

[0113] 1. Specific application fields or related products of the present invention.

[0114] The present invention can be widely used in UAV communication systems, especially in complex electromagnetic environments and low interference-to-noise ratio (INR) conditions, and can effectively identify various complex interference signals and improve the stability and security of communications. The technology can be applied to communications between UAVs and ground stations, and between UAVs and other aircraft, to ensure the reliability and anti-interference capability of communication links.

[0115] The present invention also has important application value in the field of electronic warfare (EW). Through the effective identification of composite interference signals, the technology of the present invention can be applied to electronic warfare equipment to enhance the ability to perceive and respond to enemy interference signals and improve the effect of electronic warfare.

[0116] 2. Relevant evidence of the technical effects obtained by the embodiments of the present invention.

[0117] The signal model includes two modal forms: a model of the composite interference signal of UAV communication and a sequence and image of the composite interference signal of UAV communication.

[0118] 1. Composite interference signal model for UAV communication

[0119] In the field of UAV communication, composite jammers are generated through tactical collaborative operations by independent jammers or multiple distributed jammers to interfere with the normal operation of the UAV communication system. This paper assumes that there are M independent jammers sending jammers at the same time, where each jammer is modulated with the same probability. The received discrete UAV communication composite jammer signal model can be expressed as:

[0120]

[0121] Among them, r(n) is the received UAV communication composite interference signal, * represents the convolution operation, x i (n) represents the interference signal sent by the i-th independent interferer, and h(n) is the channel impulse response corresponding to the interference signal. oi ,θ oi They represent the frequency deviation and phase deviation of the interference signal sent by the i-th independent jammer, T s represents the sampling period, w(n) is the complex-valued Gaussian white noise, and N is the number of sampling points. Typical types of UAV communication interference signals sent include single-tone interference (ST), binary phase shift keying (BPSK), noise phase modulation interference (NPM), linear swept frequency interference (LSM) and cosine swept frequency interference (SSM).

[0122] 1) Single tone interference:

[0123] x(t)=Aexp[j(2πf I t+θ I )] (13)

[0124] Where x(t) represents the received interference signal, f I and θ I represent the center frequency and initial phase of the interference signal respectively.

[0125] 2) BPSK interference: The interference signal that uses the same modulation method as the interfered signal is considered to be the best interference. Therefore, BPSK is used as one of the components of the composite interference signal of UAV communication.

[0126] x(t)=Aexp[j(2πf I t+θ P +θ I )] (14)

[0127] Among them, θ P is the modulation phase of BPSK, which is θ P =0, π represents bit 0 and bit 1 respectively, and θ P With θ I Independent

[0128] 3) Noise phase modulation interference: The amplitude, frequency and phase of the interference signal are random and irregular.

[0129] x(t)=Aexp[j(2πf I t+K PM u(t)+θ I )] (15)

[0130] Among them, K PM is the phase modulation index, u(t) has a mean of 0 and a variance of Gaussian noise modulated signal, effective phase shift

[0131] 4) Linear frequency sweep interference: The signal frequency of FM interference changes over time and is a non-stationary interference signal.

[0132]

[0133] Where T is the sweep period of the linear sweep, f H and f L Represents the start frequency and cutoff frequency respectively

[0134] 5) Cosine swept frequency interference: The frequency of the interference signal shows a cosine variation trend over time.

[0135]

[0136] Among them, K FM is the frequency modulation proportional constant, f m is the carrier frequency of the modulated signal, and the frequency modulation index β FM =K FM / f m In summary, the combination method M of the UAV communication composite signal received in this paper a There are 31 types in total.

[0137] 2. Two modal forms of composite interference signals in UAV communication

[0138] There are two modes of composite interference signals in UAV communication: frequency domain data (FD, fft data) and time frequency image (TFI, time frequency image).

[0139] 1) FD: The received UAV communication composite interference signal r(n) is the original IQ sequence expressed as:

[0140]

[0141] Among them, I(n)=real[r(n)],Q(n)=imag[r(n)],n=0,1,...,N-1,real(·) and imag(·) represent the real part and the imaginary part respectively. Performing discrete Fourier transform on r(n) yields R(k):

[0142]

[0143] The real and imaginary parts of the FFT data are expressed as: The frequency domain data of the composite interference of UAV communication with a signal length of N can finally be expressed as:

[0144]

[0145] 2) Time-frequency diagram

[0146] STFT and Gabor transform have limited performance due to the invariability of narrow windows in the time domain. Choi-Williams distribution (CWD) is an improved Wigner-Ville distribution (WVD) algorithm that has a better effect of suppressing cross-interference terms and can also reduce the impact of noise. The expression of CWD is:

[0147]

[0148] represents the instantaneous autocorrelation of the signal, τ represents the time shift, and σ represents the attenuation coefficient.

[0149] In order to facilitate subsequent processing and analysis, the time-frequency diagram of the composite signal of UAV communication is converted into a grayscale diagram to reduce the burden of the classifier. The time-frequency diagram of the composite interference signal of UAV communication is resized to 224×224 through the bilateral interpolation algorithm. The time-frequency diagram of the processed composite interference signal is shown in the figure below. Fig.14 shown.

[0150] Experiment and result analysis

[0151] Firstly, the simulation settings are given, including the composite interference signal dataset of UAV communication and the training parameter settings, and then several performance evaluation indicators of multi-label recognition models are defined. Finally, the performance of the proposed method is compared with the existing composite interference signal recognition methods, and degradation experiments are carried out.

[0152] 1 Dataset

[0153] Since there is currently no public dataset related to the identification of composite interference in UAV communications, this paper uses MATLAB to construct a UAV-CCI_31 (UAV-compoundcommunication interference) composite interference signal dataset for UAV communications under Gaussian white noise conditions. The composite interference signal is composed of 5 types of single communication interference signals additively composited, and the parameters are shown in Table 1. In order to ensure the diversity of samples, each interference signal is randomly generated within a specific parameter range, and there are 31 composite interference styles. The sampling frequency of the interference signal is set to 10MHz, and the time width is 102.4μs. The time and frequency smoothing windows in CWD both use hamming windows, with window lengths of 102 and 256, respectively. After time-frequency transformation and image processing, the size of each time-frequency graph is 224×224×1. The interference-to-noise ratio in this paper is defined as INR=10log 10 (∑ M P i / P N ), ranging from -10 to 30 dB, with an interval of 4. 100 frames of frequency domain data and 100 time-frequency diagrams were simulated and generated under each type of interference-to-noise ratio.

[0154] Table 1 Signal simulation parameter settings

[0155]

[0156] 2 Experimental setup

[0157] The sample number ratio of the training set, validation set, and test set for the scenario with sufficient samples is 5:1:4. All experiments are simulated under the PyTorch architecture, and the hardware device uses GPU P100. A total of 30 rounds of training are conducted, using a dynamic learning rate, with the maximum learning rate set to 0.003, the L2 coefficient to 2e-4, and the adam optimizer is used for iteration. The minimum number of training batches is 128.

[0158] 3 Evaluation indicators

[0159] 1) Overall accuracy (OA):

[0160]

[0161] in, represents the predicted label after multi-threshold optimization, y (i) represents the actual label, and L represents the number of samples. Z(·) is an indicator function, and its value is 1 when the label vector and the predicted result vector completely match. It is worth noting that the value of 0-1 loss is (1-OA).

[0162] 2) Precision

[0163]

[0164] 3) Recall:

[0165]

[0166] 4) F1 Score:

[0167]

[0168] 5) Hamming Score:

[0169]

[0170] 4 Simulation results analysis

[0171] Experiment 1: Composite interference recognition performance in a scenario with sufficient samples

[0172] In this section, we compare the recognition performance of composite interference in drone communications using multimodal and unimodal methods, and compare the recognition performance of the multimodal interference recognition model MIRNet proposed in this paper with other baseline models. In particular, MIRNet is degraded into MIRNet-T and MIRNet-F, which represent two unimodal methods, namely, using only time-frequency diagrams as input to TFINet and using only frequency domain data as input to FDNet. In addition, MIRNet-IQ represents a unimodal method that uses IQ data as input to FDNet. Unless otherwise specified, subsequent experiments in this paper are conducted over the entire interference-to-noise ratio range.

[0173] 1) Single-mode and multi-modal recognition methods: Figure 7 It can be observed from the accuracy curve of composite interference recognition of drone communication in that in the low interference-to-noise ratio range of INR≤2dB, MIRNet is improved compared with MIRNet-T based on time-frequency graph input. Specifically, when INR=-10dB, the recognition accuracy of MIRNet is 8.13% higher than that of MIRNet-T. In the low interference-to-noise ratio range, the average recognition accuracy of MIRNet is 2.79% and 7.16% higher than that of MIRNet-T and MIRNet-F. This shows that MIRNet effectively utilizes the potential complementary information between the two modalities, thereby improving the interference recognition accuracy of the model in the low interference-to-noise ratio range. As the interference-to-noise ratio increases, the recognition accuracy curve of MIRNet tends to be stable. When INR>2dB, the recognition accuracy of MIRNet and MIRNet-T methods is similar, while the recognition accuracy of MIRNet-F is lower than both, which shows that the network can timely select important features from simple splicing and fusion features, thereby achieving a higher recognition accuracy.

[0174] 2) Comparison with other composite interference identification methods: Figure 8As shown. We observe that when INR≤2dB, the recognition accuracy of MIRNet is significantly higher than that of other comparison models. This shows that MIRNet has strong robustness in the low interference-to-noise ratio range. In particular, when INR=-10dB, the recognition accuracy of MIRNet is significantly higher than that of MLCNN by 24.9%. This shows that the local features learned by the stacked CNN module limit the recognition accuracy of the model for composite interference in drone communications. When INR>2dB, the recognition accuracy of MIRNet is higher than that of other baseline models. The average recognition accuracy of the lightweight models MobileNet v3 and MobileNet v4 is 2.0% and 2.1% lower than that of MIRNet. At the same time, it can be observed from the figure that the recognition accuracy of MobileVitv3 is second only to MIRNet, which shows that the D-FAF module we designed can effectively extract global information, thereby achieving recognition performance that is not inferior to the CNN+transformer model in the task of composite interference recognition in drone communications.

[0175] Table 2 Composite interference identification results (INR = -10:4:30dB)

[0176]

[0177] 3) Performance evaluation of the model: As shown in Table 2, we give the parameters of the baseline model and the evaluation indicators of the model. When the multimodal method proposed in this paper degenerates to MIRNet-T, which only uses a single time-frequency graph, it can be seen that the OA of MIRNet-T is at least 1.92% higher than the baseline model, indicating that the network structure designed in this paper can better extract and learn the features of the time-frequency graph. MIRNet achieves the global optimal interference recognition accuracy by combining the two modalities of frequency domain data and time-frequency graphs. On the basis of MIRNet-T, by introducing 0.086M model parameters, the OA of the model is improved by 1.02%.

[0178] In summary, MIRNet can learn the intrinsic complementary information of the two modalities of frequency domain data and time-frequency diagram from simple splicing and fusion features, thereby improving the recognition accuracy of composite interference signals of UAV communication under low interference-to-noise ratio conditions. The MIRNet designed in this paper has better recognition performance of composite interference in UAV communication than other baseline models, and the neural network model uses the least number of parameters, which is easy to deploy on UAVs.

[0179] Experiment 2: Composite interference recognition performance in the scenario with insufficient samples

[0180] This section discusses the impact of insufficient training samples on the recognition performance of the MIRNet network and compares it with the baseline model. 5%, 3%, and 1% of the data of each interference type are randomly selected from the dataset for training, that is, 5, 3, and 1 pictures or frame sequences are randomly selected for each type and each interference-to-noise ratio.

[0181] 1) Single-modal and multi-modal recognition methods: Fig. 9 Compared with the simulation results in (a), we can observe that as the number of samples decreases to 3% of the sample data, the recognition accuracy of the two methods using sequence modality decreases more significantly than that of the image modality method. Specifically, Fig. 9 As shown in (b), when INR = 30dB, when the multimodal method in this paper degenerates to MIRNet-F using only frequency domain data, the recognition accuracy of MIRNet-F drops from 68.88% to 47.64%, and the recognition accuracy of MIRNet-IQ method drops from 39.04% to 28.25%. When the multimodal method in this paper degenerates to MIRNet-T using only time-frequency graphs, the recognition accuracy of MIRNet-T drops from 88.38% to 84.16%. This means that the method using sequence training requires a large number of training samples to learn the mapping from input to output, resulting in a low recognition accuracy of drone communication interference composite recognition in scenes with insufficient samples. The time-frequency graph itself contains rich information, so the decline in recognition accuracy is small.

[0182] As the number of training samples is further reduced, only one image or one frame sequence is used for training for each category and each interference-noise ratio. Fig. 9 As shown in (c), when INR = 6dB, the recognition accuracy of MIRNet-T is 27.45% and 34.01% higher than that of MIRNet-F and MIRNet-IQ, respectively. Furthermore, it can be observed that when 1% of the training sample data is used, the recognition accuracy of the proposed multimodal MIRNet method is higher than that of MIRNet-T in the entire interference-to-noise ratio range. In particular, when INR = 30dB, it is 5.44% higher than MIRNet-T, which shows that in the case of very few samples, MIRNet can still improve the recognition accuracy of the model's UAV communication composite interference by fusing the features of the two modalities and mining the available features from them.

[0183] 2) Comparison with other composite interference identification methods: Fig.10As shown in (a), when the number of samples is 5%, the composite interference recognition accuracy curve of MIRNet in the entire interference-to-noise ratio range is higher than that of the baseline model. In particular, when INR≤2dB, the average interference recognition accuracy of MIRNet under low interference-to-noise ratio conditions is 56.15%, which is 32.62% higher than MLCNN and 5.35% higher than ShuffleNet. Fig.10 As shown in (b), when the number of samples is reduced to 3%, the recognition accuracy curve of MIRNet is higher than that of other baseline models in the entire interference-to-noise ratio range. When only 1% of the sample size is used for training data, it can be seen from Fig.10 (c) It can be seen intuitively that the recognition accuracy of the MIRNet designed in this paper is significantly higher than that of the comparison model. In the low interference-to-noise ratio range of INR≤2dB, the recognition accuracy of JRNet is second only to MIRNet, but its average recognition accuracy is 5.37% lower than that of MIRNet. When INR>2dB, the recognition accuracy of MobileVitv3 is second only to MIRNet, but its average accuracy is 12.25% lower than that of MIRNet.

[0184] 3) Performance evaluation of the model: Fig.11 The OA and other evaluation indicators of all models are shown when trained with 5%, 3%, and 1% sample sizes. It can be seen that as the amount of training sample data decreases, the difference in recognition performance between MIRNet and the baseline model increases significantly. This trend shows that MIRNet can effectively learn key features with very few training samples. Fig.11 As shown in (c), the F1-scores of MIRNet and Mobilevitv3 are 85.12% and 81.21% respectively. This shows that compared with Mobilevitv3 of CNN+transformers, the MIRNet designed in this paper is more suitable for the composite interference signal dataset of drone communication. At the same time, it also reflects that the D-FAF module designed in this paper can effectively integrate global features.

[0185] In summary, by fusing the time-frequency diagram and frequency domain data of the composite interference signal of UAV communication, even using one picture and one frame of frequency domain data, the MIRNet designed in this paper can still learn the subtle differences in different components of the composite interference signal of UAV communication within the low interference-to-noise ratio range, thereby improving the recognition accuracy of the composite interference signal of UAV communication while maintaining a small number of neural network model parameters.

[0186] Experiment 3: Composite communication interference recognition performance under different training methods

[0187] In order to verify the generalization ability of MIRNet for different training methods, a simulation experiment was conducted using the scenario with sufficient samples as an example. For a dataset of M candidate UAV communication composite interference signals, the number of signal components k in the received composite interference signal is in the range of [1, M]. In this paper, M is taken as 5. Two training strategies are adopted according to the value of the variable k, one is top-down and the other is bottom-up. The top-down training strategy establishes a new training sample dataset by gradually adding more low-level interference signal combinations. It can be obtained When n=1, the corresponding sample set contains the most interference signal combination types. n When training the data set, the ratio of training set, validation set and test set is 5:1:4. Similarly, the same division idea is used to construct the bottom-up data set.

[0188] from Fig.12 It can be seen that the MIRNet designed in this paper performs better than the baseline model in both training strategies. Specifically, for the top-down training strategy, in TD 5 In the dataset, the recognition accuracy of all models is 100%. 5 There are only five superposition forms of components in the dataset, that is, there is only one type of UAV communication composite interference training sample in the dataset. 4 When n is further reduced, the average recognition accuracy of all models decreases significantly, especially MLCNN, which drops from 100% to 33.72%. This is because the training set contains 6 combinations, and the data set at this time involves relatively complex interference signal combinations, which naturally leads to a decrease in the recognition accuracy of the model. As the value of n is further reduced, relatively simple interference signal combinations are introduced into the data set, which helps the neural network model learn the single component features of the composite interference signal combination, and on this basis, it can better understand the components of more complex UAV communication composite interference signals.

[0189] For the bottom-up training strategy, when BU is adopted 1 When n is used, the average recognition accuracy of MIRNet is 97.11%. This is because the data set contains only five single components of interference signals, which makes it easy for the model to distinguish and obtain a higher average interference recognition accuracy. As the value of n increases, the components of the communication interference signal affect each other, resulting in a slight decrease in recognition accuracy. However, when using BU 5 When the dataset is constructed, the combination of more UAV communication composite interference signals also makes it easier for the model to capture and learn mixed features, thereby improving the model's composite interference recognition accuracy.

[0190] Experiment 4: Ablation experiment

[0191] In order to verify the effectiveness of the asymmetric deep convolution, D-FAF module design and multi-threshold optimization algorithm designed in this paper, we conducted some degradation experiments. The experiment is divided into four cases: MIRNet-A (removing the input of the sequence mode), MIRNet-B (continue to remove the multi-threshold optimization algorithm), MIRNet-C (based on the previous step, replace the asymmetric deep convolution module with ordinary convolution), and MIRNet-D (continue to remove the D-FAF module).

[0192] like Fig.13 As shown in the figure, when using 1% of the number of training samples, we compared the recognition accuracy of drone communication composite interference under four different degradation conditions. In the case of multimodal input, the average recognition accuracy of MIRNet is 48.04%. When degenerated to MIRNet-A, the average recognition accuracy of the model dropped by 2.06%. Since the additional information gain brought by the frequency domain data cannot be utilized, the model can only rely on a single image modality to extract features, resulting in a decrease in recognition performance. When the multi-threshold optimization algorithm in the training process is removed, the average recognition accuracy of MIRNet-B drops by 1.94% compared with MIRNet-A. As the degradation conditions are further expanded, the recognition accuracy of the model drops significantly. Specifically, when MIRNet-B degenerates to MIRNet-C and MIRNet-D, the average recognition accuracy drops by 2.75% and 8.27%, respectively. When the asymmetric deep convolution module in the model is removed, the network lacks sensitivity to subtle differences, resulting in a weakened ability of the network to recognize details. In particular, when INR≤2dB, the average recognition accuracy of MIRNet-D decreased by 5.40% compared with MIRNet-C. When the D-FAF module was removed, the network further lacked the ability to perceive the global information of the composite interference signal of UAV communication. It was difficult to effectively screen and utilize the key features of the input composite communication interference signal by relying only on the extracted local features, resulting in a significant decrease in the recognition accuracy in the low interference-to-noise ratio range. These results show that the design of MIRNet and the training process of the multi-threshold optimization algorithm can effectively explore and utilize the feature information of different modes of composite interference in UAV communication, thereby improving the recognition accuracy of composite interference signals for UAV communication under low interference-to-noise ratio.

[0193] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. It can be understood by a person of ordinary skill in the art that the above-mentioned devices and methods can be implemented using computer executable instructions and / or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0194] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. A method for identifying composite interference in wireless communication under low interference-to-noise ratio conditions, characterized in that: The following steps are involved: Step 1: preprocess the IQ sequence of the input UAV communication composite interference to obtain two modal data: time-frequency diagram and frequency domain data; Step 2: Build a lightweight neural network MIRNet and extract features from two independent input data respectively; Step 3: Use the connection operation to jointly process the input features of the two modalities to generate multi-modal joint features and achieve feature fusion; Step 4: Use the sigmoid function to map the output of the fully connected layer to the probability corresponding to the components of the composite interference signal of drone communication. Apply the multi-threshold optimization algorithm to obtain the predicted category of the interference component contained in the composite interference signal, and calculate the final loss through the binary cross entropy loss function.

2. The wireless communication composite interference identification method applicable to low interference-to-noise ratio conditions according to claim 1, characterized in that: MIRNet is composed of two branch networks with similar structures, including the EConv module, the D-FAF module consisting of the downsampling module and the FAF module, and the feature fusion module; Firstly, asymmetric convolution, EConv, CBAM and downsampling modules are cascaded to extract shallow features of input data. Secondly, three D-FAF modules are stacked to extract deep features of time-frequency graphs, and one D-FAF module is used to extract deep features of frequency domain data. Then, after a global pooling operation, one-dimensional feature vectors of the time-frequency graph branch network TFINet and the frequency domain data branch network FDNet are obtained respectively. Finally, the features of the two modes are concatenated and fused through a connection operation, and activated using the sigmoid function to obtain the final probability output, which is convenient for the subsequent identification of composite interference in UAV communications.

3. The wireless communication composite interference identification method applicable to low interference-to-noise ratio conditions according to claim 2, characterized in that: The FAF module is composed of the LN layer, the Channel Conv module, and the AF module in sequence; the LN layer is used to alleviate the gradient vanishing problem caused by the depth of the neural network; the Channel Conv is used to improve the model's ability to capture feature details; the AF module, as the core of the FAF module, first performs Fourier transform on the input features, then passes through the CBAM module, and finally performs inverse Fourier transform; that is, given a feature After being processed by the AF module, The processing is expressed as: in and denote Fourier transform and inverse Fourier transform respectively, ⊙ denotes Hadamard product, and C(·) denotes CBAM. From the above formula, we know that X F The features at different spatial positions of correspond to different frequency components of X, which contain the global information from X.

4. The wireless communication composite interference identification method applicable to low interference-to-noise ratio conditions according to claim 2, characterized in that: For TFINet, the FAF module completes the global feature extraction of the TFI features of the UAV composite communication interference signal, that is, the feature map output by the LN layer is converted from the spatial domain to the frequency domain after Fourier transform; secondly, the CBAM module is used to adaptively adjust the attention of the feature map to enhance the expression ability of the feature; finally, IFFT is used to complete the conversion between the frequency domain and the spatial domain, and complete the adaptive filtering of the input features; For FDNet, the FAF module can enhance the expressiveness of feature representation and extract time series correlation from the composite communication interference signal of UAV.

5. The wireless communication composite interference identification method applicable to low interference-to-noise ratio conditions according to claim 2, characterized in that: The D-FAF module consists of a downsampling module and a FAF module. The downsampling module is used to reduce the spatial dimension of features and effectively extract local features of input data. The FAF module adaptively utilizes spatial context information, which helps to understand long-distance dependencies in features and extract global features. By splicing local features and global features, detail information and overall context information are effectively combined to obtain a more comprehensive feature representation.

6. The wireless communication composite interference identification method applicable to low interference-to-noise ratio conditions according to claim 1, characterized in that: In step 3, the output feature vector F of FDNet is FD Indicated as M i =[m1,m2,...,m i ], the output feature vector F of TFINet TFI Indicated as N j =[n1,n2,...,n j ], the joint feature is represented by stacking two one-dimensional vectors F FD and F TFI , the feature fusion process is expressed as follows: in Represents the connection function.

7. The wireless communication composite interference identification method applicable to low interference-to-noise ratio conditions according to claim 1, characterized in that: The multi-threshold optimization algorithm in step 4 optimizes the multi-threshold vector V based on the input label vector Y of the training sample X and the corresponding output probability vector P; the output vector P of the neural network model = [p1, p2, ... p M ] According to the decision function F(·), it is converted into a label decision vector Y=F(P)=[f(p1),f(p2),...,f(p M )]; the decision function f(·) is expressed as: where ε i represents the decision threshold of the i-th class label, ε i ∈(0,1).

8. The wireless communication composite interference identification method applicable to low interference-to-noise ratio conditions according to claim 7, characterized in that: The optimal threshold of each label is calculated by direct search algorithm; it is iterated with a certain step size within the given threshold range, and the F1 score of each label under the current threshold is calculated; finally, the threshold that makes the F1 score reach the maximum is selected as the optimal threshold of the label; that is, in the model training process, the threshold of the i-th label obtained from L training samples is as follows: where y ji , Represents the training samples x j The true label and predicted label of . M represents the number of independent distractors.

9. The wireless communication composite interference identification method applicable to low interference-to-noise ratio conditions according to claim 1, characterized in that: The loss function in step 4 is a binary cross entropy loss function, which calculates the L sample output labels independently: in, represents the predicted label after multi-threshold optimization, y (i) Indicates the actual label; Adam optimizer is used in the training process, and the weights and biases are adjusted iteratively using the gradient of Loss; in, represents the kth trainable weight in the l+1 layer, is the corresponding bias and η is the step size.

10. A method for identifying composite interference in wireless communication under low interference-to-noise ratio conditions according to any one of claims 1 to 9, characterized in that: The following steps are involved: The data preprocessing module is used to preprocess the IQ sequence of the input UAV communication composite interference to obtain two modal data: time-frequency diagram and frequency domain data; The feature extraction module builds a lightweight neural network MIRNet and extracts features from two independent input data respectively; The feature fusion module is used to jointly process the input features of the two modalities using a connection operation to generate multi-modal joint features and realize feature fusion; The signal recognition module is used to map the output of the fully connected layer to the probability corresponding to the components of the composite interference signal of the drone communication using the sigmoid function. The multi-threshold optimization algorithm is applied to obtain the predicted category of the interference component contained in the composite signal, and the final loss is calculated through the binary cross entropy loss function.

Images