An interference identification and suppression method based on CBMA-MU-Net

Abstract

The application discloses a kind of interference identification and suppression method based on CBMA-MU-Net, comprising: the data preprocessing operation of original signal is separated into odd and even frequency points, and the even frequency point data and odd frequency point data of original data are obtained;CBAM-M-Net interference identification network is constructed, and the time-frequency matrix of even frequency point data is input into CBAM-M-Net interference identification network to obtain interference identification result, CBAM-M-Net interference identification network includes Mobile-Net model and CBAM module;Interference identification result and the time-frequency matrix of even frequency point data are input into U-Net interference prediction network simultaneously, and interference signal reconstruction is completed;By receiving end original odd frequency point signal and reconstructed interference signal difference, interference cancellation is realized.The application effectively combines the characteristics of transmission system and the advantages of deep learning, realizes the identification and reconstruction of communication system interference signal based on image recognition technology, and completes effective interference identification and suppression.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of interference identification, and particularly relates to an interference identification and suppression method based on CBMA-MU-Net. BACKGROUND

[0002] The ionospheric scattering channel has strong invulnerability and can realize ultra-long distance information transmission of up to thousands of kilometers, and has extremely important application in the field of military communication. However, the ionospheric scattering channel has large loss and serious spatial, temporal and frequency selective fading problems, which limit its practical application and promotion. In view of the non-linear characteristics and large loss of the ionospheric scattering channel, the existing research has verified that the SC-IFDMA (Single-carrier Frequency-Division Multiple Access) transmission system adopts frequency domain equalization to effectively resist channel multipath fading, and has low peak-to-average ratio, which is the best choice for adapting to the ionospheric scattering channel.

[0003] At present, the research on interference of the ionospheric scattering system is very rare, and the existing technology is divided into two parts of interference identification and interference suppression, both of which contain traditional methods and processing methods combined with neural networks.

[0004] In the interference identification research, the existing traditional methods mainly include interference identification methods based on decision tree and support vector machine. The method taking the decision tree as the classifier mainly completes feature extraction through the fractional Fourier transform (FRFT) and the periodogram method, and the feature extraction taking the support vector machine (SVM) as the system classifier mainly uses the K-means clustering algorithm. The subsequent research introduces the FastICA algorithm, mixes the modulated data signal and the system interference signal, uses FastICA to separate the mixed signal, and completes feature extraction and interference identification based on the separated signal. In the feature parameter extraction, the interference features are often extracted in the time-frequency spectrum, and the interference identification is carried out, and the subsequent research proposes an interference type automatic identification method based on the interference fractional Fourier transform spectrum feature parameter and using hierarchical decision. The subsequent extraction evolves to different signal domains.

[0005] The existing traditional interference identification method is very dependent on whether the feature value selection has representativeness, difference and separation. Different communication systems, different types of interference and different application environments all need specific feature value selection and classifier design. For the ionospheric scattering SC-IFDMA system, the traditional method is difficult to select interference feature parameters with separation, and because of the complex channel environment, the extracted interference features are not complete, and the effective information loss rate is very high, so the traditional interference identification effect is very unsatisfactory.

[0006] The research of interference recognition based on deep learning mainly focuses on signal data enhancement, input format selection and deep learning network construction. There are existing researches on interference recognition based on deep learning for three kinds of wireless interference signals of Bluetooth, Zigbee and Wifi, including performance analysis and optimization of four networks based on CNN, ResNet, convolutional long short-term memory deep neural network (CLDNN) and long short-term memory network (LSTM); followed by energy detection based on system level combined with deep learning network recognition algorithm, Inception recognition network, DCNN interference network with interference spectrum as network input, recognition algorithm for adaptively extracting deep features of interference time-frequency spectrum, network algorithm for modulated signal recognition using graph convolution network, classification recognition algorithm introducing attention module network, and CCNN, bidirectional RNN and ACGAN network.

[0007] Through simulation verification, the performance of the algorithm based on deep learning network is better than that of the traditional method on a specific data set. However, the communication environment applied in the existing interference recognition method based on deep learning does not have the characteristics of ionospheric scattering channel, and cannot be directly generalized.

[0008] Traditional interference recognition methods need to select characteristic parameters with high separation, but in scattering communication, the channel environment is complex and the signal components are mixed, making it difficult to select characteristic parameters that can easily represent the characteristics of interference. Moreover, the extraction of traditional algorithms is mostly shallow and has weak generalization ability. In the case of multiple types of interference signals and similar intensity of interference and noise, traditional interference recognition algorithms will not work. However, the current interference recognition method based on deep learning is only applicable to pure interference without signal, and cannot be applied to ionospheric scattering SC-IFDMA systems with signal and interference coexisting.

[0009] In the interference cancellation research, the traditional method includes: using linear FIR filter to fit the link characteristics, realizing the minimum LMS output of the system, and completing the elimination of the interference signal of the system; dividing the access users into odd and even groups, and processing the data of the two groups of users by dynamic SIC multi-user detection, which slightly improves the interference cancellation performance but has high complexity and high implementation cost; introducing preambles and postambles, using EM algorithm to complete parameter estimation, combining signal decision and channel parameter estimation to complete interference signal reconstruction, and realizing interference cancellation, which can only be applied to time-invariant channels and depends on signal recognition and reconstruction results.

[0010] There is no systematic research on interference cancellation for ionospheric scattering SC-IFDMA systems, traditional interference recognition methods have no adaptive interference parameters, decision thresholds have no interference experimental detection and simulation data, and conventional interference cancellation algorithms have great difficulty in migrating in scattering channel conditions, making it difficult to maintain interference cancellation performance.

[0011] The interference cancellation method based on deep learning is less researched in the field of ionospheric scatter communication. The related researches are concentrated in the field of radar interference, and the achievements include: a radar gesture interference recognition and suppression method based on frequency modulation continuous wave (FMCW), application of effective attitude distance and Doppler parameter estimation, gesture feature extraction and classification of radar interference signals by using a TS-I3D network; a radio frequency interference suppression network for a synthetic aperture radar system; an ISRNet network architecture, an interference suppression and recognition network for interference suffered by a synthetic aperture radar.

[0012] The above interference cancellation methods are deeply combined with the characteristics of radar communication systems, have special application, and cannot be generalized to ionospheric scatter SC-IFDMA systems, but the core idea of the significant research achievements of the interference suppression network combined with the application of deep learning in the radar communication environment can be further abstracted and systematized. It can be predicted that there are more research possibilities for interference cancellation combined with deep learning. SUMMARY

[0013] In view of the problems that the existing interference recognition and suppression methods cannot select interference characteristic parameters with high separation in the ionospheric scatter channel, the interference signal feature extraction is incomplete, the effective information loss rate is high, the interference recognition and suppression effect is not ideal, and some methods cannot be generalized, after the ionospheric scatter channel environment and the SC-IFDMA transmission system are determined, a deep learning-based interference recognition and suppression method is proposed by combining the transmission system characteristics and the deep learning network. The technical problems to be solved by the present application are realized by the following technical solutions:

[0014] The present application provides a CBMA-MU-Net-based interference recognition and suppression method, comprising:

[0015] S1: performing data preprocessing operation of odd-even frequency point separation on the original signal to obtain time-frequency matrix of even frequency point data of the original data;

[0016] S2: constructing a CBAM-M-Net interference recognition network, and inputting the time-frequency matrix of the even frequency point data into the CBAM-M-Net interference recognition network to obtain an interference recognition result, wherein the CBAM-M-Net interference recognition network comprises a Mobile-Net model and a CBAM module;

[0017] S3: inputting the interference recognition result and the time-frequency matrix of the even frequency point data into a U-Net interference prediction network at the same time to complete interference signal reconstruction;

[0018] S4: realizing interference cancellation by subtracting the reconstructed interference signal from the original odd frequency point signal at the receiving end.

[0019] In one embodiment of the present invention, the S1 includes:

[0020] Perform discrete Fourier transform on the original frequency domain signal and map it to subcarriers at uniform intervals. The signals are placed at intervals on the subcarriers. Assuming that the signals are mapped to odd subcarriers in the frequency point, there is no transmission signal data on the even subcarriers in the frequency domain, thereby obtaining a time-frequency matrix containing only even frequency point data of the interference signal.

[0021] In one embodiment of the present invention, the Mobile-Net model includes a depth-separable convolution module, a pooling layer, a fully connected layer and a Softmax function layer connected in sequence.

[0022] In one embodiment of the present invention, the depthwise separable convolution module includes a 3×3 depthwise separable convolution layer, a first normalization layer, a first ReLU function layer, a 3×3 convolution layer, a second normalization layer, and a second ReLU function layer connected in sequence.

[0023] In one embodiment of the present invention, the CBAM module includes a channel attention unit and a spatial attention unit connected in sequence, which are used to perform attention operations in the channel and space dimensions respectively.

[0024] In one embodiment of the present invention, the channel attention unit is specifically used to:

[0025] The input feature map passes through the parallel maximum pooling layer and average pooling layer. The maximum pooling layer corresponds to the global maximum pooling and outputs F max C ; The average pooling layer corresponds to average pooling, output F avg C The output of the maximum pooling layer and the output of the average pooling layer are simultaneously input into the multi-layer perceptron to compress the number of channels and obtain two activation outputs through the ReLU activation function. The two activation outputs are summed element by element, activated by the Sigmoid function, and the CAM feature map M is output. C (F).

[0026] In one embodiment of the present invention, the spatial attention unit is specifically used to:

[0027] The result F′ of element-wise multiplication of the CAM feature map and the input feature map F of the channel attention unit is passed through the maximum pooling layer and the average pooling layer in sequence to obtain the feature map F max S With F avg S ; Output feature map F of the maximum pooling layer max S And the output feature map F of the average pooling layeravg S The splicing processing is performed, 7*7 convolution dimension reduction is utilized, a feature map with an output channel number of 1 is output, and then the Sigmoid function is activated to complete size conversion and output the SAM feature map M S (F′).

[0028] In an embodiment of the present application, the U-Net interference prediction network comprises a plurality of VGG modules, each VGG module is used for inputting different interference identification types, and interference prediction is performed according to the input interference identification type, and the predicted interference after the U-Net interference prediction network is continuously reduced with the actual odd frequency point interference error, and interference signal reconstruction is completed.

[0029] Compared with the prior art, the present application has the following beneficial effects:

[0030] 1. The present application proposes a data preprocessing method based on ionospheric scattering SC-IFDMA system and verified by simulation. Based on the SC-IFDMA system, the frequency domain resources are mapped to the subcarriers with uniform interval, and the frequency domain distribution characteristics of the data signals placed at intervals on the subcarriers are proposed. The data preprocessing of odd and even frequency point separation is carried out on the time-frequency signal, and the even frequency point data is used as the input of interference identification. The odd and even frequency point data of each interference type signal is extracted, and the spectrum characteristics are observed to verify the rationality of the preprocessing method. Only the signals on the even subcarriers are used for interference identification and prediction of the system, which is approximately equivalent to using only the data of the interference on the interval frequency points for interference identification and prediction, and has more accurate identification results compared with the case where the interference signal and the data signal coexist. At the same time, combining the data correlation of even frequency point interference and odd frequency point interference in a very narrow bandwidth, the even frequency point data obtained by preprocessing and interference identification results are used to predict the adjacent odd frequency point interference, and the reconstructed interference is formed by deep learning training. Finally, the system interference cancellation is realized by subtracting the odd frequency point data and the interference prediction data.

[0031] 2、The application provides a CBAM-M-Net-based interference identification method, which takes a pretreated time-frequency matrix as network input, takes Mobile-Net as a main interference identification framework, and fuses a CBAM module to improve identification efficiency. Related research results in the field of image identification are reasonably generalized, communication system characteristics are combined, and system processing of interference identification by means of image identification is realized. Simulation verification shows that the CBAM-M-Net interference identification performance is excellent (99% under a training set and 98% under a test set) for ten types of interference, including single-frequency interference, multi-frequency interference, sweep-frequency interference, noise amplitude modulation interference, noise frequency modulation interference, and a composite type obtained by combination of two types of interference. Meanwhile, performance comparison between the CBAM-M-Net interference identification network and interference identification algorithms based on a CNN network, a VGG network and a Mobile-Net is given, and the result shows that the interference identification algorithm based on time-frequency data preprocessing and the CBAM-M-Net has the optimal interference identification performance.

[0032] 3、The application provides a CBAM-MU-Net-based interference suppression algorithm. Based on the CBAM-M-Net interference identification network, a CBAM-MU-Net interference cancellation algorithm fusing data preprocessing is proposed by taking U-Net as a main interference prediction framework. The interference signal is output with a type identification result by the identification module, and is given different weights according to different interference types. A plurality of VGG networks are stored in the U-Net interference prediction network, different networks are used for interference prediction of different types of interference, the prediction interference after the network is continuously reduced, and the error between the actual odd frequency point interference and the prediction interference is continuously reduced, so that the interference signal is reconstructed. Finally, the interference cancellation of the system is realized by subtracting the reconstructed interference signal from the original odd frequency point signal at the receiving end. Simulation verification shows that the new interference cancellation method can effectively cancel a plurality of interference signals for the ionospheric scattering SC-IFDMA system. The whole interference identification and suppression model realizes interference signal identification of the communication system by means of image feature identification, completes interference signal reconstruction by means of a deep learning network, and gives a unique system data preprocessing based on a communication transmission system, so that the vacancy of the ionospheric scattering communication system interference research is made up.

[0033] The application will be further described in detail in combination with the drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 is a flowchart of a CBMA-MU-Net-based interference identification and suppression method provided by the application;

[0035] Figure 2 is an SC-IFDMA frequency domain subcarrier mapping schematic diagram provided by the application;

[0036] Figure 3 is a processing process schematic diagram of a CBAM-M-Net interference identification network provided by an embodiment of the application;

[0037] Figure 4 is a structure schematic diagram of a Mobile-Net model provided by an embodiment of the application;

[0038] Figure 5 is a structure schematic diagram of a deep separable convolutional layer provided by an embodiment of the application;

[0039] Figure 6 is a processing process schematic diagram of a traditional convolution and a deep separable convolution provided by an embodiment of the application;

[0040] Figure 7 is a structure schematic diagram of a CBAM module provided by an embodiment of the application;

[0041] Figure 8 is a structure schematic diagram of a CBAM-MU-Net provided by an embodiment of the application;

[0042] Figure 9 is a structure schematic diagram of a U-Net network;

[0043] Figure 10 is a structure schematic diagram of a U-Net network based on VGG-19 provided by an embodiment of the application;

[0044] Figure 11 is a single-frequency interference odd-even frequency point spectrum provided by an embodiment of the application;

[0045] Figure 12 is a three-frequency interference odd-even frequency point spectrum provided by an embodiment of the application;

[0046] Figure 13 is a narrow-band noise interference odd-even frequency point spectrum provided by an embodiment of the application;

[0047] Figure 14 is a linear sweep interference odd-even frequency point spectrum provided by an embodiment of the application;

[0048] Figure 15 is a narrow-band noise sweep interference odd-even frequency point spectrum provided by an embodiment of the application;

[0049] Figure 16 is a noise amplitude modulation interference odd-even frequency point spectrum provided by an embodiment of the application;

[0050] Figure 17 is a noise frequency modulation interference odd-even frequency point spectrum provided by an embodiment of the application;

[0051] Figure 18is a comparison curve of interference recognition rates obtained by different networks. DETAILED DESCRIPTION

[0052] In order to further illustrate the technical means and effects taken by the present application to achieve the predetermined object, the interference recognition and suppression method based on CBMA-MU-Net according to the present application is described in detail below in combination with the drawings and specific embodiments.

[0053] The foregoing and other technical contents, features and effects of the present application can be clearly presented in the detailed description of the specific embodiments below in combination with the drawings. Through the description of the specific embodiments, the technical means and effects taken by the present application to achieve the predetermined object can be more deeply and specifically understood, however, the accompanying drawings are provided for reference and illustration only, and are not intended to limit the technical solutions of the present application.

[0054] It should be noted that in this paper, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variants are intended to cover non-exclusive inclusion, so that the article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed. Without more limitation, the element defined by the sentence "including a…" does not exclude the presence of other identical elements in the article or device including the element.

[0055] Please refer to Figure 1 , Figure 1 is a flowchart of an interference recognition and suppression method based on CBMA-MU-Net provided by an embodiment of the present application. The interference recognition and suppression method comprises:

[0056] S1: data preprocessing operation of separating odd and even frequency points of the original signal, to obtain even frequency point data and odd frequency point data of the original data.

[0057] In combination with the frequency domain distribution characteristics of the SC-IFDMA system shown in Figure 2 , there are idle frequency points in the frequency domain, and interference information features can be obtained within the band, a new data preprocessing method is proposed at the system level: the odd and even frequency point data are pre-separated, and the time-frequency matrix of the even frequency points of the system is taken as the network input.

[0058] Specifically, the SC-IFDMA system performs DFT (Discrete Fourier Transform) transformation on the original signal, maps the frequency domain resources to subcarriers at uniform intervals, and places the data signals at intervals on the subcarriers. If it is assumed that the odd subcarriers in the frequency points are mapped, there will be no transmission signal data on the even subcarriers in the frequency domain. This data mapping characteristic makes the signal on the odd subcarriers in the system transmission process a superposition of transmission data and interference signals, while the even subcarriers only have interference signals (here, the influence of system noise is not considered).

[0059] For a conventional communication system, the greatest difficulty in interference identification is that the superposition of interference signals and transmission data signals makes it difficult to extract pure interference waveforms / characteristics, and the interference identification result is greatly affected by channel conditions and transmission signals. In the ionospheric scattering channel SC-IFDMA system, the signal on the even subcarriers is only generated and affected by the interference signal, and has a great theoretical basis advantage in interference characteristic analysis. From a theoretical point of view, if the data preprocessing operation of separating the odd and even frequency points of the time-frequency signal is performed, and only the data on the even subcarriers is used for interference estimation and prediction, it is approximately equivalent to using only the data of the interference signal on the interval frequency points for interference identification and prediction, and has a more accurate identification result than the case where the interference signal and the transmission signal coexist. In combination with the data correlation of the even frequency point interference and the odd frequency point interference within a very narrow bandwidth, the even frequency point data obtained by preprocessing and the interference identification result are used to predict the interference of the adjacent two sides of the odd frequency point, and even the interference prediction of the whole frequency band can be formed through network training. Finally, the system interference suppression is realized by subtracting the interference prediction data from the odd frequency point data.

[0060] In the interference identification research, based on the characteristics of the time-frequency spectrum of the interference signal and the data analysis ability of deep learning, a lightweight Mobile-Net is selected as the main body of the identification module, a CBAM module is introduced to improve the classification recognition rate, and a CBAM-M-Net interference identification network is proposed, and the algorithm flow block diagram is as shown in Figure 3

[0061] S2: Construct a CBAM-M-Net interference identification network, and input the time-frequency matrix of the even frequency point data into the CBAM-M-Net interference identification network to obtain an interference identification result, wherein the CBAM-M-Net interference identification network comprises a Mobile-Net model and a CBAM module.

[0062] ​The CBAM-M-Net interference identification network of the embodiment is for an ionospheric scattering SC-IFDMA system, and realizes the network architecture of interference identification of a communication system by combining a Mobile-Net model with a CBAM module (Convolutional Block Attention Module).

[0063] The Mobile-Net model is a light-weight, low-latency, high-precision, small-computing-parameter / parameter convolutional neural network model proposed by Google in 2017, which can be flexibly applied to mobile terminals and embedded devices. Compared with traditional neural network models, the Mobile-Net model only needs to use 1 / 32 of the parameters under the same data set. Please refer to Figure 4 , Figure 4 is a structural schematic diagram of a Mobile-Net model provided by an embodiment of the present application. The Mobile-Net model comprises a depthwise separable convolution module, a pooling layer, a full connection layer and a Softmax function layer connected in sequence.

[0064] Generally, a CNN architecture is composed of convolutional layers and full connection layers, and the main weights (main network parameters) are located in the full connection layers, and the main computing amount is located in the convolutional layers. In order to improve the network speed, the Mobile-Net model mainly cuts into the convolutional layers, and introduces a depthwise separable convolution (DSC). Please refer to Figure 5 , Figure 5 is a structural schematic diagram of a depthwise separable convolution layer provided by an embodiment of the present application. The depthwise separable convolution module comprises a 3x3 depthwise separable convolution layer, a first normalization layer, a first ReLU function layer, a 3x3 convolution layer, a second normalization layer and a second ReLU function layer connected in sequence. The processing processes of traditional convolution and depthwise separable convolution are compared as shown in Figure 6 The structural parameter table of the Mobile-Net model is shown in Table 1.

[0065] Table 1. Structural parameter table of the Mobile-Net model

[0066]

[0067]

[0068] Further, please refer to Figure 7 , Figure 7 is a structural schematic diagram of a CBAM module provided by an embodiment of the present application. The CBAM module comprises a channel attention (Channel Attention Module, CAM) unit and a spatial attention (Spatial Attention Module, SAM) unit connected in sequence, and is respectively used for performing attention operations in two dimensions of channels and spaces, i.e., the CAM performs channel attention, and the SAM performs spatial attention. Such a structure also provides convenience for module integration. The CBAM module combines the channel attention unit and the spatial attention unit, and can perform attention operations in two dimensions of space (Spatial) and channel (Channel). The module combines the attention mechanisms of the two units in accordance with the front-back order, simultaneously performs CAM (Class Activation Mapping) visualization, more directly pays attention to a target image, and effectively improves the feature extraction efficiency of a convolutional layer. Parameters of the CBAM module are shown in Table 2.

[0069] Table 2. CBAM module parameter table

[0070]

[0071] The CBAM module is added to a Mobile-Net interference recognition module, the input time-frequency graph size is 512x120, after the CBAM-M-Net network, the signal transverse dimension is 32, the longitudinal dimension is 12, and the channel direction dimension is 240.

[0072] As shown in Figure 7 , the input of the CBAM module is a feature graph F∈R C×H×W extracted by a deep separable convolutional layer of a Mobile-Net model. First, one-dimensional convolution M C ∈R C×1×1 is performed on the feature graph F by the CAM unit, and an output F' is obtained. F and F' are taken as inputs of the SAM unit, the SAM is two-dimensional convolution M S ∈R 1×H×W , and the module output is ( represents element-wise multiplication).

[0073] As shown in Figure 7As shown in the box a in the figure, the input feature map F first passes through the parallel maximum pooling layer and average pooling layer. The maximum pooling layer corresponds to the global maximum pooling, and the output F max C ; The average pooling layer corresponds to average pooling, output F avg C , the output size of both channels is 1×1×C. The outputs are simultaneously input into the network module composed of MLP (multi-layer perceptron). In this module, the number of channels is compressed to restore, and two activation outputs are obtained by ReLU activation function. The two results are summed element by element, and then activated by Sigmoid function to output the feature map M of the CAM unit. C (F), as shown in formula (1).

[0074]

[0075] Here, σ, W0, and W1 represent the weights of the activation function and the multilayer perceptron, respectively. The most significant feature of the CAM unit is the introduction of parallel pooling layers, which makes the extracted deep features more comprehensive and effective. In the CAM unit, the channel dimension remains unchanged, while the spatial dimension is compressed. The entire unit focuses on the basis for signal classification.

[0076] Further, if Figure 7 As shown in box b, the input of the SAM unit is the CAM output M C (F) The result F′ of element-by-element multiplication with the input feature map F is passed through the maximum pooling layer and the average pooling layer in turn to obtain F max S With F avg S , the size is 1×H×W. The output feature map F of the maximum pooling layer max S And the output feature map F of the average pooling layer avg S Perform splicing processing, use 7*7 convolution to reduce the dimension, output the feature map with a channel number of 1, and then activate it through the Sigmoid function to complete the size conversion and output the SAM feature map M S (F′), the processing process is shown in formula (2).

[0077]

[0078] Where σ represents the activation function. In the SAM unit, the spatial dimension does not change, the channel dimension is compressed, and the entire SAM unit focuses on the location information of the target signal.

[0079] The CAM unit and the SAM unit are arranged in a "CAM first and SAM then serial" order, a combination layer is formed by a convolution layer, a regularization layer, a ReLU layer, a CBAM layer and a maximum pooling layer, two combination layers are placed in succession in the system, and end-to-end training is carried out after integrating each convolution layer. The size of the input time-frequency diagram is 512*120, the CBAM-M-Net network, the signal horizontal dimension is 32, the vertical dimension is 12, and the channel direction dimension is 240.

[0080] S3: inputting the interference identification result and the time-frequency matrix of the even frequency point data into the U-Net interference prediction network simultaneously to complete interference signal reconstruction.

[0081] Please refer to Figure 8 , Figure 8 is a structure diagram of a CBAM-MU-Net provided by an embodiment of the application. The U-Net interference prediction network of the embodiment includes a plurality of VGG modules, each VGG module is used for inputting different interference identification types, and interference prediction is carried out according to the input interference identification type, and the predicted interference after the U-Net interference prediction network is continuously reduced, and interference signal reconstruction is completed.

[0082] Please refer to Figure 9 and Figure 10 , Figure 9 is a structure diagram of a U-Net network, Figure 10 is a structure diagram of a U-Net network based on VGG-19 provided by an embodiment of the application. The structure of the interference prediction based on the shallow U-Net network proposed in the embodiment uses a VGG-19 encoder to replace the original encoder, as shown in Figure 10 .

[0083] The search path of the U-Net is two VGG-19 convolution blocks, and each convolution layer in the network architecture uses ReLU as an activation function. The expansion path in the symmetric structure is also processed accordingly, and the convolution layer output is passed through ReLU and batch normalization. The convolution layer realizes global integration of information by means of continuous convolution and pooling.

[0084] The interference prediction network parameters of the U-Net network are shown in Table 3.

[0085] Table 3. U-Net interference prediction network parameter settings

[0086]

[0087] S4: realizing interference cancellation by subtracting the reconstructed interference signal from the original odd frequency point signal of the receiving end.

[0088] Specifically, the overall network takes a time-frequency diagram containing signal time-frequency domain characteristics as the network input, and completes the identification and prediction of the interference signal through image recognition and prediction. The preprocessed time-frequency matrix is taken as the network input and prediction label of the CBAM-M-Net interference identification network, and the correlation between the even-frequency point interference data and the odd-frequency point interference data is found by means of the prediction idea, and the interference type identification result is output. Then the interference identification result of the CBAM-M-Net interference identification network and the preprocessed data pass through the U-Net interference prediction network, and the interference identification result will be given different weights according to different interference types. There are multiple VGG modules in the U-Net interference prediction network, and different VGG modules are used for interference prediction of different types of interference, so that the prediction interference after passing through the VGG module continuously reduces the error with the actual odd-frequency point interference, and the interference signal reconstruction is completed. Finally, the original odd-frequency point signal and the reconstructed interference signal are subtracted at the receiving end to realize the interference cancellation of the system, and the whole process of interference cancellation is completed.

[0089] Further, the effect of the interference identification and suppression method based on CBMA-MU-Net is verified by simulation experiment.

[0090] (1) Data preprocessing verification

[0091] Table 4. Interference signal parameter table

[0092]

[0093] The time-frequency waterfall diagram of the seven interference signals after preprocessing is shown in Figures 11 to 17 The above seven groups of diagrams show that in the ionospheric scattering channel, the frequency domain distribution characteristics of the system receiving end signal after removing the cyclic prefix are obvious after transmission by the SC-IFDMA transmission system: the even frequency points are data idle frequency points, and only interference signals are present on them, and the diagram clearly shows the characteristics of the interference signal; the odd frequency points are frequency points containing effective data, and the diagram is the superposition of data and interference, and when the data signal is severely submerged by the interference. Using even frequency point data as network training input will be more beneficial to the identification and prediction of interference signals, and the preprocessing rationality is verified.

[0094] (2). CBAM-M-Net interference identification network performance

[0095] The performance comparison between the CBAM-M-Net interference identification network and the interference identification network based on VGG and Mobile-Net is given, and seven single-class interferences including single-frequency interference, three-frequency interference, linear sweep frequency interference, narrowband noise sweep frequency interference, narrowband noise interference, noise amplitude modulation interference and noise frequency modulation interference are selected for performance comparison.

[0096] Two indicators are selected to measure the performance of the algorithm: recognition accuracy and system time consumption. Recognition accuracy represents the percentage of correct samples in the total number of samples, which directly reflects the correct recognition rate of the interference recognition network; system time consumption refers to the time required for the system to recognize a single sample, which directly reflects the system performance. In addition to comparing different interference recognition networks, the recognition performance of the network before and after data preprocessing is also compared. The performance indicators of each network and input are compared as shown in Table 5, and the interference recognition rate curves are compared as shown in Figure 18

[0097] Table 5. Comparison of different network performances

[0098]

[0099] Based on the comparison of the table and the curve, the performance of the proposed interference recognition algorithm is summarized as the following three points:

[0100] 1. Comparing the data in the first and second rows of the table and the curves ① and ②, the introduction of the CBAM module effectively improves the feature expression ability of the network. Compared with the pure Mobile-Net model, the interference recognition accuracy is improved by 4.72%. And the introduction of CBAM has a very low loss, and the system time consumption only increases by 0.26ms.

[0101] 2. Comparing the data in the first and third rows of the table and the curves ① and ③, the data preprocessing method proposed in the embodiment of the application effectively improves the interference recognition accuracy. Compared with the case without preprocessing, the system interference recognition accuracy is improved by 7.19%.

[0102] 3. Comparing the data in the first, fourth and fifth rows of the table and the curves ①, ④ and ⑤, the CBAM-M-Net interference recognition algorithm proposed in this paper has a great performance improvement compared with the traditional neural network-based interference recognition algorithm. Compared with the traditional VGG interference recognition method, the interference recognition rate is improved by 16.05%, and the system time consumption is shortened by 2.24ms; compared with the Mobile-Net interference recognition method, the interference recognition rate is improved by 10.27%, and the system time consumption is only increased by 0.24ms.

[0103] (3). Performance of CBMA-MU-Net interference cancellation network

[0104] Table 6. Reduction of single-frequency interference power

[0105]

[0106] Table 7. Reduction of three-frequency interference power

[0107]

[0108] Table 8. Reduction of linear sweep interference power

[0109] ​

[0110] Table 9. Narrowband noise sweep interference power reduction

[0111]

[0112] Through simulation verification, the CBAM-MU-Net interference cancellation proposed in the application has excellent interference cancellation effect on single-frequency interference, multi-frequency interference, linear sweep interference and narrowband noise sweep interference.

[0113] The application proposes a data preprocessing method based on ionospheric scattering SC-IFDMA system and verifies it through simulation. Based on the SC-IFDMA system, the frequency domain resources are mapped to subcarriers at uniform intervals, and the frequency domain distribution characteristics of the data signal are placed at intervals on the subcarriers. The data preprocessing of separating the odd and even frequency points of the time-frequency signal is proposed, and the even frequency point data is used as the input of interference identification. The odd and even frequency point data of each interference type signal is extracted, and its spectrum characteristics are observed to verify the rationality of the preprocessing method. Only the signals on the even subcarriers are used for interference identification and prediction of the system, which is approximately equivalent to using only the data of the interference on the interval frequency points for interference identification and prediction. Compared with the case where the interference signal and the data signal coexist, it has more accurate identification results. At the same time, combined with the data correlation of the even frequency point interference and the odd frequency point interference within the extremely narrow bandwidth, the even frequency point data obtained by preprocessing and the interference identification results are used to predict the adjacent two sides of the odd frequency point interference, and the reconstructed interference is formed by means of deep learning training. Finally, the interference cancellation of the system is realized by subtracting the odd frequency point data and the interference prediction data.

[0114] Yet another embodiment of the present application provides a storage medium having stored therein a computer program for performing the steps of the CBMA-MU-Net based interference identification and suppression method described in the above embodiments. In yet another aspect, the present application provides an electronic device comprising a memory having stored therein a computer program and a processor which, when invoking the computer program stored in the memory, implements the steps of the CBMA-MU-Net based interference identification and suppression method described in the above embodiments. Specifically, the integrated modules implemented in the form of software functional modules described above can be stored in a computer readable storage medium. The software functional modules described above are stored in a storage medium, including a number of instructions for causing an electronic device (which can be a personal computer, a server, or a network device, etc.) or a processor to execute part of the steps of the method described in the various embodiments of the present application. The aforementioned storage medium includes various media that can store program codes, such as a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0115] The above is a further detailed description of the present application in combination with specific preferred embodiments, and it cannot be considered that the specific implementation of the present application is limited to these descriptions. For ordinary skilled persons in the technical field to which the present application belongs, a number of simple deductions or replacements can be made without departing from the concept of the present application, and all of them should be considered to fall within the protection scope of the present application.

Claims

1. A method for interference identification and suppression based on CBMA-MU-Net, characterized in that: include: S1: Perform data preprocessing to separate the odd and even frequency points of the original signal to obtain the time-frequency matrix of the even frequency point data of the original data; S2: Construct a CBAM-M-Net interference identification network, and input the time-frequency matrix of the even-frequency data into the CBAM-M-Net interference identification network to obtain an interference identification result. The CBAM-M-Net interference identification network includes a Mobile-Net model and a CBAM module; S3: Input the interference identification result and the time-frequency matrix of the even-frequency point data into the U-Net interference prediction network at the same time to complete the interference signal reconstruction; S4: Interference cancellation is achieved by subtracting the original odd frequency signal from the receiving end and the reconstructed interference signal; Said S1 comprises: Perform discrete Fourier transform on the original frequency domain signal and map it to subcarriers at uniform intervals. The signals are placed at intervals on the subcarriers. Assuming that the signals are mapped to odd subcarriers in the frequency point, there is no transmission signal data on the even subcarriers in the frequency domain, thereby obtaining a time-frequency matrix containing only even frequency point data of the interference signal.

2. The interference identification and suppression method based on CBMA-MU-Net according to claim 1, characterized in that The Mobile-Net model includes a depth-wise separable convolution module, a pooling layer, a fully connected layer, and a Softmax function layer connected in sequence.

3. The interference identification and suppression method based on CBMA-MU-Net according to claim 2, characterized in that: The depthwise separable convolution module includes a 3×3 depthwise separable convolution layer, a first normalization layer, a first ReLU function layer, a 3×3 convolution layer, a second normalization layer and a second ReLU function layer connected in sequence.

4. The interference identification and suppression method based on CBMA-MU-Net according to claim 3, characterized in that The CBAM module includes a channel attention unit and a spatial attention unit connected in sequence, which are used to perform attention operations in the channel and spatial dimensions respectively.

5. The interference identification and suppression method based on CBMA-MU-Net according to claim 4, characterized in that: The channel attention unit is specifically used to: The input feature map is passed through the parallel maximum pooling layer and average pooling layer. The maximum pooling layer corresponds to the global maximum pooling, and the output ; The average pooling layer corresponds to average pooling, output The output of the maximum pooling layer and the output of the average pooling layer are simultaneously input into the multi-layer perceptron to compress the number of channels and obtain two activation outputs through the ReLU activation function. The two activation outputs are summed element by element, activated by the Sigmoid function, and the CAM feature map is output. .

6. The interference identification and suppression method based on CBMA-MU-Net according to claim 5, characterized in that: The spatial attention unit is specifically used to: Combine the CAM feature map with the input feature map of the channel attention unit The result of element-wise multiplication , pass through the maximum pooling layer and the average pooling layer in turn, and obtain the feature maps ; Output feature map of the maximum pooling layer And the output feature map of the average pooling layer Perform splicing and use Convolution reduces the dimension and outputs a feature map with a channel number of 1. It is then activated by the Sigmoid function to complete the size conversion and output the SAM feature map. .

7. The interference identification and suppression method based on CBMA-MU-Net according to claim 5, characterized in that The U-Net interference prediction network includes multiple VGG modules, each of which is used to input a different interference identification type and perform interference prediction based on the input interference identification type, so that the error between the predicted interference after passing through the U-Net interference prediction network and the actual odd frequency point interference is continuously reduced, thereby completing the interference signal reconstruction.

Images