Neural network-based interference signal identification method, device, medium and apparatus
By using a neural network-based multimodal signal processing method, the problem of low accuracy in identifying interference signals in existing technologies has been solved, achieving higher identification accuracy and faster training speed.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIJING FORESTRY UNIVERSITY
- Filing Date
- 2024-03-27
- Publication Date
- 2026-07-07
AI Technical Summary
Existing interference signal identification methods use only single-mode data and ignore the characteristics of other mode data, resulting in low identification accuracy.
A neural network-based approach is adopted, which utilizes a parallel three-layer convolutional attention module, a multi-branch fusion module, a dual-branch hybrid parallel attention module, and a fully connected layer to process time-frequency images. It also combines a two-dimensional deep convolution module, a channel attention module, a spatial attention module, and a multi-head self-attention module to extract features of multimodal signals.
It improves the accuracy of interference signal recognition, enhances the model's classification performance and generalization ability, simplifies the model structure, and increases training speed.
Smart Images
Figure CN120730352B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of communication technology, and in particular to a method, apparatus, medium and device for identifying interference signals based on neural networks. Background Technology
[0002] In the field of modern wireless communication, with the rapid development of technology and the wide expansion of its application, communication systems are increasingly affected by various types of interference. These interferences come from diverse sources, including natural environmental noise, electromagnetic interference generated by equipment operation, and malicious human interference. This interference not only degrades communication quality but can also lead to communication outages, severely impacting the normal operation of commercial activities, security monitoring, and emergency rescue. Therefore, anti-interference technology has become a crucial technical means to ensure the stable operation of wireless communication systems.
[0003] The prerequisite for interference mitigation is the accurate and real-time classification and identification of interference signals in the received signal. Interference signal identification, as a crucial component of wireless communication interference mitigation, can support subsequent interference cancellation and suppression methods, thereby improving the performance of satellite communication systems.
[0004] Existing interference signal identification methods typically use different networks for different modal data; for example, Model 1 is used for time-domain data and Model 2 is used for frequency-domain data. However, this method only uses single-modal data during the identification process, ignoring features in other modal data, resulting in low identification accuracy. Summary of the Invention
[0005] This application provides a method, apparatus, medium, and device for identifying interference signals based on neural networks, to solve the problem of low accuracy in interference signal identification in related technologies. The technical solution is as follows:
[0006] According to a first aspect of this application, a method for identifying interference signals based on a neural network is provided, the method comprising:
[0007] The signal to be identified is obtained by superimposing at least a noise signal, a useful signal, and interference signals of n types, where n is an unknown number greater than or equal to 1.
[0008] Obtain a pre-trained neural network, wherein the neural network includes at least three parallel convolutional attention modules, a multi-branch fusion module, a dual-branch hybrid parallel attention module, and a fully connected layer; the convolutional attention module includes at least a two-dimensional deep convolution module, a channel attention module, a two-dimensional pointwise convolution module, and a spatial attention module; the multi-branch fusion module includes at least a two-dimensional convolution module, a compression and activation attention module, and an h-swish activation function; and the dual-branch hybrid parallel attention module includes at least a parallel multi-head self-attention module, a convolutional block attention module, a one-dimensional layer normalization layer, and a feedforward neural network.
[0009] The signal to be identified is converted into a corresponding time-frequency image;
[0010] The time-frequency image is processed using modules in the neural network to obtain n types of interference signals in the signal to be identified.
[0011] In one possible implementation, when the neural network includes three parallel convolutional attention modules, the process of using each module in the neural network to process the time-frequency image includes:
[0012] The time-frequency image is convolved using the two-dimensional depth convolution module to obtain a first intermediate result;
[0013] The channel attention module is used to perform channel attention calculation on the first intermediate result to obtain the second intermediate result;
[0014] The second intermediate result is convolved using the two-dimensional pointwise convolution module to obtain the third intermediate result.
[0015] The spatial attention module is used to perform spatial attention operations on the third intermediate result to obtain the output result of the convolutional attention module.
[0016] In one possible implementation, when the neural network includes a multi-branch fusion module and other modules, the processing of the time-frequency image using the modules in the neural network includes:
[0017] The outputs of the three parallel convolutional attention modules are concatenated into a feature map;
[0018] The feature map is convolved using the two-dimensional convolution module to obtain the fourth intermediate result;
[0019] The compression and activation attention module is used to perform attention operations on the fourth intermediate result to obtain the fifth intermediate result;
[0020] The h-swish activation function is used to process the fifth intermediate result to obtain the sixth intermediate result;
[0021] The sixth intermediate result is processed using other modules to obtain the output results of those other modules.
[0022] In one possible implementation, when the neural network includes a dual-branch hybrid parallel attention module and a fully connected layer, the processing of the time-frequency image using the modules in the neural network includes:
[0023] The one-dimensional normalization layer in the dual-branch hybrid parallel attention module is used to normalize the output results of the other modules. The normalization results are multiplied by three different weight matrices to obtain the query, key, and value. The multi-head self-attention module in the dual-branch hybrid parallel attention module is used to perform attention operations and concatenation on the query, the key, and the value to obtain the seventh intermediate result.
[0024] The channel attention module in the convolutional block attention module is used to perform channel attention operation on the output results of the other modules, and the spatial attention module in the convolutional block attention module is used to perform spatial attention operation on the operation results to obtain the eighth intermediate result;
[0025] The sum of the seventh intermediate result and the eighth intermediate result is normalized using the one-dimensional normalization layer to obtain the ninth intermediate result;
[0026] The feedforward neural network is used to process the ninth intermediate result to obtain the output result of the dual-branch hybrid parallel attention module;
[0027] The fully connected layer is used to calculate the output of the dual-branch hybrid parallel attention module to obtain n types of interference signals in the signal to be identified.
[0028] In one possible implementation, converting the interference signal into a corresponding time-frequency image includes:
[0029] The signal to be identified is converted into a corresponding time-frequency image based on the short-time Fourier transform; or...
[0030] The signal to be identified is converted into a corresponding time-frequency image based on wavelet transform.
[0031] In one possible implementation, the interference type includes at least one of single-tone interference, multi-tone interference, linear sweep frequency interference, pulse interference, comb spectrum interference, narrowband noise interference, and broadband noise interference.
[0032] In one possible implementation, the method further includes:
[0033] Multiple sets of training samples are obtained. Each set of training samples includes interference signal samples and pre-labeled labels. The interference signal samples are obtained by superimposing at least noise signals, useful signals and interference signals of at least one type of interference. The labels are used to label all the actual interference types corresponding to the interference signal samples.
[0034] Create the neural network;
[0035] For each training sample, the interference signal samples in the training samples are processed using the modules in the neural network to obtain the corresponding predicted interference type; the loss value is calculated based on the predicted interference type, the actual interference type, and the preset loss function.
[0036] The parameters of the neural network are adjusted based on the loss value until the parameters meet a preset condition, at which point training stops.
[0037] According to a second aspect of this application, a neural network-based interference signal identification device is provided, the device comprising:
[0038] The acquisition module is used to acquire the signal to be identified, which is obtained by superimposing at least a noise signal, a useful signal and interference signals of n types of interference, where n is an unknown number greater than or equal to 1;
[0039] The acquisition module is further configured to acquire a pre-trained neural network, which includes at least three parallel convolutional attention modules, a multi-branch fusion module, a dual-branch hybrid parallel attention module, and a fully connected layer. The convolutional attention module includes at least a two-dimensional deep convolution module, a channel attention module, a two-dimensional pointwise convolution module, and a spatial attention module. The multi-branch fusion module includes at least a two-dimensional convolution module, a compression and activation attention module, and an h-swish activation function. The dual-branch hybrid parallel attention module includes at least a parallel multi-head self-attention module, a convolutional block attention module, a one-dimensional layer normalization layer, and a feedforward neural network.
[0040] The conversion module is used to convert the signal to be identified into a corresponding time-frequency image;
[0041] The identification module is used to process the time-frequency image using modules in the neural network to obtain n types of interference signals in the signal to be identified.
[0042] According to a third aspect of this application, a computer-readable storage medium is provided, wherein at least one instruction is stored therein, the at least one instruction being loaded and executed by a processor to implement the neural network-based interference signal identification method as described above.
[0043] According to a fourth aspect of this application, a computer device is provided, the computer device including the above-described neural network-based interference signal identification device.
[0044] The beneficial effects of the technical solution provided in this application include at least the following:
[0045] Since the DSC (Depthwise Separable Convolution) module includes a two-dimensional depthwise convolution module and a two-dimensional pointwise convolution module, by improving the DSC module by adding a channel attention module after the two-dimensional depthwise convolution module and a spatial attention module after the two-dimensional pointwise convolution module, the DSC module can be improved into a convolutional attention module. This simplifies the model structure while maintaining performance, thereby improving training speed, and can capture the correlation of data in a larger spatial range, thereby improving recognition accuracy.
[0046] The dual-branch hybrid parallel attention module includes a parallel multi-head self-attention module and a convolutional block attention module. The multi-head self-attention module can extract long-distance dependencies and enhance the information connection between features. The channel attention module and spatial attention module in the convolutional block attention module can strengthen effective feature information and weaken ineffective feature information, thereby improving the model's classification performance and generalization ability, and achieving feature optimization selection.
[0047] The signal to be identified includes n types of interference signals. When the value of n is different, the neural network can identify the multimodal signal to be identified and consider the characteristics of different modes of the signal to be identified during the identification process, thereby improving the identification accuracy. Attached Figure Description
[0048] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0049] Figure 1 This is a schematic diagram of the structure of a neural network according to some exemplary embodiments;
[0050] Figure 2 This is a schematic diagram illustrating the structure of a convolutional attention module according to some exemplary embodiments;
[0051] Figure 3 This is a schematic diagram of the structure of a dual-branch hybrid parallel attention module according to some exemplary embodiments;
[0052] Figure 4This is a flowchart of a neural network-based training method provided in one embodiment of this application;
[0053] Figure 5 This is a flowchart of a neural network-based interference signal identification method provided in one embodiment of this application;
[0054] Figure 6 This is a flowchart of a neural network-based interference signal identification method provided in one embodiment of this application;
[0055] Figure 7 This is a structural block diagram of a neural network-based interference signal identification device provided in one embodiment of this application;
[0056] Figure 8 This is a structural block diagram of a computer device provided in one embodiment of this application. Detailed Implementation
[0057] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the implementation methods of this application will be further described in detail below with reference to the accompanying drawings.
[0058] In this application, a neural network needs to be created in advance and trained so that the trained neural network can identify multimodal interference signals. The structure and training process of the neural network are described below.
[0059] (1) Structure of neural networks
[0060] The neural network in this application includes at least a parallel three-layer convolutional attention module (DCPS), a multi-branch fusion module (MFFU), a dual-branch hybrid parallel attention module (HPA), and a fully connected layer (FC). It may also include a max pooling layer, a fusion module (FB), an average pooling layer (Avg Pooling), a patch embedding, and a vertical strip embedding.
[0061] In one example, the connections between the modules in the neural network are as follows: Figure 1 As shown, the neural network sequentially includes three parallel convolutional attention modules, a multi-branch fusion module, a max pooling layer, a convolutional attention module, a fusion module, an average pooling layer, a convolutional attention module, a fusion module, an average pooling layer, a convolutional attention module, a fusion module, parallel block embeddings and vertical strip embeddings, a dual-branch hybrid parallel attention module, and two fully connected layers.
[0062] In one example, the structure of a convolutional attention module is as follows: Figure 2As shown, a convolutional attention module includes at least a two-dimensional deep convolutional module (DConv), a channel attention module (CA), a two-dimensional pointwise convolutional module (PConv), and a spatial attention module (SA). The kernel of the two-dimensional deep convolutional module is k×k, and the value of k varies in different convolutional attention modules. For example, in a neural network with three parallel convolutional attention layers, the kernels of the three two-dimensional deep convolutional modules are 3×3, 5×5, and 7×7, respectively. The kernel of the two-dimensional pointwise convolutional module is 1×1.
[0063] In one example, the multi-branch fusion module includes at least a two-dimensional convolution module, a compression and activation attention module, and an h-swish activation function.
[0064] In one example, the structure of the dual-branch hybrid parallel attention module is as follows: Figure 3 As shown, the dual-branch hybrid parallel attention module includes at least a parallel multi-head self-attention module, a convolutional block attention module, a one-dimensional layer normalization layer, and a feedforward neural network. Specifically, the parallel multi-head self-attention module includes at least a one-dimensional layer normalization layer (Norm) and a multi-head self-attention module (MSA), the convolutional block attention module includes at least a cascaded channel attention module (CA) and a spatial attention module (SA), and the feedforward neural network includes at least two fully connected layers.
[0065] (2) Training process of neural networks
[0066] like Figure 4 As shown, the training process for a neural network includes the following steps:
[0067] Step 401: Obtain multiple sets of training samples. Each set of training samples includes interference signal samples and pre-labeled labels. The interference signal samples are obtained by superimposing at least noise signals, useful signals, and interference signals of at least one type of interference. The labels are used to label all the actual interference types corresponding to the interference signal samples.
[0068] Each training sample set includes one interference signal sample and one label. The interference signal sample can be a single interference signal sample or a composite interference signal sample. A single interference signal sample is obtained by superimposing a noise signal, a useful signal, and interference signals of any interference category. A composite interference signal sample is obtained by superimposing a noise signal, a useful signal, and interference signals of at least two interference categories.
[0069] Interference signal samples can be generated using simulation software, such as MATLAB. One interference category corresponds to multiple interference signal samples with different interference-to-signal ratios. The label of each interference signal sample is its interference category; for example, a single interference signal sample is labeled with its corresponding single interference category, while a composite interference signal sample is labeled with its corresponding at least two interference categories.
[0070] In one embodiment, taking wireless communication as an example, the useful signal is a BPSK (Binary Phase Shift Keying) modulated signal, the noise signal is additive white Gaussian noise, and the signal parameters are set as follows: signal-to-noise ratio of 5dB, interference-to-signal ratio of [-30dB, 10dB], interval of 4dB, carrier frequency of 140KHz, and sampling frequency of 800KHz.
[0071] The types of interference include at least one of the following: single-tone interference, multi-tone interference, linear sweep frequency interference, pulse interference, comb spectrum interference, narrowband noise interference, and broadband noise interference.
[0072] Among them, the frequency points of single-tone interference are randomly generated within the bandwidth, the frequency set of multi-tone interference is randomly generated within the bandwidth, the initial frequency of linear sweep interference is 120KHz, the sweep rate is 8MHz, the pulse interference is a rectangular pulse with a duty cycle of 2%, the total bandwidth of comb spectrum interference accounts for more than 1 / 2, the bandwidth of narrowband noise interference accounts for 1 / 10, and the bandwidth of broadband noise interference accounts for more than 2 / 3.
[0073] For each type of interference, 250 signal samples are used at each interference-to-signal ratio, and these samples are divided into a training set, a validation set, and a test set in a 6:2:2 ratio. The training set is the set of training samples, while the validation and test sets are used to verify whether the parameters of the neural network meet the preset conditions.
[0074] Step 402, Create the neural network.
[0075] Computer devices can create, for example Figure 2 The neural network shown.
[0076] Step 403: For each group of training samples, the interference signal samples in the training samples are processed using the modules in the neural network to obtain the corresponding predicted interference type; the loss value is calculated based on the predicted interference type, the actual interference type and the preset loss function.
[0077] The computer equipment can convert the interference signal samples in each training sample into corresponding time-frequency images, and then input the time-frequency images into the neural network. The neural network processes the time-frequency images and outputs the predicted interference type. The loss value is calculated based on the predicted interference type, the actual interference type, and the preset loss function.
[0078] When converting interference signal samples into time-frequency images, computer equipment can use short-time Fourier transform to convert the signal to be identified into a corresponding time-frequency image; or, it can use wavelet transform to convert the signal to be identified into a corresponding time-frequency image. After converting the interference signal samples into time-frequency images, the advantages of time-domain features and frequency-domain features can be combined, thereby improving the recognition accuracy.
[0079] Step 404: Adjust the parameters of the neural network according to the loss value until the parameters meet the preset conditions and then stop training.
[0080] During training, the computer device calculates the loss value based on the loss function and uses the Adam optimizer, an adaptive optimization algorithm, to adaptively adjust the learning rate. The initial learning rate is 0.001, and a total of 50 iterations are performed. After training stabilizes, the validation set and test set are input into the trained neural network to test the accuracy of interference recognition.
[0081] like Figure 5 The diagram illustrates a flowchart of a neural network-based interference signal identification method according to an embodiment of this application. This neural network-based interference signal identification method can be applied to computer devices. The neural network-based interference signal identification method may include:
[0082] Step 501: Obtain the signal to be identified. The signal to be identified is obtained by superimposing at least a noise signal, a useful signal, and interference signals of n types, where n is an unknown number greater than or equal to 1.
[0083] Where n is an unknown number greater than or equal to 1, meaning that the signal to be identified may contain one type of interference signal or at least two types of interference signals.
[0084] Step 502: Obtain the pre-trained neural network.
[0085] The computer device acquires the neural network trained through the above training process.
[0086] Step 503: Convert the signal to be identified into a corresponding time-frequency image.
[0087] Step 504: Use the modules in the neural network to process the time-frequency image to obtain n types of interference signals in the signal to be identified.
[0088] The computer device inputs the time-frequency image into the neural network. Each module in the neural network processes the time-frequency image and outputs a predicted interference type. This predicted interference type is used to identify n types of interference signals in the signal to be identified.
[0089] In summary, the neural network-based interference signal recognition method provided in this application improves the DSC module by adding a channel attention module after the two-dimensional deep convolution module and a spatial attention module after the two-dimensional pointwise convolution module. This transforms the DSC module into a convolutional attention module, which simplifies the model structure while maintaining performance, thereby improving training speed. Furthermore, it captures the correlation of data over a larger spatial range, thus improving recognition accuracy.
[0090] The dual-branch hybrid parallel attention module includes a parallel multi-head self-attention module and a convolutional block attention module. The multi-head self-attention module can extract long-distance dependencies and enhance the information connection between features. The channel attention module and spatial attention module in the convolutional block attention module can strengthen effective feature information and weaken ineffective feature information, thereby improving the model's classification performance and generalization ability, and achieving feature optimization selection.
[0091] The signal to be identified includes n types of interference signals. When the value of n is different, the neural network can identify the multimodal signal to be identified and consider the characteristics of different modes of the signal to be identified during the identification process, thereby improving the identification accuracy.
[0092] like Figure 6 The diagram illustrates a flowchart of a neural network-based interference signal identification method according to an embodiment of this application. This neural network-based interference signal identification method can be applied to computer devices. The neural network-based interference signal identification method may include:
[0093] Step 601: Obtain the signal to be identified. The signal is obtained by superimposing at least a noise signal, a useful signal, and interference signals of n types, where n is an unknown number greater than or equal to 1.
[0094] Where n is an unknown number greater than or equal to 1, meaning that the signal to be identified may contain one type of interference signal, or it may contain at least two types of interference signals. The interference types include at least one of the following: single-tone interference, multi-tone interference, linear sweep frequency interference, pulse interference, comb spectrum interference, narrowband noise interference, and broadband noise interference.
[0095] Step 602: Obtain the pre-trained neural network.
[0096] The computer device acquires the neural network trained through the above training process.
[0097] Step 603: Convert the signal to be identified into a corresponding time-frequency image based on short-time Fourier transform; or, convert the signal to be identified into a corresponding time-frequency image based on wavelet transform.
[0098] In one example, the resolution of the time-frequency image is 3×224×224.
[0099] Converting interference signal samples into time-frequency images can combine the advantages of time-domain and frequency-domain features, thereby improving recognition accuracy.
[0100] Step 604: Use the modules in the neural network to process the time-frequency image to obtain n types of interference signals in the signal to be identified.
[0101] When a neural network includes three parallel convolutional attention modules, processing the time-frequency image using each module in the neural network can include: performing convolution operations on the time-frequency image using a two-dimensional depth convolution module to obtain a first intermediate result; performing channel attention calculations on the first intermediate result using a channel attention module to obtain a second intermediate result; performing convolution operations on the second intermediate result using a two-dimensional pointwise convolution module to obtain a third intermediate result; and performing spatial attention calculations on the third intermediate result using a spatial attention module to obtain the output result of the convolutional attention module.
[0102] The input to the convolutional attention module is T∈R C×H×W First, a three-layer two-dimensional depthwise convolution module with a kernel size of k×k is used to obtain the first intermediate result T1∈R. C×H×W The first intermediate result is input into the channel attention module to obtain the second intermediate result T2∈R. C×H×W Then, the second intermediate result is input into the two-dimensional pointwise convolution module and the spatial attention module to obtain the output result T3∈R with channel and spatial attention. C0×H×W The entire process can be summarized as follows:
[0103] DCPS(T,k)=SA(PConv(CA(DConv(T,k)))) (1)
[0104] Where T represents the feature map (time-frequency image) of the input convolutional attention module, k represents the kernel size in the two-dimensional deep convolution module, C represents the number of channels, H represents the height of the feature map, W represents the width of the feature map, T1 represents the feature map after processing by two-dimensional deep convolution, T2 represents the feature map after processing by the channel attention mechanism, and T3 represents the feature map after processing by pointwise convolution and spatial attention mechanism.
[0105] Equation (1) means that the feature map T is sequentially passed through a two-dimensional depthwise convolution with a kernel of k×k, channel attention, two-dimensional pointwise convolution, and spatial attention.
[0106] Two-dimensional depthwise convolution modules and two-dimensional pointwise convolution modules are used to reduce the number of model parameters, thereby reducing the computational complexity of the model.
[0107] The channel attention module focuses on the relationships between different channels of the time-frequency image, assigning different weights to each channel and strengthening the attention to important channels, thereby helping to improve the model's representational ability.
[0108] The spatial attention module focuses on the spatial structure within the time-frequency image, assigning different weights to each pixel location, enabling the model to focus on important regions in the time-frequency image.
[0109] When a neural network includes a multi-branch fusion module and other modules, processing time-frequency images using these modules can include: concatenating the outputs of three parallel convolutional attention modules into a feature map; performing convolution operations on the feature map using a two-dimensional convolution module to obtain a fourth intermediate result; performing attention operations on the fourth intermediate result using a compression and activation attention module to obtain a fifth intermediate result; performing operations on the fifth intermediate result using the h-swish activation function to obtain a sixth intermediate result; and processing the sixth intermediate result using other modules to obtain the outputs of the other modules.
[0110] The outputs of the three convolutional attention modules are F1, F2, and F3, each with a size of 3×H×W. These three outputs are concatenated along the channel dimension to obtain a 9×H×W feature map, which incorporates feature information from each convolutional attention module. Then, a 1×1 two-dimensional convolutional layer is used to obtain a preliminary feature fusion result, and the number of channels is varied to obtain a fourth intermediate result. Finally, the sixth intermediate result is obtained through compression, activation of attention, and the h-swish activation function. The entire process can be summarized as follows:
[0111]
[0112] Here, F1, F2, and F3 represent the output results of the three-layer convolutional attention module, SE represents the compression and activation attention module, HS represents the h-swish activation function, and concat represents the concatenation operation of the three output results.
[0113] After obtaining the sixth intermediate result through the multi-branch fusion module, this sixth intermediate result needs to be input into other modules for processing, ultimately yielding the input feature map of the dual-branch hybrid parallel attention module. For example... Figure 2 As shown, other modules include max pooling layers, convolutional attention modules, fusion modules, average pooling layers, convolutional attention modules, fusion modules, average pooling layers, convolutional attention modules, fusion modules, parallel block embeddings, and vertical strip embeddings.
[0114] When a neural network includes a dual-branch hybrid parallel attention module and a fully connected layer, processing time-frequency images using the modules in the neural network can include: normalizing the outputs of other modules using a one-dimensional normalization layer in the dual-branch hybrid parallel attention module, multiplying the normalized results by three different weight matrices to obtain the query, key, and value; performing attention operations and concatenation on the query, key, and value using a multi-head self-attention module in the dual-branch hybrid parallel attention module to obtain a seventh intermediate result; performing channel attention operations on the outputs of other modules using a channel attention module in a convolutional block attention module, and performing spatial attention operations on the results using a spatial attention module in a convolutional block attention module to obtain an eighth intermediate result; normalizing the sum of the seventh and eighth intermediate results using a one-dimensional normalization layer to obtain a ninth intermediate result; processing the ninth intermediate result using a feedforward neural network to obtain the output of the dual-branch hybrid parallel attention module; and calculating the n types of interference signals in the signal to be identified using a fully connected layer based on the output of the dual-branch hybrid parallel attention module.
[0115] The input feature map is X∈R C×M First, the input feature map is processed by a multi-head self-attention module and a convolutional block attention module, respectively, to obtain the seventh and eighth intermediate results. Then, the seventh and eighth intermediate results are summed and processed by a one-dimensional normalization layer and a feedforward neural network to obtain the output result X1∈R of the dual-branch hybrid parallel attention module. C×M M represents the length of the feature sequence. The entire process can be summarized as follows:
[0116] X' = X MSA +X CBAM (3)
[0117] X”=FFN(LN(X'))+X' (4)
[0118] Among them, X MSA Characterizing the multi-head self-attention module, X CBAM The convolutional block attention module is represented by LN, the one-dimensional normalization layer is represented by LN, and the feedforward neural network is represented by FFN.
[0119] For the multi-head self-attention module, the input feature map is first processed through a one-dimensional normalization layer. Then, the normalization result is multiplied by different weight matrices to obtain the query, key, and value, which are then input into the multi-head self-attention module to obtain the seventh intermediate result. The entire process can be summarized as follows:
[0120]
[0121] head i=Attention(QW i Q ,KW i K VW i V (6)
[0122] MSA=concat(head1,head2,…,head n )W (7)
[0123] Where Q represents the query, K represents the key, V represents the value, and d k The dimension of the key is represented by softmax, the normalization exponential function is represented by W, the weight matrix is represented by W, and concat is represented by concatenation operation.
[0124] The convolutional block attention module consists of a channel attention module and a spatial attention module connected in sequence. The input to the channel attention module is the input feature map F∈R. C×M The output is a one-dimensional channel attention feature map M. C ∈R C×1 The one-dimensional channel attention feature map is input into the spatial attention module to obtain a one-dimensional spatial attention feature map M. S ∈R 1×M This is the eighth intermediate result. The entire process can be summarized as follows:
[0125]
[0126]
[0127] Where F represents the feature map of the input channel attention module, R represents the real number field, C represents the number of channels, H represents the height of the feature map, W represents the width of the feature map, F' represents the feature sequence after processing by the channel attention mechanism, and F” represents the feature sequence after processing by the spatial attention mechanism.
[0128] Equation (8) means: the feature sequence after processing by the channel attention mechanism = channel attention feature sequence × original feature sequence.
[0129] Equation (9) means: the feature sequence after spatial attention mechanism processing = spatial attention feature sequence × feature sequence after channel attention mechanism processing.
[0130] The fully connected layer uses the softmax activation function to map the outputs of multiple neurons to probabilities in [0,1] to achieve multi-label classification, and the loss function is the cross-entropy loss function.
[0131] In summary, the neural network-based interference signal recognition method provided in this application improves the DSC module by adding a channel attention module after the two-dimensional deep convolution module and a spatial attention module after the two-dimensional pointwise convolution module. This transforms the DSC module into a convolutional attention module, which simplifies the model structure while maintaining performance, thereby improving training speed. Furthermore, it captures the correlation of data over a larger spatial range, thus improving recognition accuracy.
[0132] The dual-branch hybrid parallel attention module includes a parallel multi-head self-attention module and a convolutional block attention module. The multi-head self-attention module can extract long-distance dependencies and enhance the information connection between features. The channel attention module and spatial attention module in the convolutional block attention module can strengthen effective feature information and weaken ineffective feature information, thereby improving the model's classification performance and generalization ability, and achieving feature optimization selection.
[0133] The signal to be identified includes n types of interference signals. When the value of n is different, the neural network can identify the multimodal signal to be identified and consider the characteristics of different modes of the signal to be identified during the identification process, thereby improving the identification accuracy.
[0134] like Figure 7 The diagram illustrates a structural block diagram of a neural network-based interference signal identification device according to an embodiment of this application. This neural network-based interference signal identification device can be applied to computer equipment. The neural network-based interference signal identification device may include:
[0135] The acquisition module 710 is used to acquire the signal to be identified. The signal to be identified is obtained by superimposing at least a noise signal, a useful signal and interference signals of n types of interference, where n is an unknown number greater than or equal to 1.
[0136] The acquisition module 710 is also used to acquire a pre-trained neural network, which includes at least three parallel convolutional attention modules, a multi-branch fusion module, a dual-branch hybrid parallel attention module, and a fully connected layer. The convolutional attention module includes at least a two-dimensional deep convolution module, a channel attention module, a two-dimensional pointwise convolution module, and a spatial attention module. The multi-branch fusion module includes at least a two-dimensional convolution module, a compression and activation attention module, and an h-swish activation function. The dual-branch hybrid parallel attention module includes at least a parallel multi-head self-attention module, a convolutional block attention module, a one-dimensional layer normalization layer, and a feedforward neural network.
[0137] The conversion module 720 is used to convert the signal to be identified into a corresponding time-frequency image;
[0138] The identification module 730 is used to process the time-frequency image using modules in the neural network to obtain n types of interference signals in the signal to be identified.
[0139] In an optional embodiment, when the neural network includes three parallel convolutional attention modules, the recognition module 730 is further configured to:
[0140] The time-frequency image is convolved using a two-dimensional depth convolution module to obtain the first intermediate result;
[0141] The channel attention module is used to calculate the channel attention of the first intermediate result to obtain the second intermediate result;
[0142] The second intermediate result is convolved using a two-dimensional pointwise convolution module to obtain the third intermediate result;
[0143] Spatial attention is applied to the third intermediate result using the spatial attention module to obtain the output of the convolutional attention module.
[0144] In an optional embodiment, when the neural network includes a multi-branch fusion module and other modules, the recognition module 730 is further configured to:
[0145] The outputs of the three parallel convolutional attention modules are concatenated into a feature map;
[0146] The feature map is convolved using a two-dimensional convolution module to obtain the fourth intermediate result.
[0147] The fifth intermediate result is obtained by performing attention operations on the fourth intermediate result using the compression and activation attention modules.
[0148] The h-swish activation function is used to process the fifth intermediate result to obtain the sixth intermediate result;
[0149] The sixth intermediate result is processed using other modules to obtain the output results of those other modules.
[0150] In an optional embodiment, when the neural network includes a dual-branch hybrid parallel attention module and a fully connected layer, the recognition module 730 is further configured to:
[0151] The one-dimensional normalization layer in the dual-branch hybrid parallel attention module is used to normalize the output of other modules. The normalization result is multiplied by three different weight matrices to obtain the query, key and value. The multi-head self-attention module in the dual-branch hybrid parallel attention module is used to perform attention operation and concatenation on the query, key and value to obtain the seventh intermediate result.
[0152] The channel attention module in the convolutional block attention module is used to perform channel attention operations on the output results of other modules, and the spatial attention module in the convolutional block attention module is used to perform spatial attention operations on the operation results to obtain the eighth intermediate result;
[0153] The sum of the seventh and eighth intermediate results is normalized using a one-dimensional normalization layer to obtain the ninth intermediate result.
[0154] The ninth intermediate result is processed using a feedforward neural network to obtain the output of the dual-branch hybrid parallel attention module;
[0155] The output of the dual-branch hybrid parallel attention module is calculated using a fully connected layer to obtain n types of interference signals in the signal to be identified.
[0156] In an optional embodiment, the conversion module 720 is further configured to:
[0157] The signal to be identified is converted into a corresponding time-frequency image based on the short-time Fourier transform; or...
[0158] The signal to be identified is converted into a corresponding time-frequency image based on wavelet transform.
[0159] In one optional embodiment, the interference type includes at least one of single-tone interference, multi-tone interference, linear sweep interference, pulse interference, comb spectrum interference, narrowband noise interference, and broadband noise interference.
[0160] In an optional embodiment, the device further includes a training module for:
[0161] Multiple sets of training samples are obtained. Each set of training samples includes interference signal samples and pre-labeled labels. The interference signal samples are obtained by superimposing at least noise signals, useful signals and interference signals of at least one type of interference. The labels are used to label all the actual interference types corresponding to the interference signal samples.
[0162] Create a neural network;
[0163] For each training sample, the interference signal samples in the training samples are processed using modules in the neural network to obtain the corresponding predicted interference type; the loss value is calculated based on the predicted interference type, the actual interference type, and the preset loss function.
[0164] The parameters of the neural network are adjusted based on the loss value until the parameters meet the preset conditions, at which point training stops.
[0165] In summary, the interference signal recognition device based on neural networks provided in this application improves the DSC module by adding a channel attention module after the two-dimensional deep convolution module and a spatial attention module after the two-dimensional pointwise convolution module. This improves the DSC module into a convolutional attention module, which simplifies the model structure while maintaining performance, thereby improving training speed. Furthermore, it can capture the correlation of data in a larger spatial range, thereby improving recognition accuracy.
[0166] The dual-branch hybrid parallel attention module includes a parallel multi-head self-attention module and a convolutional block attention module. The multi-head self-attention module can extract long-distance dependencies and enhance the information connection between features. The channel attention module and spatial attention module in the convolutional block attention module can strengthen effective feature information and weaken ineffective feature information, thereby improving the model's classification performance and generalization ability, and achieving feature optimization selection.
[0167] The signal to be identified includes n types of interference signals. When the value of n is different, the neural network can identify the multimodal signal to be identified and consider the characteristics of different modes of the signal to be identified during the identification process, thereby improving the identification accuracy.
[0168] like Figure 8 As shown, it illustrates a structural schematic diagram of a computer device 800 suitable for implementing embodiments of the present invention. Figure 8 The computer device 800 shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.
[0169] like Figure 8 As shown, the computer device 800 includes a central processing unit (CPU) 801, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 802 or a program loaded from a storage section 808 into a random access memory (RAM) 803. The RAM 803 also stores various programs and data required for the operation of the system 800. The CPU 801, ROM 802, and RAM 803 are interconnected via a bus 804. Input / output (I / O) interfaces are also connected to the bus 804.
[0170] The following components are connected to I / O interface 805: an input section 806 including a keyboard, mouse, etc.; an output section 807 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 808 including a hard disk, etc.; and a communication section 809 including a network interface card such as a LAN card, modem, etc. The communication section 809 performs communication processing via a network such as the Internet. A drive 810 is also connected to I / O interface 805 as needed. A removable medium 811, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 810 as needed so that computer programs read from it can be installed into storage section 808 as needed.
[0171] In particular, according to the embodiments disclosed in this invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 809, and / or installed from removable medium 811. When the computer program is executed by central processing unit (CPU) 801, it performs the functions defined above in the system of this invention.
[0172] It should be noted that the computer-readable medium shown in this invention can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this invention, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.
[0173] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.
[0174] The modules described in the embodiments of the present invention can be implemented in software or hardware. The described modules can also be housed in a processor; for example, a processor can be described as including a sending module, an acquisition module, a determining module, and a first processing module. The names of these modules do not necessarily limit the module itself; for example, the sending module can also be described as "a module that sends an image acquisition request to a connected server."
[0175] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can occur depending on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.
Claims
1. A method for identifying interference signals based on neural networks, characterized in that, The method includes: The signal to be identified is obtained by superimposing at least a noise signal, a useful signal, and interference signals of n types, where n is an unknown number greater than or equal to 1. Obtain a pre-trained neural network, wherein the neural network includes at least three parallel convolutional attention modules, a multi-branch fusion module, a dual-branch hybrid parallel attention module, and a fully connected layer; the convolutional attention module includes at least a two-dimensional deep convolution module, a channel attention module, a two-dimensional pointwise convolution module, and a spatial attention module; the multi-branch fusion module includes at least a two-dimensional convolution module, a compression and activation attention module, and an h-swish activation function; and the dual-branch hybrid parallel attention module includes at least a parallel multi-head self-attention module, a convolutional block attention module, a one-dimensional layer normalization layer, and a feedforward neural network. The signal to be identified is converted into a corresponding time-frequency image; The time-frequency image is processed using modules in the neural network to obtain n types of interference signals in the signal to be identified; When the neural network includes a multi-branch fusion module and other modules, the process of using the modules in the neural network to process the time-frequency image includes: concatenating the outputs of three parallel convolutional attention modules into a feature map; performing convolution operations on the feature map using the two-dimensional convolution module to obtain a fourth intermediate result; performing attention operations on the fourth intermediate result using the compression and activation attention module to obtain a fifth intermediate result; performing operations on the fifth intermediate result using the h-swish activation function to obtain a sixth intermediate result; and processing the sixth intermediate result using other modules to obtain the outputs of the other modules. When the neural network includes a dual-branch hybrid parallel attention module and a fully connected layer, the processing of the time-frequency image using the modules in the neural network includes: using a one-dimensional normalization layer in the dual-branch hybrid parallel attention module to normalize the output results of the other modules; multiplying the normalized results by three different weight matrices to obtain the query, key, and value; using a multi-head self-attention module in the dual-branch hybrid parallel attention module to perform attention operations and concatenation on the query, the key, and the value to obtain a seventh intermediate result; and using a channel attention module in the convolutional block attention module to process the... The outputs of other modules are subjected to channel attention operations. The spatial attention module in the convolutional block attention module is used to perform spatial attention operations on the results to obtain the eighth intermediate result. The sum of the seventh and eighth intermediate results is normalized using the one-dimensional normalization layer to obtain the ninth intermediate result. The ninth intermediate result is processed using the feedforward neural network to obtain the output of the dual-branch hybrid parallel attention module. The fully connected layer is used to calculate the output of the dual-branch hybrid parallel attention module to obtain n types of interference signals in the signal to be identified.
2. The interference signal identification method based on neural networks according to claim 1, characterized in that, When the neural network includes three parallel convolutional attention modules, the process of using each module in the neural network to process the time-frequency image includes: The time-frequency image is convolved using the two-dimensional depth convolution module to obtain a first intermediate result; The channel attention module is used to perform channel attention calculation on the first intermediate result to obtain the second intermediate result; The second intermediate result is convolved using the two-dimensional pointwise convolution module to obtain the third intermediate result. The spatial attention module is used to perform spatial attention operations on the third intermediate result to obtain the output result of the convolutional attention module.
3. The interference signal identification method based on neural networks according to claim 1, characterized in that, The step of converting the signal to be identified into a corresponding time-frequency image includes: The signal to be identified is converted into a corresponding time-frequency image based on the short-time Fourier transform; or... The signal to be identified is converted into a corresponding time-frequency image based on wavelet transform.
4. The interference signal identification method based on neural networks according to claim 1, characterized in that, The interference types include at least one of single-tone interference, multi-tone interference, linear sweep frequency interference, pulse interference, comb spectrum interference, narrowband noise interference, and broadband noise interference.
5. The interference signal identification method based on neural networks according to any one of claims 1 to 4, characterized in that, The method further includes: Multiple sets of training samples are obtained. Each set of training samples includes interference signal samples and pre-labeled tags. The interference signal samples are obtained by superimposing at least noise signals, useful signals and interference signals of at least one type of interference. The tags are used to label all the actual interference types corresponding to the interference signal samples. Create the neural network; For each training sample, the interference signal samples in the training samples are processed using the modules in the neural network to obtain the corresponding predicted interference type; the loss value is calculated based on the predicted interference type, the actual interference type, and the preset loss function. The parameters of the neural network are adjusted based on the loss value until the parameters meet a preset condition, at which point training stops.
6. A neural network-based interference signal identification device, characterized in that, The device includes: The acquisition module is used to acquire the signal to be identified, which is obtained by superimposing at least a noise signal, a useful signal and interference signals of n types of interference, where n is an unknown number greater than or equal to 1; The acquisition module is further configured to acquire a pre-trained neural network, which includes at least three parallel convolutional attention modules, a multi-branch fusion module, a dual-branch hybrid parallel attention module, and a fully connected layer. The convolutional attention module includes at least a two-dimensional deep convolution module, a channel attention module, a two-dimensional pointwise convolution module, and a spatial attention module. The multi-branch fusion module includes at least a two-dimensional convolution module, a compression and activation attention module, and an h-swish activation function. The dual-branch hybrid parallel attention module includes at least a parallel multi-head self-attention module, a convolutional block attention module, a one-dimensional layer normalization layer, and a feedforward neural network. The conversion module is used to convert the signal to be identified into a corresponding time-frequency image; The identification module is used to process the time-frequency image using modules in the neural network to obtain n types of interference signals in the signal to be identified; When the neural network includes a multi-branch fusion module and other modules, the recognition module is further configured to: concatenate the outputs of the parallel three-layer convolutional attention modules into a feature map; perform convolution operations on the feature map using the two-dimensional convolution module to obtain a fourth intermediate result; perform attention operations on the fourth intermediate result using the compression and activation attention module to obtain a fifth intermediate result; perform operations on the fifth intermediate result using the h-swish activation function to obtain a sixth intermediate result; and process the sixth intermediate result using other modules to obtain the outputs of the other modules. When the neural network includes a dual-branch hybrid parallel attention module and a fully connected layer, the recognition module is further configured to: normalize the output results of the other modules using a one-dimensional normalization layer in the dual-branch hybrid parallel attention module; multiply the normalization results by three different weight matrices to obtain the query, key, and value; perform attention operations and concatenation on the query, key, and value using a multi-head self-attention module in the dual-branch hybrid parallel attention module to obtain a seventh intermediate result; perform channel attention operations on the output results of the other modules using a channel attention module in the convolutional block attention module; perform spatial attention operations on the operation results using a spatial attention module in the convolutional block attention module to obtain an eighth intermediate result; normalize the sum of the seventh and eighth intermediate results using the one-dimensional normalization layer to obtain a ninth intermediate result; process the ninth intermediate result using the feedforward neural network to obtain the output result of the dual-branch hybrid parallel attention module; and calculate the output result of the dual-branch hybrid parallel attention module using the fully connected layer to obtain n types of interference signals in the signal to be recognized.
7. A computer-readable storage medium, characterized in that, The storage medium stores at least one instruction, which is loaded and executed by a processor to implement the neural network-based interference signal identification method as described in any one of claims 1 to 5.
8. A computer device, characterized in that, The computer device includes: the neural network-based interference signal identification device as described in claim 6.