A communication interference signal identification method based on a Mobile-Former network

Abstract

The application relates to a communication interference signal recognition method based on a Mobile-Former network, and relates to the field of communication.The application is used to solve the problem that, when composite interference signals are recognized, the number of interference signal categories exponentially increases, the training complexity of the network increases, the difference between the extracted features decreases, and the interference signal recognition performance decreases.The application uses a Mobile-Former network as a feature extraction network, extracts the features of a signal time-frequency graph and a frequency spectrum graph, and builds a feature fusion network to fuse the extracted time-frequency graph features and frequency spectrum graph features, so that more abundant features are obtained, the interference signal recognition performance under a low signal-to-noise ratio is effectively improved, a multi-label classification method is used to realize the recognition of single interference and composite interference signals, the output vector of the network is judged by selecting a proper threshold value, and the final recognition result is obtained.

Description

Technical Field

[0001] This invention belongs to the field of communications, and particularly relates to the identification of communication interference signals. Background Technology

[0002] Identification of communication interference signals is one of the key technologies for sensing complex electromagnetic environments. It is the prerequisite and foundation for anti-interference communication and has wide applications in both civilian and military fields. Only by formulating corresponding anti-interference strategies according to different types of interference signals can the security of wireless communication systems be effectively guaranteed.

[0003] With the continuous development of modern electronic and intelligent technologies, the electromagnetic environment is becoming increasingly complex, filled with various dynamic electromagnetic interference signals. These signals are characterized by their ubiquitous spatial presence and dense, overlapping temporal frequencies. To better utilize interference resources and achieve more effective interference, interfering parties often employ multiple jammers to collaboratively generate composite interference signals with overlapping interference signals, thereby disrupting communication systems. Therefore, communication interference identification systems must not only effectively identify individual interference signals but also the specific components within composite interference signals.

[0004] Traditional methods for identifying interference signals rely on manually extracted signal features, which are then classified using algorithms such as support vector machines and decision trees. However, manually extracted signal features are often incomplete and insufficient, making it difficult to accurately classify interference signals. In recent years, deep learning has been increasingly used in interference signal identification. However, the input to neural networks mostly uses single data, such as time-domain signals, spectra, or time-frequency graphs. A few researchers have fused multi-domain features of the same signal, but this fusion method merely concatenates multiple feature vectors, failing to fully integrate features from multiple domains. Currently, research on composite interference signal identification is limited, and most studies treat composite interference signals as a new type of single interference signal. The characteristic that composite interference signals are composed of multiple single interference signals is not properly utilized. This leads to an exponential increase in the number of interference signal categories, increasing network training complexity, reducing the differences in extracted features, and degrading interference signal identification performance. Summary of the Invention

[0005] This invention aims to address the problem that treating composite interference signals as single-signal interference during composite interference signal identification leads to an exponential increase in the number of interference signal categories, increasing network training complexity, reducing the differences in extracted features, and degrading interference signal identification performance. The invention provides a communication interference signal identification method based on Mobile-Former networks.

[0006] A method for identifying communication interference signals based on Mobile-Former networks is as follows:

[0007] Short-time Fourier transform and Fourier transform are performed on the interference signals to be identified to obtain the time-frequency diagram and the spectrum diagram.

[0008] The time-frequency graph and the spectrum graph are respectively input into two trained Mobile-Former networks to extract the time-frequency graph feature vector and the spectrum graph feature vector;

[0009] The time-frequency graph feature vector and the spectrum graph feature vector are input into the trained feature fusion network and fused to obtain a fusion result. The fusion result is a 1×N one-dimensional vector, where N is the total number of interference types that can be included in the interference signal to be identified, and each element represents an interference type.

[0010] Each element in the fusion result is compared with a decision threshold. Elements greater than the decision threshold are marked as 1, and elements less than the decision threshold are marked as 0, thus obtaining the recognition result. If there is an element with a value of 1 in the recognition result, the interference signal has the interference type corresponding to that element.

[0011] Furthermore, the process of extracting feature vectors using the Mobile-Former network described above is as follows:

[0012] The original input image is preprocessed to obtain an initial two-dimensional local feature map;

[0013] The initial local feature map and the initial global feature token are updated sequentially through multiple Mobile-Former modules to obtain the updated local feature map and global feature token;

[0014] The updated local feature map is transformed into a one-dimensional vector by global average pooling (AGP) and then concatenated with the updated global feature token.

[0015] The concatenated result is output through a fully connected layer (FC) to obtain the extracted feature vector;

[0016] The input original image is a time-frequency graph or a spectrum graph;

[0017] The initial global feature token is a set of randomly initialized one-dimensional vectors used to represent the global information of the original image.

[0018] Furthermore, the preprocessing of the input raw image described above includes:

[0019] The input original image is sequentially subjected to 3×3 Conv convolution, batch normalization, and ReLU6 activation to obtain a two-dimensional initial local feature map.

[0020] Furthermore, the number of Mobile-Former modules mentioned above is I, and the local feature map X output by the i-th Mobile-Former module is... i With global feature token Z i As the input of the (i+1)th Mobile-Former module, i = 0, 1, ..., I-1. When i = 0, the input of the Mobile-Former module is the initial local feature map and the initial global feature token.

[0021] The Mobile-Former module includes: a Mobile submodule, a Former submodule, a Mobile-to-Former submodule, and a Former-to-Mobile submodule;

[0022] The input to the Mobile submodule is the local feature map X. i The local feature map X i The output is sequentially passed through a first 1×1 Conv convolutional layer, a first DY-ReLU activation function layer, a 3×3 DW-Conv convolutional layer, a second DY-ReLU activation function layer, and a second 1×1 Conv convolutional layer, and then input to the Former to Mobile submodule.

[0023] The Mobile-to-Former submodule is used to transfer the local feature map X i Integrating into global feature token Z i The results are then input into the Former submodule.

[0024] The Former submodule is a Transformer block, and its output is a global feature token Z. i+1 ;

[0025] The Former to Mobile submodule is used to transfer the global feature token Z output by the Former submodule. i+1 The result is incorporated into the output of the Mobile submodule to obtain the local feature map X. i+1 .

[0026] Furthermore, the above will use the local feature map X i Integrating into global feature token Z i Including:

[0027] The global feature token Z i Linear processing is performed through a linear layer, followed by comparison with the local feature map X. i Perform attention calculations.

[0028] Furthermore, the global feature token Z output by the Former submodule mentioned above... i+1The results incorporated into the output of the Mobile submodule include:

[0029] Global feature token Z i+1 Linear processing is performed through a linear layer, followed by attention calculation with the local features passed through the Mobile submodule.

[0030] Furthermore, the global feature token Z output by the aforementioned Former submodule... i+1 After passing through a multilayer perceptron (MLP), the adjustment parameter θ = {a} is obtained. 1 ,a 2}, using this adjustment parameter θ={a 1 ,a 2 To adjust the first and second DY-ReLU activation functions;

[0031] The expressions for both the first and second DY-ReLU activation functions are:

[0032] f(x) = max{a 1 x,a 2 x},

[0033] Among them, a 1 and a 2 To adjust the two components of parameter θ, x is a variable of the DY-ReLU activation function.

[0034] Furthermore, the fusion process of the above feature fusion network is as follows:

[0035] The time-frequency graph feature vector, the spectrum graph feature vector, and the average value of the time-frequency graph feature vector and the spectrum graph feature vector are linearly processed through a linear layer to obtain matrices V, K, and Q, respectively.

[0036] The matrices V, K, and Q are fused using an attention mechanism to obtain the fused feature vector y:

[0037]

[0038] The feature vector y is then added to the feature vectors of the time-frequency plot and the spectrum plot. The resulting sum is then passed through a fully connected layer (FC) and a sigmoid activation function to obtain the fusion result.

[0039] Furthermore, the loss functions for both the Mobile-Former network and the feature fusion network mentioned above are:

[0040]

[0041] in, p is a label for the nth type of single interference, with a value of 1 or 0.n Let be the output value of the nth type of single disturbance in the output vector of the neural network.

[0042] Furthermore, the expression for the above decision threshold ε is as follows:

[0043]

[0044] Among them, F1 n Let be the average F1 score of the nth type of interference, and we have:

[0045]

[0046] Precision n and Recall n Let be the precision and recall of the nth type of interference, respectively, and their expressions are as follows:

[0047]

[0048]

[0049] TP n For the number of samples correctly identified as the nth type of interference, FP n FN represents the number of other interference samples identified as the nth type of interference. n This represents the number of nth type of interference samples that were identified as other types of interference.

[0050] The beneficial effects of this invention are as follows:

[0051] This invention provides a communication interference signal identification method based on a Mobile-Former network. Using the Mobile-Former network as the feature extraction network, features from the signal's time-frequency plot and spectrum plot are extracted separately. A feature fusion network is then constructed to fuse the extracted time-frequency plot features with the spectrum plot features, obtaining richer features and effectively improving the interference signal identification performance under low interference-to-noise ratio (NNR). Furthermore, a multi-label classification method is used to identify single and composite interference signals. By selecting an appropriate threshold to determine the network's output vector, the final identification result is obtained.

[0052] In summary, this invention extracts the time-frequency graph features and spectrum graph features of the signal separately and then fuses them to obtain more complete and richer features, effectively improving the interference signal recognition performance under low interference-to-noise ratio. In addition, the recognition of single interference signals and composite interference is regarded as a multi-label classification problem. By making full use of the feature that composite interference signals are composed of single interference signals, the recognition of single interference signals and composite interference signals is achieved by using the multi-label classification idea. Attached Figure Description

[0053] Figure 1 This is a block diagram illustrating the principle of a communication interference signal identification method based on a Mobile-Former network as described in this invention.

[0054] Figure 2 This is a schematic diagram of the Mobile-Former network structure;

[0055] Figure 3 This is a structural diagram of the Mobile-Former module;

[0056] Figure 4 This is a schematic diagram of the feature fusion network structure;

[0057] Figure 5 This is a schematic diagram of the confusion matrix for the interference signal identification results when the interference-to-noise ratio is -16dB.

[0058] Figure 6 This is a schematic diagram of the confusion matrix for the interference signal identification results when the interference-to-noise ratio is -10dB.

[0059] Figure 7 The F1 score curves show the identification results of 12 types of interference signals. Detailed Implementation

[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0061] Specific implementation method one: Refer to Figures 1 to 4 This embodiment describes a method for identifying communication interference signals based on a Mobile-Former network, as detailed below:

[0062] S1: Perform short-time Fourier transform and Fourier transform on the interference signal to be identified to obtain the time-frequency diagram and the spectrum diagram.

[0063] S2: Input the time-frequency graph and the spectrum graph into two trained Mobile-Former networks respectively to extract the feature vectors of the time-frequency graph and the spectrum graph. The process of feature vector extraction by the Mobile-Former network is as follows:

[0064] The time-frequency image or spectrogram is used as the original image, and the original image is preprocessed to obtain a two-dimensional initial local feature map, including: performing 3×3 Conv convolution, batch normalization and ReLU6 activation on the input original image in sequence to obtain a two-dimensional initial local feature map.

[0065] A set of randomly initialized one-dimensional vectors is used to represent the global information of the original image, serving as the initial global feature token. Then, the initial local feature map and the initial global feature token are sequentially updated through 12 Mobile-Former modules to obtain the updated local feature map and global feature token. The local feature map X output by the i-th Mobile-Former module is... i With global feature token Z i As the input to the (i+1)th Mobile-Former module, i = 0, 1, ..., 11, when i = 0, the input to the Mobile-Former module is the initial local feature map and the initial global feature token.

[0066] The Mobile-Former module includes: the Mobile submodule, the Former submodule, the Mobile-to-Former submodule, and the Former-to-Mobile submodule.

[0067] The input to the Mobile submodule is the local feature map X. i The local feature map X i The output is sequentially passed through a first 1×1 Conv convolutional layer, a first DY-ReLU activation function layer, a 3×3 DW-Conv convolutional layer, a second DY-ReLU activation function layer, and a second 1×1 Conv convolutional layer, and then input to the Former to Mobile submodule.

[0068] The Mobile to Former submodule will transfer the global feature token Z i Linear processing is performed through a linear layer, followed by comparison with the local feature map X. i Attention calculation is performed so that the local feature map X i Integrating into global feature token Z i Then, the Mobile module inputs the result into the Former submodule.

[0069] The Former submodule is a Transformer block, and its output is the global feature token Z. i+1 The global signature token Z output by the Former submodule. i+1 After passing through a multilayer perceptron (MLP), the adjustment parameter θ = {a} is obtained. 1 ,a 2}, using this adjustment parameter θ={a 1 ,a 2 To adjust the first and second DY-ReLU activation functions in the Mobile submodule.

[0070] The expressions for both the first and second DY-ReLU activation functions are:

[0071] f(x) = max{a 1 x,a 2 x},

[0072] Among them, a 1 and a 2 To adjust the two components of parameter θ, x is a variable of the DY-ReLU activation function.

[0073] The Former to Mobile submodule will use the global feature token Z i+1 Linear processing is performed through a linear layer, followed by attention calculation with the local features from the Mobile submodule, resulting in a global feature token Z output by the Former submodule. i+1 The result is incorporated into the output of the Mobile submodule to obtain the local feature map X. i+1 .

[0074] The local feature maps updated by the 12 Mobile-Former modules are transformed into one-dimensional vectors using global average pooling (AGP) and concatenated with the updated global feature tokens. The concatenated result is then output through a fully connected layer (FC) to obtain the extracted feature vector.

[0075] S3: Input the time-frequency graph feature vector and the spectrum graph feature vector into the trained feature fusion network. The feature fusion network performs linear processing on the time-frequency graph feature vector, the spectrum graph feature vector, and the average value of the time-frequency graph feature vector and the spectrum graph feature vector through a linear layer to obtain matrices V, K, and Q respectively.

[0076] The matrices V, K, and Q are fused using an attention mechanism to obtain the fused feature vector y:

[0077]

[0078] The feature vector y is then added to the feature vectors of the time-frequency plot and the spectrum plot. The resulting sum is then passed through a fully connected layer (FC) and a sigmoid activation function to obtain the fusion result. This fusion result is a 1×N one-dimensional vector, where N is the total number of interference types that the interference signal to be identified can contain, and each element represents an interference type.

[0079] The loss functions for both the Mobile-Former network and the feature fusion network described above are:

[0080]

[0081] in, p is a label for the nth type of single interference, with a value of 1 or 0. n Let be the output value of the nth type of single disturbance in the output vector of the neural network.

[0082] S4: Compare each element in the fusion result with the decision threshold ε. Record the element greater than the decision threshold as 1 and the element less than the decision threshold as 0 to obtain the recognition result. If there is an element with a value of 1 in the recognition result, the interference signal has the interference type corresponding to that element.

[0083] The expression for the decision threshold ε is as follows:

[0084]

[0085] Among them, F1 n Let be the average F1 score of the nth type of interference, and we have:

[0086]

[0087] Precision n and Recall n Let be the precision and recall of the nth type of interference, respectively, and their expressions are as follows:

[0088]

[0089]

[0090] TP n For the number of samples correctly identified as the nth type of interference, FP n FN represents the number of other interference samples identified as the nth type of interference. n This represents the number of nth type of interference samples that were identified as other types of interference.

[0091] This implementation uses the Mobile-Former network as the backbone, which is a combination of the MobileNet and Transformer networks. It leverages the advantages of MobileNet in extracting local features and the advantages of Transformer in extracting global features to extract features from the signal's time-frequency plot and spectrogram, respectively. This implementation constructs a feature fusion network based on an attention mechanism, efficiently fusing the extracted time-frequency plot features and spectrogram features to obtain richer and more comprehensive features, thereby effectively improving the recognition performance of interference signal types under low interference-to-noise ratio (JNR). In the network output section, this implementation treats the recognition problem of composite interference signals as a multi-label classification problem, applying a threshold decision to the network's output vector to obtain the recognition results for single and composite interference signals. The decision threshold is selected to maximize the average F1 score of the recognition results. The F1 score can be seen as a weighted average of the accuracy and recall of the recognition results.

[0092] Specific Implementation Method Two: The communication interference signal identification method based on deep learning described in this implementation method is as follows:

[0093] Step 1: Construct an interference signal dataset;

[0094] Six types of single interference were generated using MATLAB: single-tone (ST) interference, multi-tone (MT) interference, linear frequency sweep (LFM) interference, quadratic frequency sweep (QFM) interference, frequency hopping (FH) interference, and partial frequency band noise (PBN) interference. Considering the frequency characteristics of the interference signals, some single interferences were selected and combined to obtain composite interference, including multi-tone + linear frequency sweep (MT / LFM) interference, multi-tone + frequency hopping (MT / FH) interference, multi-tone + frequency hopping + linear frequency sweep (MT / FH / LFM) interference, multi-tone + partial frequency band (MT / PBN) interference, frequency hopping + linear frequency sweep (FH / LFM) interference, and linear frequency sweep + partial frequency band (LFM / PBN) interference. Each type of interference signal was superimposed with additive white Gaussian noise in the interval with an interference-to-noise ratio of [-16dB, 10dB], generating M (500) time-domain signal samples every 2dB, with N (4096) sampling points. Short-time Fourier transform and Fourier transform are performed on the sampled signals to obtain time-frequency plots and spectrum plots. These plots are used as the interference signal dataset. Each sample in the dataset is labeled with a 1×6 one-dimensional vector, where each element represents a single type of interference, in the following order: quadratic frequency sweep, single tone, multi-tone, linear frequency sweep, frequency hopping, and partial frequency band. Values ​​are 0 or 1, where 0 indicates the sample does not contain this type of interference, and 1 indicates the sample contains it. For example, [0,0,1,0,1,0] represents multi-tone + frequency hopping interference. The dataset is divided into training, validation, and test sets in a 6:2:2 ratio. The validation set is used to check the network's state and convergence during training and to determine when to stop training.

[0095] Step 2: Build the Mobile-Former network;

[0096] This implementation utilizes a Mobile-Former network for feature extraction from the time-frequency graph and the spectrogram, i.e. Figure 1 The feature extraction network in the example, the Mobile-Former network structure is as follows: Figure 2As shown, the Mobil-Former network parallelizes MobileNet and Transformer, connecting them via a bidirectional bridge. The input to the Mobile-Former network consists of two parts: the original image and trainable global feature tokens. In this implementation, the original image is a time-frequency map or a spectrogram, and the global feature tokens are a set of randomly initialized one-dimensional vectors used to represent the global information of the input image. The values ​​of the global feature tokens are continuously adjusted during network training to better reflect the global information of the image. The stem block preprocesses the input original image through convolutional layers. The preprocessing process involves sequentially performing 3×3 Conv convolutions, batch normalization, and ReLU6 activation on the input original image to obtain a two-dimensional initial local feature map. The Mobile-Former network is composed of 12 stacked Mobile-Former modules. The specific structural details of the Mobile-Former modules are as follows... Figure 3 As shown. The input to the Mobile-Former module is a two-dimensional local feature map X and a one-dimensional global feature token Z. The output is the local feature map X' updated by the Mobile-Former module and the updated global feature token Z'. The Mobile-Former module consists of four sub-modules: the Mobile block, the Former block, the Mobile→Former block, and the Mobile←Former block. The Mobile block takes the local feature map X as input and consists of a 1×1 Conv convolutional layer, a 3×3 DW-Conv convolutional layer, and a DY-ReLU activation function layer. θ is used to adjust the parameters of the DY-ReLU activation function. The Former block is a standard Transformer block, consisting of a multi-head attention (Multi-head Attn) and a feedforward neural network (FFN). The Mobile→Former block is used to integrate the local feature map X into the global feature token Z through a linear transformation matrix W. Q 'A linear transformation is performed on Z to obtain matrix Q', where the linear transformation expression is Q' = ZW Q' W Q' It is a randomly initialized and trainable parameter matrix. Then, attention is calculated between Q' and X. The expression for this attention calculation is: The Mobile←Former block is used to incorporate the global feature token Z' output by the Former block into the local feature X processed by the Mobile block. M In the process, through the linear transformation matrix W K '、W V' After performing linear transformations on Z', we obtain matrices K' and V', respectively. The linear transformation expressions are K' = ZW. K' V' = ZW V' Among them, WK' W V' All parameters are randomly initialized and trainable matrices. Then, attention is calculated between K', V', and the local features processed by the Mobile block. The expression for this attention calculation is: Figure 2 In the Mobile-Former network, the two-dimensional local feature map is transformed into a one-dimensional vector by global average pooling (AGP) at the output end and concatenated with the global feature token. The extracted features are then output through a fully connected layer (FC), which is either a time-frequency map feature vector (FeatureTF) or a spectrum map feature vector (FeatureF).

[0097] Step 3: Construct a feature fusion network;

[0098] This implementation extracts features from the time-frequency graph and spectrogram using a Mobile-Former network. A feature fusion network is then constructed to fuse the time-frequency graph features and spectrogram features extracted in step two. The feature fusion network structure is as follows: Figure 4 As shown, the input to the feature fusion network is the output of two feature extraction networks: the time-frequency graph feature vector (FeatureTF) and the spectrum graph feature vector (FeatureF). The average feature vector of these two vectors is then calculated. Using linear transformation matrices, respectively, FeatureTF, FeatureF, and A linear transformation yields matrices V, K, and Q, which are then fused using an attention mechanism to obtain the fused feature vector. The feature vector y, the time-frequency plot feature vector FeatureTF, and the spectrum plot feature vector FeatureF are then added together. Finally, the result is passed through a fully connected layer (FC) and a sigmoid activation function to output a 1×6 one-dimensional vector.

[0099] Step 4: Train the network;

[0100] In this embodiment, the neural network extracts features from the time-frequency graph and the spectrogram using two Mobile-Former networks, and then fuses these features using a feature fusion network. During training, the two feature extraction networks are first trained separately to achieve the best possible recognition performance using individual features. Then, the feature fusion network is added after the trained feature extraction networks, keeping the parameters of the feature extraction networks unchanged. Only the feature fusion network is trained, fusing the features extracted by the two feature extraction networks to obtain better recognition performance.

[0101] In this implementation, the loss function for both the feature extraction network and the feature fusion network during training is the same, and is expressed as:

[0102]

[0103] in, p is a label for the nth type of single interference, with a value of 1 or 0. n For example, the output value of the nth type of single disturbance in the output vector of the neural network, such as Figure 4 As shown, since the 1×6 output vector of the neural network has been processed by the sigmoid activation function, therefore p n The value is in the range (0,1), and its magnitude matches the 0 and 1 in the label, thus calculating the loss.

[0104] The goal of this loss function is to minimize the difference between the network output vector and the actual label vector, so that the value of the interference signal type present in the sample in the output vector approaches 1, while the value of the interference signal type not present in the sample in the output vector approaches 0.

[0105] During training, the neural network adjusts its parameters using backpropagation and gradient descent based on the loss function of the test set, continuously improving its recognition performance. After each training epoch, the network's performance is monitored in real time using data from the validation set to check its state, convergence, and determine when to stop training.

[0106] Step 5: Select the decision threshold and output the recognition result;

[0107] In this implementation, a threshold decision is applied to the network's output vector to obtain the final recognition result. The neural network output is a 1×6 one-dimensional vector, where each element represents a single type of interference, and the value of each element is within the interval (0,1). Values ​​in the output vector greater than the threshold are set to 1 to indicate the presence of this single type of interference, and values ​​less than the threshold are set to 0 to indicate the absence of this single type of interference. For example, if the threshold is 0.7 and the output vector is [0.1,0.2,0.8,0.1,0.9,0.2], then the recognition result after threshold decision is [0,0,1,0,1,0], which is identified as multi-tone + frequency hopping interference.

[0108] The decision threshold ε is selected as the value that maximizes the average F1 score of the network recognition results, i.e.:

[0109]

[0110]

[0111]

[0112]

[0113] Among them, TPn For the number of samples correctly identified as the nth type of interference, FP n FN represents the number of other interference samples identified as the nth type of interference. n This represents the number of nth type of interference samples that were identified as other types of interference.

[0114] This implementation uses Python to build the network, with a GPU (NVIDIA RTX A5000) as the simulation environment and PyTorch 2.0.1 as the deep learning framework. The network's input batch size is 16, and the learning rate is 0.0001. During testing, the time-frequency graph and spectrogram of the test set are input into the trained network to obtain the identification results of interference signals.

[0115] Figure 5 , Figure 6 These are the confusion matrices at interference ratios (IRRs) of -16dB and -10dB. At an IRR of -16dB, the confusion matrix is ​​relatively clear overall, indicating that the network still has good recognition performance at low IRRs. At an IRR of -10dB, the recognition performance of interference signals is further significantly improved.

[0116] Figure 7 The F1 score curves show the identification results of 12 types of interference signals. When the interference-to-noise ratio (IRR) is -16dB, the F1 scores are concentrated between 0.6 and 0.7. The F1 scores of single-tone and multi-tone + partial frequency band interference are greater than 0.8, and only frequency hopping + linear sweep is below 0.5. As the IRR increases, the F1 scores gradually increase. When the IRR is greater than -10dB, the F1 scores of all interference signal identification results are close to 1, indicating that there is still good interference signal identification performance under low IRR.

[0117] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.

Claims

1. A communication jamming signal identification method based on a Mobile-Former network, characterized in that, Specifically as follows: Short-time Fourier transform and Fourier transform are performed on the interference signals to be identified to obtain the time-frequency diagram and the spectrum diagram. The time-frequency graph and the spectrum graph are respectively input into two trained Mobile-Former networks to extract the time-frequency graph feature vector and the spectrum graph feature vector; The time-frequency graph feature vector and the spectrum graph feature vector are input into the trained feature fusion network and fused to obtain a fusion result. The fusion result is a 1×N one-dimensional vector, where N is the total number of interference types that can be included in the interference signal to be identified, and each element represents an interference type. Each element in the fusion result is compared with a decision threshold. Elements greater than the decision threshold are marked as 1, and elements less than the decision threshold are marked as 0, thus obtaining the recognition result. If there is an element with a value of 1 in the recognition result, the interference signal has the interference type corresponding to that element.

2. The communication jamming signal identification method based on the Mobile-Former network according to claim 1, characterized in that, The process of extracting feature vectors by the Mobile-Former network is as follows: The original input image is preprocessed to obtain an initial two-dimensional local feature map; The initial local feature map and the initial global feature token are updated sequentially through multiple Mobile-Former modules to obtain the updated local feature map and global feature token; The updated local feature map is transformed into a one-dimensional vector by global average pooling and then concatenated with the updated global feature token. The concatenated result is output through a fully connected layer to obtain the extracted feature vector; The input original image is a time-frequency graph or a spectrum graph; The initial global feature token is a set of randomly initialized one-dimensional vectors used to represent the global information of the original image.

3. The communication jamming signal identification method based on the Mobile-Former network according to claim 2, characterized in that, The preprocessing of the input raw image includes: The input original image is sequentially subjected to 3×3 Conv convolution, batch normalization, and ReLU6 activation to obtain a two-dimensional initial local feature map.

4. The communication jamming signal identification method based on the Mobile-Former network according to claim 2, characterized in that, The number of the Mobile-Former modules is I, and the local feature map X output by the i-th Mobile-Former module i with the global feature token Z i as the input of the i+1-th Mobile-Former module, i=0, 1,..., I-1, and when i=0, the input of the Mobile-Former module is the initial local feature map and the initial global feature token; The Mobile-Former module includes: a Mobile submodule, a Former submodule, a Mobile-to-Former submodule, and a Former-to-Mobile submodule; The input of the Mobile submodule is a local feature map X i The local feature map X i The output result obtained by sequentially passing through a first 1*1Conv convolutional layer, a first DY-ReLU activation function layer, a 3*3DW-Conv convolutional layer, a second DY-ReLU activation function layer and a second 1*1Conv convolutional layer is input to the Former-to-Mobile submodule; The Mobile-to-Former submodule is configured to incorporate the local feature maps X i into a global feature token Z i and input the result to the Former submodule. The Former submodule is a Transformer block, and the output is a global feature token Z i+1 ; The Former-to-Mobile submodule is configured to input the global feature tokens Z output by the Former submodule i+1 into the Mobile submodule output, to obtain the local feature map X i+1 .

5. The communication jamming signal identification method based on the Mobile-Former network according to claim 4, characterized in that, The local feature map X i is fused into the global feature token Z i , comprising: The global feature token Z i Linear processing is performed by a linear layer, followed by local feature maps X i Attention calculation is performed.

6. The communication jamming signal identification method based on the Mobile-Former network according to claim 4 or 5, characterized in that, The global feature token Z output by the Former submodule i+1 is fused into the result output by the Mobile submodule, including: The global feature token Z i+1 Linear processing is performed by a linear layer, and then attention calculation is performed with the local features that have passed through the Mobile sub-module.

7. The communication jamming signal identification method based on the Mobile-Former network according to claim 6, characterized in that, The global feature token Z output by the Former submodule i+1 After the multilayer perceptron, the adjustment parameter θ = {a 1 ,a 2} is obtained, and the adjustment parameter θ = {a 1 ,a 2} is used to adjust the first DY-ReLU activation function and the second DY-ReLU activation function. The expressions for both the first and second DY-ReLU activation functions are: f(x) = max{a 1 x,a 2 x} where a 1 and a 2 are two components of the adjustment parameter θ, and x is the variable of the DY-ReLU activation function.

8. The communication jamming signal identification method based on the Mobile-Former network according to claim 1, characterized in that, The fusion process of the feature fusion network is as follows: The time-frequency graph feature vector, the spectrum graph feature vector, and the average value of the time-frequency graph feature vector and the spectrum graph feature vector are linearly processed through a linear layer to obtain matrices V, K, and Q, respectively. The matrices V, K, and Q are fused using an attention mechanism to obtain the fused feature vector y: y = softmax(QK T V; The feature vector y is then added to the feature vectors of the time-frequency plot and the spectrum plot. The resulting sum is then passed through a fully connected layer and a sigmoid activation function to obtain the fusion result.

9. The communication jamming signal identification method based on the Mobile-Former network according to claim 1, characterized in that, The loss functions for both the Mobile-Former network and the feature fusion network are: wherein, is the label of the nth single interference and takes the value of 1 or 0, p n is the output value of the nth single interference in the output vector of the neural network.

10. The communication jamming signal identification method based on the Mobile-Former network according to claim 1, characterized in that, The expression for the decision threshold ε is as follows: F1 n F1is the average F1 score for the nth type of interference, and has: Precision n and Recall n The precision and recall of the nth class of interference are respectively expressed as follows: TP n the number of samples correctly identified as the nth class of interference, FP n the number of samples of other classes of interference identified as the nth class of interference, FN n the number of samples of the nth class of interference identified as other classes of interference.

Images