A system and method for identifying jammed electromagnetic signals based on diffusion Sigformer

Abstract

This invention belongs to the field of electromagnetic signal recognition technology and discloses a system and method for identifying interfered electromagnetic signals based on diffusion Sigformer. The system for identifying interfered electromagnetic signals includes a diffusion model DDPM and a classification model Sigformer. The classification model Sigformer includes a 1D Patch Embedding module, a Sigformer Block module, and a classifier module. The diffusion model DDPM denoises the interfered electromagnetic signal to obtain the denoised signal. The 1D Patch Embedding module divides the denoised signal sequence into blocks, adds positional encoding, and preserves the temporal relationship of the original sequence. Residual convolution is introduced into the original Transformer Encoder to construct the Sigformer Block module, which is used to extract features from the denoised signal to obtain information on different modulation categories. This allows the system to better focus on local features without losing global information. The classifier module receives the extracted deep features and outputs the predicted probability of each modulation category.

Description

Technical Field

[0001] This invention belongs to the field of electromagnetic signal recognition technology, and relates to a system and method for identifying interfered electromagnetic signals based on diffused Sigformer. Background Technology

[0002] With the rapid development of information technology, electromagnetic signals, as carriers of information transmission, are becoming increasingly important. Electromagnetic signal identification (EMI), as a crucial branch of signal processing, not only has profound theoretical significance but also demonstrates broad application value in both civilian and military fields. In the civilian sector, EII technology can be used to monitor civilian communication signals, identify interference, and manage the electromagnetic spectrum. By monitoring and analyzing electromagnetic signals in real time, it is possible to supervise the lawful operation of civilian communications and monitor unauthorized users' encroachment and misappropriation of satellite transponder resources. Furthermore, EII technology can be used in smart homes, the Internet of Things (IoT), and other fields to achieve intelligent control and information interaction of devices. In the military field, accurate signal identification is essential for conducting electromagnetic countermeasures in space and gaining information superiority. By identifying and analyzing enemy electromagnetic signals, it is possible to intercept enemy forces, conduct electromagnetic interference and attacks, providing strong support for military operations. In addition, EII technology can also be used in radar signal processing, electronic warfare, and other fields to improve operational efficiency and security.

[0003] In recent years, with the continuous development of artificial intelligence, deep learning technology has attracted widespread attention in many fields, and deep learning methods have become a research hotspot in the field of electromagnetic signal recognition. By constructing neural network models, deeper information can be learned from different electromagnetic signal data, improving the accuracy and efficiency of recognition. However, with the rapid development of wireless communication technology, diverse and numerous electromagnetic waves overlap and change in multiple dimensions such as time, frequency, and space. Furthermore, wireless channels are highly time-varying, with numerous interference factors such as channel fading and background noise. The electromagnetic environment is becoming increasingly complex, and electromagnetic signals are dynamically unstable. These factors pose significant challenges to signal recognition models. Deep learning models are easily affected by various interferences, leading to a decrease in recognition accuracy. To meet practical needs, developing a disturbance-resistant electromagnetic signal recognition model is crucial.

[0004] Existing electromagnetic signal identification methods have the following problems:

[0005] 1. Deep learning models are easily affected by various noises, leading to decreased accuracy and poor robustness, which cannot meet practical needs.

[0006] 2. CNN-based methods cannot establish long-range dependencies on global information and cannot fully extract temporal and detailed information from signals. Summary of the Invention

[0007] To address the technical problems existing in the prior art, this invention provides a system and method for identifying interfered electromagnetic signals based on diffused Sigformer.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] This invention provides a system for identifying interfered electromagnetic signals based on the diffusion Sigformer, comprising: a diffusion model DDPM for denoising the interfered electromagnetic signal to obtain a clean signal; and a classification model Sigformer for identifying the modulation scheme of the clean signal.

[0010] Furthermore, the classification model Sigformer includes: a 1D Patch Embedding module: dividing the clean signal sequence into blocks, converting it into a form that the Transformer Encoder can receive, adding positional encoding, and preserving the temporal relationship of the original sequence; a Sigformer Block module: used to extract features from the clean signal to obtain information on different modulation categories; and a classifier module: used to receive the extracted deep features and output the predicted probability of each modulation category.

[0011] Furthermore, the 1D Patch Embedding module includes a patch one-dimensional convolution and a learnable position encoding vector; the patch one-dimensional convolution divides the interference-free signal into a series of patches; the learnable position encoding vector is used to retain the positional information between each patch.

[0012] Furthermore, the classifier module consists of linear layers, taking the final output of the Sigformer Block module as input, and outputting the predicted probability of each modulation category.

[0013] Furthermore, the diffusion model DDPM is trained using the mean squared error loss function, and the classification model Sigformer is trained using the cross-entropy loss function.

[0014] Furthermore, the expression for the mean squared error loss function is:

[0015]

[0016] Where N is the number of samples, and ∈ represents the true noise. To predict noise.

[0017] Furthermore, the expression for the cross-entropy loss function is:

[0018]

[0019] Where N is the number of samples, K is the number of categories, and y ik Let i be the true label of the i-th sample belonging to the k-th category. Let be the predicted probability that the i-th sample belongs to the k-th category.

[0020] Furthermore, the backbone of the diffusion model DDPM is Unet, which is used to predict the amount of added noise.

[0021] This invention also provides a method for identifying interfered electromagnetic signals based on a diffusion Sigformer, comprising the following steps: inputting the disturbed signal into a diffusion model DDPM for de-perturbation, outputting a clean signal as input to a 1D Patch Embedding module, wherein the 1D Patch Embedding module converts the clean signal into a token, adds position encoding and a class token for classification, and inputs it to a Sigformer Block module for feature extraction, wherein the Sigformer Block module uses the final output as input to a classifier module, and outputs the predicted probability of each modulation category.

[0022] Furthermore, the Sigformer Block module first performs a self-attention operation on the output of the 1D Patch Embedding module, adds the original input and the self-attention result to the one-dimensional convolution, adds the one-dimensional convolution result to the one-dimensional convolution input, and inputs it to the multilayer perceptron. Finally, the result of the multilayer perceptron is added to the input of the multilayer perceptron as the final output of the Sigformer Block module.

[0023] Compared with the prior art, the present invention has the following beneficial technical effects:

[0024] This invention discloses an interference-affected electromagnetic signal identification system based on the diffusion Sigformer. The system uses the diffusion model DDPM as the interference removal model. During the training process, noise is added to the clean signal for processing. The noise-added signal is used as input to predict the added noise. During the testing process, the system receives the interference signal as input and performs noise removal processing to obtain the interference-free signal.

[0025] This invention discloses an interference-affected electromagnetic signal identification system based on diffusion Sigformer. The 1D PatchEmbedding module converts the interference-free signal into token form and adds position encoding, while retaining the timing information in the original signal.

[0026] This invention provides a system for identifying interfered electromagnetic signals based on diffused Sigformer. The Sigformer Block module is used to extract features from the signal. Based on Transformer Encoder, residual convolution is added after self-attention, so that the classification model Sigformer can better focus on local detailed features without losing global information. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the operation of an interference-affected electromagnetic signal identification system based on diffused Sigformer according to the present invention.

[0028] Figure 2 This is a schematic diagram of the working process of the diffusion model DDPM in an embodiment of the present invention;

[0029] Figure 3 This is a schematic diagram illustrating the operation of the Sigformer classification model in an embodiment of the present invention. Detailed Implementation

[0030] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0031] Example 1

[0032] like Figure 1 He Ru Figure 2 As shown, this invention discloses an interference-affected electromagnetic signal identification system based on a diffusion Sigformer, comprising a diffusion model DDPM and a classification model Sigformer. The classification model Sigformer consists of a 1D Patch Embedding module, a Sigformer Block module, and a classifier module. The diffusion model DDPM receives the interference-affected electromagnetic signal as input and performs denoising processing to obtain the denoised signal. Specifically, firstly, noise is added to the clean signal. The intensity and type of noise can be adjusted according to the actual application scenario. The denoised signal is then used as input to the diffusion model DDPM to predict the added noise. During testing, the interference-affected signal is received as input and denoised to obtain the denoised signal.

[0033] The Diffusion-Driven Propagation Model (DDPM) has achieved great success in image generation, and research based on it continues to emerge. We apply it to the field of electromagnetic signals, and its basic idea can be explained as follows: It progressively adds noise to a clean signal. As the amount of noise increases, the signal becomes increasingly chaotic. When the noise is sufficiently large, the original clean signal almost becomes pure noise. During the noise addition process, neural networks such as Unet are used to predict the amount of noise added. For denoising, a noisy signal is input, and the neural network progressively denoises it, ultimately obtaining the de-noising signal.

[0034] Specifically, the diffusion model DDPM treats the noise addition process as a Markov process, modeling the process from x1 to x0 given x0. t By understanding the joint distribution of these distributions, we can model the entire process. Combining this with the Markov property, we obtain:

[0035]

[0036] Where x1 is the value of x at time 1, T is the final time, x0 is the value of x at the initial time, i.e., the input of the diffusion model DDPM, t is a certain time, q is the true distribution, and x t Let x be the value of x at time t. t-1 Let x be the value of x at time t-1, N be a normal distribution, and β be the value of x at time t-1. t I represents a series of fixed values ​​that vary linearly or cosinely, where I is the unit variance.

[0037] As can be seen, each state node in a Markov chain conforms to a Gaussian distribution and is only related to the previous state. That is, the representation of the current distribution is determined by the previous observation. If the current distribution is known, the value of the current sample can be obtained using the reparameterization technique.

[0038]

[0039] Where, α t =1-β t , In the diffusion model DDPM, reparameterization can better represent the recursive relationship between random variables. After reparameterization, x can be recursively obtained from the original input x0. T :

[0040]

[0041] Where, x T The value of x at time T is already pure Gaussian noise.

[0042] The noise reduction process involves converting pure Gaussian noise into x TIterate back to x0 step by step. For the true backward transition distribution function, use Bayes' theorem:

[0043]

[0044] Since the initial distribution is unknown, the marginal distribution cannot be directly derived from the joint distribution. However, given the transition probability distributions of each state, and using the properties of Markov chains, we can obtain:

[0045]

[0046] After reparameterization, we can obtain:

[0047]

[0048] in,

[0049] Since the noise is unknown, fitting the distribution of x0 is essentially a problem of finding the optimal parameter estimate using a probabilistic model. The most classic method is maximum likelihood estimation. Therefore, the optimization objective is to minimize the negative log-likelihood of the sample - log(p). θ (x0)) is equivalent to minimizing KL[q(x0)]. t-1 |x t ,x0)|p θ (x t-1 |x t )).

[0050] Where KL is the KL divergence, p θ This is the predicted distribution.

[0051] KL[q(x t-1 |x t ,x0)|p θ (x t-1 |x t Simplify to:

[0052]

[0053] Where, ∈ θ This is a function of the predicted noise fitted to the model.

[0054] This transforms the problem into an optimization problem for a noise predictor. During the forward noise addition process of the DDPM diffusion model, the ratio of noise to signal coefficients changes with time t. To satisfy the defined perturbation rate relationship, we fix the time t and set the coefficients of both noise and signal to a constant. Thus, the coefficient ratio between the two is 1. Therefore, to set different perturbation rates, we only need to change the coefficient of noise ∈.

[0055] We define the signal perturbation rate as the ratio of the root mean square amplitude of the perturbation noise to the root mean square amplitude of the original electromagnetic signal:

[0056]

[0057] We investigate the impact of three types of noise disturbances on the accuracy of electromagnetic signal recognition: Gaussian noise, Rayleigh noise, and periodic noise. Periodic noise is noise that repeats with a specific frequency and amplitude. It usually presents as a sine or cosine waveform and has a distinct periodic structure.

[0058] The definitions of Gaussian noise and Rayleigh noise are as follows:

[0059]

[0060] Where σ is the variance, x is the observed value, and μ is the mean.

[0061] Inspired by the ViT concept, we also segment the signal into several patches. However, unlike images, signals are one-dimensional data, so we perform 1D patch embedding on them. Similarly, we add positional encoding to preserve the temporal information in the signal.

[0062] Furthermore, we found that directly using the Transformer encoder to extract signal features resulted in mediocre classification accuracy and unstable training. Therefore, we modified the Transformer encoder. Specifically, we added residual convolutions between the Attention and MLP layers to enable the model to better focus on fine-grained local features without losing global contextual information, while also making training more stable. We named the modified module the Sigformer Block module, as follows: Figure 3 As shown.

[0063] Let the input X∈R C*L Its forward propagation process is as follows:

[0064] X = Attention(X) + X

[0065] X = Conv(X) + X

[0066] X = MLP(X) + X

[0067] Where Attention(·) represents self-attention operation, Conv(·) represents convolution operation, and MLP(·) represents multilayer perceptron.

[0068] The output of the last Sigformer Block is used as the input to the classifier, and the classifier model outputs the predicted probability for each class.

[0069] The process of training the diffusion model DDPM essentially involves reducing the distributional difference between predicted noise and actual noise. We optimize this process using the mean squared error loss function (MSE):

[0070]

[0071] Where ∈ represents real noise. To predict noise.

[0072] The 1D Patch Embedding module divides the noise-reduced signal sequence into blocks, converts it into a format that the Transformer Encoder can receive, adds positional encoding, and preserves the temporal relationship of the original sequence. The Sigformer Block module extracts features from the noise-reduced signal to obtain information about different modulation categories. The classifier module receives the extracted deep features and outputs the predicted probability for each modulation category. The 1D Patch Embedding module, Sigformer Block module, and classifier module together form the Sigformer classification model, which is optimized using the cross-entropy loss function.

[0073]

[0074] Where N is the number of samples, K is the number of categories, and y ik Let i be the true label of the i-th sample belonging to the k-th category. Let be the predicted probability that the i-th sample belongs to the k-th category.

[0075] Experiments were conducted using data with a signal-to-noise ratio of 18dB selected from the RML2016.10a and RML2016.10b datasets to investigate the recognition capability of Sigformer on relatively clean signals.

[0076] Table 1. Recognition Accuracy of RML2016.10a

[0077]

[0078] Table 2. RML 2016.10b Recognition Accuracy

[0079]

[0080] Tables 1 and 2 show the results of Sigformer and the comparison algorithms on the RML2016.10a and RML2016.10b datasets. It can be seen that Sigformer outperforms the other algorithms in accuracy on both datasets.

[0081] Using data with a signal-to-noise ratio of 18dB from the RML2016.10a dataset, three types of noise were selected, and 20 consecutive feature points in half of the samples in each batch were perturbed to explore the recognition effect of the present invention on the disturbed signal.

[0082] Table 3. Accuracy of Sigformer in identifying interference signals (Gaussian noise)

[0083] Disturbance rate 1 2 3 4 5 6 7 8 9 10 Before noise reduction 0.707 0.620 0.595 0.572 0.560 0.557 0.555 0.551 0.549 0.546 After noise reduction 0.853 0.860 0.858 0.857 0.853 0.852 0.862 0.854 0.855 0.851

[0084] Table 4. Accuracy of Sigformer in identifying interference signals (Rayleigh noise)

[0085] Disturbance rate 1 2 3 4 5 6 7 8 9 10 Before noise reduction 0.709 0.690 0.677 0.662 0.649 0.641 0.635 0.632 0.629 0.623 After noise reduction 0.794 0.795 0.798 0.790 0.794 0.791 0.788 0.792 0.797 0.802

[0086] Table 5. Accuracy of Sigformer in identifying interference signals (periodic noise)

[0087] Disturbance rate 1 2 3 4 5 6 7 8 9 10 Before noise reduction 0.648 0.614 0.588 0.582 0.582 0.582 0.575 0.566 0.563 0.562 After noise reduction 0.838 0.846 0.849 0.847 0.852 0.845 0.845 0.842 0.845 0.837

[0088] Tables 3-5 show the recognition results of the diffused Sigformer for interfered signals under three types of noise perturbations. Before denoising, the Sigformer's recognition accuracy was tested directly using data with added noise perturbations, to study the dynamic relationship between model accuracy and perturbation rate as the perturbation rate varied from 1 to 10. After denoising, the DDPM diffusion model was used for denoising before model recognition, and the denoised data was then input into the model for recognition. The data in the tables show that as the perturbation rate changes from 1 to 10, the more noise perturbation added, the lower the model's recognition accuracy becomes. However, when using the DDPM diffusion model for denoising before recognition, the model's recognition accuracy remained at a high level, demonstrating the effectiveness of using the diffusion model for interference removal. However, the diffusiond Sigformer showed slightly different denoising effects for the three types of noise perturbations, performing better on Gaussian perturbations and worse on Rayleigh perturbations. This is related to the characteristics of the three signals themselves and requires further investigation.

[0089] In summary, electromagnetic signal identification (ESI) holds significant importance and has broad application value in fields such as communication, radar detection, spectrum management, information security, and military applications. In wireless communication, ESI is fundamental to ensuring correct signal demodulation and extracting useful information. By identifying the modulation scheme, parameters, and other characteristics of signals, the performance of communication systems can be optimized, improving reliability and stability. In the military field, ESI plays a crucial role in space electromagnetic warfare and information superiority. Accurately identifying the modulation type of enemy communication signals enables effective electromagnetic interference and intelligence interception, providing strong support for military operations. In radar systems, ESI technology is used to identify the modulation type of radar echo signals, thereby obtaining information such as the distance, speed, and azimuth of target objects. This is of great significance for radar detection and target tracking. Furthermore, ESI helps spectrum management agencies effectively monitor and manage spectrum resources. By identifying the spectrum occupancy of different signals, problems of spectrum resource waste and abuse can be promptly detected and addressed, ensuring the rational and efficient use of spectrum resources. As ESI technology continues to develop and improve, it will play an important role in more fields, bringing greater convenience and security to people's lives and work.

[0090] The backbone of the diffusion model DDPM is Unet, which is used to predict the amount of noise added. Specifically, Unet, as a powerful convolutional neural network architecture, is often used as the backbone of a model. In the diffusion model DDPM, Unet's task is not to directly predict the denoised signal, but to predict the amount of noise that needs to be removed to recover the original clean signal from the current noisy signal.

[0091] Example 2

[0092] This invention also provides a method for identifying interfered electromagnetic signals based on a diffusion Sigformer, comprising: inputting the disturbed signal into a diffusion model DDPM for de-perturbation, outputting the de-perturbed signal as input to a 1D Patch Embedding module, wherein the 1D Patch Embedding module converts the de-perturbed signal into a token, adds position encoding and a class token for classification, and inputs it to a Sigformer Block module for feature extraction, wherein the Sigformer Block module uses the final output as input to a classifier module, and outputs the predicted probability of each modulation category.

[0093] The Disturbed Signal Model (DDPM) takes the following inputs: the Disturbed Signal; the processing involves feeding the signal into the trained DDPM model and gradually removing disturbances through the model's denoising process; and the output is the Denoised Signal.

[0094] The 1D Patch Embedding module takes the following input: the noise-reduced signal. The processing steps include: dividing the signal into a series of small patches, each patch being considered a small signal segment; converting each patch into a fixed-size embedding vector (token), typically achieved through linear transformations or convolutional layers; adding positional encoding (such as sinusoidal positional encoding) to preserve the order information between patches; and adding a class token for classification, which is placed before or after all patch tokens for subsequent classification tasks. The output is a sequence of tokens containing positional encodings and the class token.

[0095] The Sigformer Block module takes a sequence of tokens from a 1D Patch Embedding as input. It processes tokens by incorporating a Transformer-based architecture that may include specific improvements for signal processing (although "Sigformer" is not a standard term, it is assumed here to be similar to a Transformer variant used for signal processing). Internally, the Block may contain self-attention mechanisms to capture dependencies between tokens, as well as feed-forward networks and layer normalization. By stacking multiple Sigformer Blocks, high-level features of the signal can be extracted progressively. The output is a sequence of tokens with extracted features.

[0096] Specifically: The Sigformer Block module first performs a self-attention operation on the output of the 1D Patch Embedding module, adds the original input and the self-attention result to the one-dimensional convolution, adds the one-dimensional convolution result to the one-dimensional convolution input, and adds the result of the one-dimensional convolution to the multilayer perceptron. Finally, the result of the multilayer perceptron is added to the input of the multilayer perceptron, which is the final output of the Sigformer Block module.

[0097] The classifier module takes the extracted deep features as input and outputs the predicted probability for each modulation category as output, along with the final output of the Sigformer Block module. In this embodiment, the interference signal identification system uses a linear layer as the classifier module.

[0098] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

Claims

1. A system for identifying interfered electromagnetic signals based on diffused Sigformer, characterized in that, include: DDPM (Dispersion Diffusion Model): This model denoises the interfered electromagnetic signal to obtain a denoised signal. The Sigformer classification model is used to identify the modulation scheme of the denoised signal. The classification model Sigformer includes a 1D Patch Embedding module, a Sigformer Block module, and a classifier module. The Sigformer Block module adds residual convolution between Attention and MLP, enabling the model to better focus on local fine-grained features without losing global contextual information, while making training more stable. The diffusion model DDPM is trained using the mean squared error loss function, and the classification model Sigformer is trained using the cross-entropy loss function. The expression for the mean squared error loss function is: Where N is the sample size. This is real noise. To predict noise; The expression for the cross-entropy loss function is as follows: Where N is the number of samples and K is the number of categories. Let i be the true label of the i-th sample belonging to the k-th category. Let be the predicted probability that the i-th sample belongs to the k-th category.

2. The interference-affected electromagnetic signal identification system based on diffused Sigformer as described in claim 1, characterized in that, The classification model Sigformer includes: The 1D Patch Embedding module is used to divide the interference-free signal sequence into blocks, convert it into a form that the Transformer Encoder can receive, add position encoding, and save the timing relationship of the original sequence. Sigformer Block module: used to extract features from the de-interference signal to obtain information on different modulation categories; Classifier module: Used to receive the extracted deep features and output the predicted probability for each modulation category.

3. The interference-affected electromagnetic signal identification system based on diffused Sigformer as described in claim 2, characterized in that: The 1D Patch Embedding module includes a patch one-dimensional convolution and a learnable positional encoding vector; The patch one-dimensional convolution divides the interference-free signal into a series of patches; The learnable positional encoding vector is used to preserve positional information between each patch.

4. The interference-affected electromagnetic signal identification system based on diffused Sigformer as described in claim 2, characterized in that: The classifier module consists of linear layers, taking the final output of the Sigformer Block module as input, and outputting the predicted probability of each modulation category.

5. The interference-affected electromagnetic signal identification system based on diffused Sigformer as described in claim 1, characterized in that: The diffusion model DDPM has a Unet backbone, which is used to predict the amount of added noise.

6. A method for identifying interfered electromagnetic signals based on diffused Sigformer, characterized in that, The interference-affected electromagnetic signal identification system based on any one of claims 1-5 includes the following steps: The perturbed signal is input into the diffusion model DDPM for deperturbation, and the output of the deperturbed signal is used as the input of the 1D Patch Embedding module. The 1D Patch Embedding module converts the deperturbed signal into a token, adds position encoding and a class token for classification, and inputs it to the Sigformer Block module for feature extraction. The Sigformer Block module uses the final output as the input of the classifier module and outputs the predicted probability of each modulation category.

7. The method for identifying interfered electromagnetic signals based on diffused Sigformer as described in claim 6, characterized in that: The Sigformer Block module first performs a self-attention operation on the output of the 1D Patch Embedding module, adds the original input and the self-attention result to a one-dimensional convolution, adds the one-dimensional convolution result to the one-dimensional convolution input, and adds the result of the one-dimensional convolution to the multilayer perceptron. Finally, it adds the result of the multilayer perceptron to the input of the multilayer perceptron, which is the final output of the Sigformer Block module.

Images