Diffusion Sigformer-based interfered electromagnetic signal identification system and method

Abstract

The invention belongs to the technical field of electromagnetic signal identification, and discloses an interfered electromagnetic signal identification system and method based on diffusion Sigformer. The interfered electromagnetic signal identification system comprises a diffusion model DDPM and a classification model Sigformer, the classification model Sigformer comprises a 1D Patch Embedding module, a Sigformer Block module and a classifier module, the diffusion model DDPM performs de-noising processing on interfered electromagnetic signals to obtain interference-removed signals, the 1D Patch Embedding module divides the interference-removed signal sequence into blocks, position codes are added, and the Sigformer Block module performs de-noising processing on the interference-removed signals to obtain an interference-removed signal sequence; the time sequence relation of an original sequence is stored, residual convolution is introduced into an original Transform Encoder, and a Sigform Block module is constructed and is used for carrying out feature extraction on signals subjected to interference removal and obtaining information of different modulation categories, so that the system can better pay attention to local features under the condition that global information is not lost; and outputting the prediction probability of each modulation category.

Description

Technical Field

[0001] The invention belongs to the technical field of electromagnetic signal recognition, and relates to a system and method for identifying interfered electromagnetic signals based on a diffusion Sigformer. Background Art

[0002] With the rapid development of information technology, the importance of electromagnetic signals as a carrier of information transmission has become increasingly prominent. As an important branch of signal processing, electromagnetic signal recognition not only has far-reaching theoretical significance, but also shows wide application value in the civil and military fields. In the civil field, electromagnetic signal recognition technology can be used to monitor civil communication signals, conduct interference identification and electromagnetic spectrum management. By real-time monitoring and analysis of electromagnetic signals, it is possible to supervise the legal operation of civil communications and monitor the occupation and theft of satellite transponder resources by illegal users. In addition, electromagnetic signal recognition technology can also be used in smart homes, the Internet of Things and other fields to achieve intelligent control of equipment and information interaction. In the military field, accurate signal recognition is a prerequisite for space electromagnetic confrontation and obtaining information rights. By identifying and analyzing enemy electromagnetic signals, it is possible to intercept the enemy, conduct electromagnetic interference and attack, etc., providing strong support for military operations. In addition, electromagnetic signal recognition technology can also be used in radar signal processing, electronic warfare and other fields to improve combat efficiency and safety.

[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 also become a research hotspot in the field of electromagnetic signal recognition. By constructing a neural network model, we can learn deeper information from different electromagnetic signal data and improve the accuracy and efficiency of recognition. However, with the rapid development of wireless communication technology, various types and large numbers of electromagnetic waves overlap and change in multiple dimensions such as time / frequency / space, and the wireless channel is highly time-varying. There are many interference factors such as channel fading and background noise. The electromagnetic environment is becoming increasingly complex, and the electromagnetic signal is dynamically unstable. These factors make the signal recognition model face major challenges. Deep learning models are easily affected by various interferences, which leads to a decrease in recognition accuracy. In order to meet practical needs, it is crucial to develop an electromagnetic signal recognition model that is resistant to disturbances.

[0004] The existing electromagnetic signal recognition has the following problems:

[0005] 1. Deep learning models are easily interfered by various noises, resulting in decreased accuracy and poor robustness, which cannot meet actual needs.

[0006] 2. The CNN-based method cannot establish long-distance dependencies on global information and cannot fully mine the timing and detail information of the signal. Summary of the invention

[0007] In order to solve the technical problems existing in the prior art, the present invention provides a system and method for identifying interfered electromagnetic signals based on a diffusion Sigformer.

[0008] In order to achieve the above object, the present invention adopts the following technical solution:

[0009] The present invention provides a system for identifying interfered electromagnetic signals based on diffusion Sigformer, comprising: a diffusion model DDPM: performing denoising processing on interfered electromagnetic signals to obtain clean signals; and a classification model Sigformer: used for identifying the modulation mode of the clean signals.

[0010] Furthermore, the classification model Sigformer includes: a 1D Patch Embedding module: which divides the clean signal sequence into blocks, converts it into a form that can be received by the Transformer Encoder, adds position coding, and preserves the temporal relationship of the original sequence; a Sigformer Block module: which is used to extract features from the clean signal and obtain information of different modulation categories; a classifier module: which is 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 de-interferenced signal into a series of patches; the learnable position encoding vector is used to retain the position information between each patch.

[0012] Furthermore, the classifier module is composed of a linear layer, takes the final output of the Sigformer Block module as input, and outputs the predicted probability of each modulation category.

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

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

[0015]

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

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

[0018]

[0019] Among them, N is the number of samples, K is the number of categories, and y ik is the true label of the i-th sample belonging to the k-th category, is 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, and the Unet is used to predict the amount of added noise.

[0021] The present invention also provides a method for identifying interfered electromagnetic signals based on diffusion Sigformer, comprising the following steps: inputting the disturbed signal into a diffusion model DDPM for de-perturbation, outputting a clean signal as an input to a 1D PatchEmbedding module, the 1D Patch Embedding module converting the clean signal into a token, adding a position code and a class token for classification, and inputting the token into a Sigformer Block module for feature extraction, the SigformerBlock module using the final output as an input to a classifier module, and outputting 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 to the self-attention result and inputs it into a one-dimensional convolution, adds the one-dimensional convolution result to the one-dimensional convolution input, and inputs it into a multi-layer perceptron, and adds the result of the multi-layer perceptron to the input of the multi-layer 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] The invention discloses an interfered electromagnetic signal recognition system based on diffusion Sigformer. The diffusion model DDPM is used as a de-interference model. Noise is added to a clean signal during training to process the clean signal. The noisy signal is used as input to predict the added noise. The interfered signal is received as input during testing to perform de-noising processing to obtain a de-interferenced signal.

[0025] The present invention discloses an interfered electromagnetic signal recognition system based on diffusion Sigformer. The 1D PatchEmbedding module converts the de-interfered signal into a token form and adds a position code, thereby retaining the timing information in the original signal.

[0026] The present invention discloses an interfered electromagnetic signal recognition system based on a diffusion Sigformer. The Sigformer Block module is used for extracting features from signals. Based on a Transformer Encoder, a residual convolution is added after self-attention, so that the classification model Sigformer can better focus on local detail features without losing global information. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a working schematic diagram of a system for identifying interfered electromagnetic signals based on diffusion Sigformer of the present invention;

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

[0029] Figure 3 Schematic diagram of the working of the classification model Sigformer in an embodiment of the present invention. DETAILED DESCRIPTION

[0030] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0031] Example 1

[0032] like Figure 1 He Ru Figure 2 As shown, the present invention is a system for identifying interfered electromagnetic signals based on diffusion Sigformer, including a diffusion model DDPM and a classification model Sigformer, wherein the classification model Sigformer is composed of a 1D Patch Embedding module, a Sigformer Block module and a classifier module. The diffusion model DDPM receives the interfered electromagnetic signal as input, performs denoising processing to obtain a denoised signal, specifically, firstly, the clean signal is subjected to noise processing, the intensity and type of the noise can be adjusted according to the actual application scenario, and the denoised signal is used as the input of the diffusion model DDPM, and the added noise is predicted. During the test process, the interfered signal is received as input, and denoising processing is performed on it to obtain a denoised signal.

[0033] The diffusion model DDPM has achieved great success in the field of 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 gradually adds noise to the clean signal. As the amount of noise continues to increase, the signal becomes more and more chaotic. When the amount of noise is large enough, the original clean signal almost becomes pure noise. During the noise addition process, a neural network such as Unet is used to predict the amount of noise added. When denoising, a noisy signal is input, and the neural network gradually denoises it, and finally a signal with interference removed is obtained.

[0034] Specifically, the diffusion model DDPM regards the noise addition process as a Markov process. Given x0, it models the process from x1 to x t The joint distribution of can model the whole process. Combined with the Markov property, we can get:

[0035]

[0036] Among them, x1 is the value of x at time 1, T is the final time, x0 is the value of x at the initial time, that is, the input of the diffusion model DDPM, t is a certain time, q is the true distribution, and x t is the value of x at time t, x t-1 is the value of x at time t-1, N is a normal distribution, β t is a series of fixed values ​​that vary linearly or cosine-variantly, and I is unit variance.

[0037] It can be seen that each state node in the Markov chain conforms to the Gaussian distribution and is only related to the previous state, that is, the expression of the current distribution is determined by the previous observation value. If the current distribution is known, the reparameterization technique can be used to obtain the value of the current sample:

[0038]

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

[0040]

[0041] Among them, x T is the value of x at time T, which is pure Gaussian noise.

[0042] The denoising process is to transform the pure Gaussian noise x TIterate back to x0 step by step, and use the Bayesian formula for the true reverse transfer distribution function:

[0043]

[0044] Since the initial distribution is unknown, it is impossible to directly obtain the marginal distribution from the joint distribution. However, the transition probability distribution of each state is known, and with the properties of the Markov chain, we can get:

[0045]

[0046] After reparameterization, we can get:

[0047]

[0048] in,

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

[0050] Among them, KL is KL divergence, p θ is the predicted distribution.

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

[0052]

[0053] Among them, ∈ θ is a function of the prediction noise fitted to the model.

[0054] In this way, it is transformed into an optimization problem of the noise predictor. In the forward noise addition process of the diffusion model DDPM, the coefficient ratio of noise to signal will change with the change of t. In order to satisfy the defined disturbance rate relationship, we fix the time t and set the coefficients of noise and signal to In this way, the ratio of the two coefficients is 1. In this way, to set different perturbation rates, we only need to change the coefficient of noise∈.

[0055] For the signal perturbation rate, we define it 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 explore the impact of three types of noise disturbances on the accuracy of electromagnetic signal recognition, namely Gaussian noise, Rayleigh noise and periodic noise. Periodic noise is a type of noise that recurs with a specific frequency and amplitude. It usually appears in the form of a sine or cosine waveform and has a clear periodic structure.

[0058] Gaussian noise and Rayleigh noise are defined as follows:

[0059]

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

[0061] Inspired by the ViT idea, we also split the signal into several patches. However, unlike images, signals are one-dimensional data, so we perform 1D Patch Embedding on them. Similarly, we add position encoding to them to preserve the temporal information in the signal.

[0062] In addition, we found that when the Transformer encoder is directly used to extract signal features, the classification accuracy is average and the training is unstable. Therefore, we modified the Transformer encoder. Specifically, we added residual convolution between Attention and MLP to enable the model to better focus on local fine-grained features without losing global context information, while making the training more stable. We named the modified module Sigformer Block module, as shown in Figure 3 shown.

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

[0064] X=Attention(X)+X

[0065] X=Conv(X)+X

[0066] X=MLP(X)+X

[0067] Among them, Attention(·) represents the self-attention operation, Conv(·) represents the convolution operation, and MLP(·) represents the multi-layer perceptron.

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

[0069] The process of training the diffusion model DDPM is essentially to reduce the distribution difference between the predicted noise and the real noise. We use the mean square error loss function MSE to optimize this process:

[0070]

[0071] Among them, ∈ is the real noise, To predict the noise.

[0072] 1D Patch Embedding module: divides the signal sequence after interference removal into blocks, converts it into a form that can be received by TransformerEncoder, adds position coding, and preserves the temporal relationship of the original sequence; Sigformer Block module: is used to extract features from the signal after interference removal and obtain information of different modulation categories; Classifier module: is used to receive the extracted deep features and output the predicted probability of each modulation category. The 1D Patch Embedding module, SigformerBlock module, and classifier module constitute the classification model Sigformer, and the cross entropy loss function is used to optimize Sigformer:

[0073]

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

[0075] We conducted experiments on data with a signal-to-noise ratio of 18 dB from the RML2016.10a and RML2016.10b datasets to explore the recognition ability of Sigformer on cleaner signals.

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

[0077]

[0078] Table 2 Recognition accuracy of RML2016.10b

[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 the accuracy of Sigformer on both datasets is better than that of other algorithms.

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

[0082] Table 3 Diffusion Sigformer recognition accuracy of disturbed signals (Gaussian noise)

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

[0084] Table 4. Recognition accuracy of diffuse Sigformer for disturbed signals (Rayleigh noise)

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

[0086] Table 5. Recognition accuracy of diffuse Sigformer for disturbed signals (periodic noise)

[0087] Disturbance rate 1 2 3 4 5 6 7 8 9 10 Before denoising 0.648 0.614 0.588 0.582 0.582 0.582 0.575 0.566 0.563 0.562 After denoising 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 diffuse Sigformer for the disturbed signal under three forms of noise disturbance. Before denoising, it refers to directly testing the recognition accuracy of Sigformer with the data with added noise disturbance, which is used to study the dynamic relationship between the model accuracy and the disturbance rate when the disturbance rate changes from 1-10 times. After denoising, it means that the diffusion model DDPM is used for denoising before the model is used for recognition, and the denoised data is input into the model for recognition. From the data in the table, it can be seen that as the disturbance rate changes from 1-10, the more noise disturbance is added, the recognition accuracy of the model gradually decreases. However, using the diffusion model DDPM for denoising first and then recognition, the recognition accuracy of the model is always maintained at a high level, which reflects the effectiveness of using the diffusion model for interference removal. However, the interference removal effect of the diffuse Sigformer on the three noise disturbances is slightly different. Its interference removal effect on Gaussian disturbance is better, and its interference removal effect on Rayleigh disturbance is poor. This is related to the characteristics of the three signals themselves, and further exploration is needed.

[0089] In general, electromagnetic signal recognition has important significance and wide application value in the fields of communication, radar detection, spectrum management, information security and military applications. In wireless communication, electromagnetic signal recognition is the basis for ensuring the correct demodulation of signals and extracting useful information. By identifying the modulation mode, parameters and other characteristics of the signal, the performance of the communication system can be optimized and the reliability and stability of communication can be improved. In the military field, electromagnetic signal recognition plays a vital role in space electromagnetic confrontation and obtaining information rights. By accurately identifying the modulation type of the enemy's communication signal, effective electromagnetic interference and intelligence interception can be implemented to provide strong support for military operations. In the radar system, electromagnetic signal recognition technology is used to identify the modulation type of the radar echo signal, so as to obtain the distance, speed, direction and other information of the target object. This is of great significance for radar detection and target tracking. In addition, electromagnetic signal recognition helps spectrum management agencies to effectively monitor and manage spectrum resources. By identifying the spectrum occupancy of different signals, the waste and abuse of spectrum resources can be discovered and dealt with in a timely manner to ensure the rational and efficient use of spectrum resources. With the continuous development and improvement of electromagnetic signal recognition technology, it will play an important role in more fields and bring more 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 added noise. Specifically, Unet, as a powerful convolutional neural network architecture, is often used as the backbone of the model. In the diffusion model DDPM, the task of Unet is not to directly predict the denoised signal, but to predict the amount of noise that needs to be removed in order to restore the current noisy signal to the original clean signal.

[0091] Example 2

[0092] The present invention also provides a method for identifying interfered electromagnetic signals based on diffusion Sigformer, including: inputting the disturbed signal into a diffusion model DDPM for de-perturbation, outputting the de-interfered signal as an input to a 1D Patch Embedding module, the 1D Patch Embedding module converting the de-interfered signal into a token, adding a position code and a class token for classification, and inputting the token into a Sigformer Block module for feature extraction, the Sigformer Block module using the final output as an input to a classifier module, and outputting the predicted probability of each modulation category.

[0093] Diffusion model DDPM, input: disturbed signal Disturbed Signal, processing: input the signal into the trained diffusion model DDPM model, gradually remove the disturbance in the signal through the denoising process of the model, output: Denoised Signal after interference.

[0094] 1D Patch Embedding module, input: signal after interference removal, processing: split the signal into a series of small segments (patches), each patch can be regarded as a small signal fragment; convert each patch into a fixed-size embedding vector (token), usually implemented by linear transformation or convolution layer; add position encoding (such as sinusoidal position encoding) to preserve the order information between patches; add a class token for classification, which is before or after all patch tokens for subsequent classification tasks. Output: token sequence containing position encoding and class token.

[0095] Sigformer Block module, input: token sequence from 1D Patch Embedding, processing: Sigformer Block is a Transformer-based architecture, but may contain some specific improvements for signal processing (although "Sigformer" is not a standard term, it is assumed here that it is similar to the Transformer variant for signal processing). The Block may contain a self-attention mechanism to capture the dependencies between tokens, and may also contain a feed-forward network and layer normalization. By stacking multiple Sigformer Blocks, high-level features of the signal can be gradually extracted. Output: feature-extracted token sequence.

[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 to the self-attention result and inputs it into a one-dimensional convolution, adds the one-dimensional convolution result to the one-dimensional convolution input, and inputs it into a multi-layer perceptron, and adds the result of the multi-layer perceptron to the input of the multi-layer perceptron as the final output of the SigformerBlock module.

[0097] Classifier module, input: the final output of the Sigformer Block module, receives the extracted deep features, and outputs the predicted probability of each modulation category. In this embodiment, the interfered signal recognition system uses a linear layer as a classifier module.

[0098] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

Claims

1. A system for identifying interfered electromagnetic signals based on diffusion Sigformer, characterized in that: include: Diffusion model DDPM: denoise the interfered electromagnetic signal to obtain the de-interferenced signal; Classification model Sigformer: used to identify the modulation mode of the signal after interference removal.

2. The interfered electromagnetic signal identification system based on diffusion Sigformer according to claim 1 is characterized in that: The classification model Sigformer includes: 1D Patch Embedding module: used to divide the signal sequence after interference removal into blocks, convert it into a form that can be received by TransformerEncoder, add position coding, and preserve the temporal relationship of the original sequence; Sigformer Block module: used to extract features from the signal after interference removal to obtain information of different modulation categories; Classifier module: used to receive the extracted deep features and output the predicted probability of each modulation category.

3. The interference electromagnetic signal identification system based on diffusion Sigformer according to claim 2 is characterized in that: The 1D Patch Embedding module includes a patch one-dimensional convolution and a learnable position encoding vector; The patch one-dimensional convolution divides the de-interferenced signal into a series of patches; The learnable position encoding vector is used to retain the position information between each patch.

4. The interference electromagnetic signal identification system based on diffusion Sigformer according to claim 2 is characterized in that: The classifier module is composed of linear layers, takes the final output of the Sigformer Block module as input, and outputs the predicted probability of each modulation category.

5. The interference electromagnetic signal identification system based on diffusion Sigformer according to claim 1 is characterized in that: The diffusion model DDPM is trained using a mean square error loss function, and the classification model Sigformer is trained using a cross entropy loss function.

6. The interference electromagnetic signal identification system based on diffusion Sigformer according to claim 5 is characterized in that: The mean square error loss function expression is: Where N is the number of samples, ∈ is the real noise, To predict the noise.

7. The interference electromagnetic signal identification system based on diffusion Sigformer according to claim 5 is characterized in that: The cross entropy loss function expression is: Among them, N is the number of samples, K is the number of categories, and y ik is the true label of the i-th sample belonging to the k-th category, is the predicted probability that the i-th sample belongs to the K-th category.

8. The interference electromagnetic signal identification system based on diffusion Sigformer according to claim 7 is characterized in that: The backbone of the diffusion model DDPM is Unet, which is used to predict the amount of added noise.

9. A method for identifying interfered electromagnetic signals based on diffusion Sigformer, characterized in that: The interfered electromagnetic signal identification system based on the diffusion Sigformer according to any one of claims 1 to 8 comprises the following steps: The disturbed signal is input into the diffusion model DDPM for de-perturbation, and the de-interferenced signal is output as the input of the 1D Patch Embedding module. The 1D Patch Embedding module converts the de-interferenced signal into a token, adds a position code and a class token for classification, and inputs it into 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.

10. The method for identifying interfered electromagnetic signals based on diffusion Sigformer according to claim 9, 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 to the self-attention result and inputs it into a one-dimensional convolution, adds the one-dimensional convolution result to the one-dimensional convolution input, inputs it into a multi-layer perceptron, and adds the result of the multi-layer perceptron to the input of the multi-layer perceptron as the final output of the Sigformer Block module.

Images