An electromagnetic signal recognition method and system based on U-Mamba

Abstract

The application relates to the technical field of electromagnetic signal modulation identification, in particular to an electromagnetic signal identification method and system based on U-Mamba. The method comprises the following steps: acquiring electromagnetic signal data to be identified; inputting the electromagnetic signal data to be identified into a constructed identification model; performing modulation identification on the disturbed signal in the electromagnetic signal data to obtain a final classification result; wherein the identification model is constructed based on a U-Mamba algorithm, and the identification model is a model trained by using an electromagnetic signal data set; the electromagnetic signal data set comprises disturbed signals and undisturbed signals; the method can not only accurately identify the disturbed signals, but also effectively prevent part of defects in the existing network.

Description

Technical Field

[0001] This invention relates to the field of electromagnetic signal modulation recognition technology, and specifically to an electromagnetic signal recognition method and system based on U-Mamba. Background Technology

[0002] As an intermediate step between signal reception and demodulation, electromagnetic signal modulation identification (EMI) is essentially a pattern recognition task, aiming to automatically identify and distinguish the modulation scheme of received unknown electromagnetic signals without requiring prior information. This has significant practical implications for cognitive radio and plays a crucial role in both military and civilian fields such as electronic reconnaissance, interference identification, spectrum monitoring, and signal reconstruction. For example, in the civilian sector, EPI is commonly used for electromagnetic spectrum monitoring and management, identifying unauthorized users occupying and interfering with frequency bands to ensure communication security. In the military sector, identifying electromagnetic signal categories can provide relevant information, assess the threat level of enemy jamming equipment, and assist friendly reconnaissance equipment. However, with the rapid development of radio communication technology, the number of electromagnetic devices is growing exponentially, and the electromagnetic environment is becoming increasingly complex and variable with the increase in the number and modulation schemes of radio signals. Numerous external factors make signals more susceptible to various noise interferences during transmission, and the interfered signals often differ significantly from the original signals, posing a significant challenge to signal identification technology. Therefore, effective interference signal identification technology has received widespread attention. How to accurately identify the modulation category of interfered signals in such a dense, complex, and diverse electromagnetic environment is a problem that urgently needs to be solved.

[0003] With the rapid development of deep learning methods, Convolutional Neural Networks (CNNs) have been widely used due to their powerful feature extraction capabilities, providing new solutions beyond traditional recognition methods. CNNs can automatically learn discriminative features from input signals for recognition and classification with high accuracy. However, while CNNs have strong local feature extraction capabilities, they are limited in extracting global features and long-range dependencies due to the limited scope of convolution operations. Therefore, transformer-based models have shown potential in computer vision tasks and have been used in the field of signal modulation recognition. However, the use of attention mechanisms typically incurs significant computational costs, especially when processing long sequences of data. How to extract discriminative features from complex electromagnetic signals more efficiently, reduce computational memory usage, and achieve more effective and accurate modulation recognition remains a challenge.

[0004] On the other hand, the electromagnetic environment is complex and variable, and modulated signals are subject to interference during transmission. Various noise interferences can cause significant differences between the modulated signal and the original signal. Models trained only on clean signals perform well in ideal channel environments, but they cannot handle situations where the distribution of test data and training data differs. Directly using these models for classifying interfered signals will lead to a gradual decrease in recognition accuracy. Furthermore, most existing methods for handling interfered signals train the classifier and denoising network separately. The denoising model is trained first, and then the denoised result is input into the classifier. This two-stage processing approach may inevitably remove some semantic features that are important to the recognition task during the denoising process, resulting in poor classification performance. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an electromagnetic signal identification method and system based on U-Mamba, which solves the problems of inaccurate identification and poor classification ability of existing methods for identifying and classifying interfered electromagnetic signals.

[0006] The objective of this invention is achieved through the following technical solutions:

[0007] In a first aspect, the present invention provides an electromagnetic signal identification method based on U-Mamba, comprising:

[0008] The electromagnetic signal data to be identified is acquired, and the electromagnetic signal data to be identified is input into the constructed identification model. The interference signal in the electromagnetic signal data is modulated and identified to obtain the final classification result.

[0009] The recognition model is constructed based on the U-Mamba algorithm. The recognition model is a model trained using an electromagnetic signal dataset, which includes both interfered and undisturbed signals. The U-Mamba algorithm is used to construct the model based on the network structure of Unet using the Mamba algorithm.

[0010] As a further improvement of the present invention, the recognition model includes an encoding module, a multi-scale learning module, a decoding module and a classification module connected in sequence;

[0011] The encoding module is used to extract the features of the input interference signal and perform downsampling to obtain deep encoded features;

[0012] The multi-scale learning module is used to process the encoded deep feature map to obtain decoded input features at different scales;

[0013] The decoding module is used to subtract the deep encoded features at symmetrical positions from the decoded input feature maps at different scales, further extract local-global features from the subtracted features, and perform upsampling operations to obtain the decoded output features.

[0014] The classification module is used to subtract the features of the decoded output feature from the features of the input interference signal to obtain the denoised features, and to perform modulation recognition on the denoised features to obtain the final classification result.

[0015] As a further improvement of the present invention, the encoding module includes a channel attention block, a Mamba block, a splicing block, and a fusion downsampling block connected in sequence:

[0016] The channel attention module is used to perform channel attention processing on the features of the input interfered signal to obtain a channel attention feature map;

[0017] The Mamba block is used to extract features from the input distorted signal to obtain a coded global feature map;

[0018] The splicing block is used to splice the channel attention feature map and the encoded global feature map to obtain the encoded spliced ​​feature map;

[0019] The fusion downsampling block is used to fuse the coded spliced ​​feature map and the input interference signal features to obtain a coded fused feature map. Then, a downsampling operation is performed on the coded fused feature map to obtain a reduced-size coded deep feature.

[0020] As a further improvement of the present invention, the channel attention block is specifically used for:

[0021] The features of the interfered signal are processed by adaptive average pooling to obtain the first processing result;

[0022] The features of the interfered signal are subjected to adaptive max pooling to obtain a second processing result;

[0023] The first processing result and the second processing result are concatenated, and then the concatenated output is processed sequentially by one-dimensional convolution, ReLU operation, one-dimensional convolution and Sigmoid function to obtain the third processing result;

[0024] The third processing result is multiplied by the features of the interfered signal to obtain the channel attention feature map.

[0025] As a further improvement of the present invention, a convolution block is also included between the splicing block and the fusion downsampling block. The convolution block is used to perform convolution processing on the encoded splicing feature map so that the number of channels of the encoded splicing feature map is adjusted to be the same as the input interference signal features.

[0026] As a further improvement of the present invention, the specific steps of the multi-scale learning module include:

[0027] The encoded deep features are processed by one-dimensional dilated convolution, and the dilation factor is set to a fixed value to obtain feature maps of different scales.

[0028] The feature maps of different scales are concatenated to obtain a concatenated multi-scale feature map. After the concatenated multi-scale feature map is processed by a one-dimensional convolution operation, a multi-scale learning feature map is finally obtained.

[0029] As a further improvement of the present invention, the decoding module includes a subtraction block, a similarity calculation block, and an upsampling block;

[0030] The subtraction block subtracts the multi-scale learning feature map output by the multi-scale learning module from the deep encoding feature map output by the previous layer encoding module to obtain the subtracted feature map.

[0031] The similarity calculation block extracts local and global decoding features from the subtracted feature map by passing it through two one-dimensional convolutions and a Mamba block, respectively. The local decoding features are multiplied by the transpose of the global high-level features to obtain the decoding similarity map. The decoding similarity map is further processed by a linear layer. The output of the linear layer is added to the original decoding similarity map as a residual. The result is multiplied by the output of the two one-dimensional convolutions to obtain the enhanced global decoding features.

[0032] The upsampling block processes the enhanced global decoding features sequentially through one-dimensional deconvolution, ReLU operation, and batch normalization layer to obtain the deep decoding feature map.

[0033] As a further improvement of the present invention, the classification module includes a one-dimensional patch embedding block, a Mamba block, and a classification header connected in sequence.

[0034] The one-dimensional patch embedding block is used to transform the input feature map of the classification module into a set of embedding forms and add position encoding to obtain the embedded feature map;

[0035] The Mamba block is used to extract global information from the embedded feature map to obtain classification features;

[0036] The classification head is used to perform layer normalization on the classification features. The last two dimensions are flattened by the flatten operation and then input into the linear layer to map the number of channels to the same number of categories, thus obtaining the final classification result.

[0037] As a further improvement of the present invention, the loss function in the recognition model is a denoising-classification joint loss function, the expression of which is:

[0038] loss Dc =lossrec +loss lae +loss cla

[0039] In the formula, loss rec Represents the reconstruction loss, loss lae The loss represents the log-absolute error loss. cla This represents the cross-entropy loss.

[0040] Secondly, the present invention also provides an electromagnetic signal identification system based on U-Mamba, for implementing a U-Mamba electromagnetic signal identification method, comprising:

[0041] The signal recognition module is used to acquire electromagnetic signal data to be identified, input the electromagnetic signal data to be identified into the constructed recognition model, modulate and identify the interference signals in the electromagnetic signal data, and obtain the final classification result.

[0042] The recognition model construction module is used to construct the recognition model based on the U-Mamba algorithm. The recognition model is a model trained using an electromagnetic signal dataset, which includes both interfered and undisturbed signals. The U-Mamba algorithm is used to construct the model based on the network structure of Unet using the Mamba algorithm.

[0043] The beneficial effects of this invention are as follows: This invention provides an electromagnetic signal identification method based on U-Mamba, which constructs an identification model using the U-Mamba algorithm to achieve the identification and classification of interfered electromagnetic signals. Compared to existing signal identification models, the U-Mamba algorithm used in this invention has efficient data processing capabilities. When processing interfered electromagnetic signals, it can complete complex calculations or analysis tasks in a shorter time. Furthermore, it can adaptively learn during the training data process, automatically adjusting its parameters or strategies based on new data or experience to optimize its performance and thus improve the accuracy of identification.

[0044] Furthermore, the recognition model in this invention, namely the U-Mamba algorithm model, includes an encoding module, a multi-scale learning module, a decoding module, and a classification module. The encoding module performs feature encoding processing on the input interference signal. By applying a channel attention mechanism, the recognition model automatically learns the importance of feature channels and adjusts the feature representation according to the importance of each channel, thereby enhancing the model's focus on key information of the input features. Global feature extraction of the input signal data is implemented based on Mamba blocks. Compared to CNNs, Mamba blocks have better global feature extraction capabilities and require less computational memory than transformers, resulting in more efficient inference. The attention mechanism is concatenated and fused with the output of the Mamba blocks and added to the input of the encoding block to prevent information loss during feature extraction. Finally, a downsampling operation is performed to reduce the feature size, thereby further reducing computational complexity and effectively preventing overfitting due to limited training data.

[0045] Furthermore, the multi-scale learning module in the recognition model can extract multi-scale features by utilizing the deep encoded features output by the encoding block. Considering that upsampling or pooling layers may lead to the loss of internal feature structure, thus failing to improve recognition accuracy, the multi-scale learning module uses dilated convolutions with different dilation factors on the deep encoded features of the encoding block to expand the receptive field. In the decoding module, multi-scale contextual information is supplemented, thereby improving the network performance.

[0046] Furthermore, the decoding module combines reconstruction loss and log-absolute error loss to denoise the interfered signal. The model predicts the noise distribution, and the final denoised features are obtained by subtracting the predicted noise features from the input features of the interfered signal. By combining reconstruction loss, log-absolute error loss, and classification loss, denoising and classification are performed jointly, enabling the denoising part to process the interfered signal in a way that is beneficial for classification. The classification module can better extract the discriminative information of the denoised features, thereby improving the recognition accuracy. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a schematic diagram of the structure of a recognition model provided in an embodiment of the present invention;

[0049] Figure 2This is a schematic diagram of a method for identifying interfered electromagnetic signals based on U-Mamba, provided in an embodiment of the present invention.

[0050] Figure 3 This is a schematic diagram of the structure of an encoding module provided in an embodiment of the present invention;

[0051] Figure 4 This is a schematic diagram of the structure of a decoding module provided in an embodiment of the present invention;

[0052] Figure 5 This is a schematic diagram of the structure of a classification module provided in an embodiment of the present invention;

[0053] Figure 6 This is a comparative curve diagram showing the modulation and identification of the interfered signal using different methods;

[0054] In the diagram, 101 is the encoding module, 102 is the multi-scale learning module, 103 is the decoding module, and 104 is the classification module. Detailed Implementation

[0055] To make the objectives and technical solutions of this invention clearer and easier to understand, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. The specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0056] Terminology Explanation:

[0057] U-Mamba: A novel selectively structured state-space model. This network is built using the Mamba algorithm based on the Unet network structure and performs exceptionally well in long sequence modeling tasks. Mamba is suitable for processing long sequence data, such as Natural Language Processing (NLP) and other tasks involving sequence modeling. Its core lies in introducing a selection mechanism that efficiently determines whether the relevance information of each label in the sequence is worth propagating or discarding. Through global receptive field and dynamic weighting, it alleviates the modeling constraints of convolutional neural networks and provides advanced modeling capabilities similar to Transformers.

[0058] RML: Radio Signal Machine Learning;

[0059] ReLU: Activation function

[0060] Encoder: Encoding module;

[0061] Multi-L: Multi-scale learning module;

[0062] Decoder: Decoding module;

[0063] CBlock: Classification module;

[0064] The flatten operation converts a multidimensional array (usually two-dimensional or three-dimensional) into a one-dimensional array.

[0065] The present invention provides a method and system for electromagnetic signal identification based on U-Mamba, the method mainly comprising:

[0066] The electromagnetic signal data to be identified is acquired, and the electromagnetic signal data to be identified is input into the constructed identification model. The interference signal in the electromagnetic signal data is modulated and identified to obtain the final classification result.

[0067] The recognition model is constructed based on the U-Mamba algorithm. The recognition model is a model trained using an electromagnetic signal dataset, which includes both interfered and undisturbed signals. The U-Mamba algorithm is used to construct the model based on the network structure of Unet using the Mamba algorithm.

[0068] The recognition model consists of an encoding module, a multi-scale learning module, a decoding module, and a classification module connected in sequence.

[0069] The encoding module extracts features from the input interference signal and performs downsampling to obtain deep encoded features. The encoding module consists of a channel attention block, a Mamba block, a concatenation block, and a fusion downsampling block connected in sequence. The channel attention block performs channel attention processing on the input interference signal features to obtain a channel attention feature map. The Mamba block extracts features from the input interference signal features to obtain a global encoded feature map. The concatenation block concatenates the channel attention feature map and the global encoded feature map to obtain a concatenated encoded feature map. The fusion downsampling block fuses the concatenated encoded feature map and the input interference signal features to obtain a fused encoded feature map, which is then downsampled to obtain a reduced-size deep encoded feature map.

[0070] Specifically, the channel attention block is used to: obtain a first processing result by performing adaptive average pooling on the disturbed signal features; obtain a second processing result by performing adaptive max pooling on the disturbed signal features; concatenate the first processing result and the second processing result, and then process the concatenated output sequentially by performing one-dimensional convolution, ReLU operation, one-dimensional convolution and Sigmoid function to obtain a third processing result; and perform a dot product between the third processing result and the disturbed signal features to obtain the channel attention feature map.

[0071] Between the splicing block and the fusion downsampling block, there is also a convolution block, which is used to perform convolution processing on the encoded splicing feature map so that the number of channels of the encoded splicing feature map is the same as the input interference signal features.

[0072] The multi-scale learning module processes the encoded deep feature maps to obtain decoded input features at different scales. The specific steps of the multi-scale learning module include:

[0073] The deep encoded features are processed by one-dimensional dilated convolution, with the dilation factor set to a fixed value, to obtain feature maps of different scales. The feature maps of different scales are then concatenated to obtain a concatenated multi-scale feature map. After the concatenated multi-scale feature map is processed by one-dimensional convolution, a multi-scale learning feature map is finally obtained.

[0074] The decoding module is used to subtract the encoded deep features at symmetrical positions from the decoded input feature maps at different scales. It then further extracts local-global features from the subtracted features and performs an upsampling operation to obtain the decoded output features. The decoding module includes a subtraction block, a similarity calculation block, and an upsampling block.

[0075] Specifically, the subtraction block subtracts the multi-scale learning feature map output by the multi-scale learning module from the deep encoding feature map output by the previous layer encoding module to obtain a subtracted feature map. The similarity calculation block extracts the decoding local features and decoding global features from the subtracted feature map by passing it through two one-dimensional convolutions and a Mamba block, respectively. The decoding local features are multiplied by the transpose of the decoding global high-level features to obtain a decoding similarity map. The decoding similarity map is further processed by a linear layer. The output of the linear layer is added to the original decoding similarity map as a residual. The result is multiplied by the output of the two one-dimensional convolutions to obtain the enhanced decoding global features. The upsampling block processes the enhanced decoding global features sequentially through a one-dimensional deconvolution, ReLU operation, and batch normalization layer to obtain the decoding deep feature map.

[0076] The classification module is used to subtract the features of the decoded output signal from the features of the input interference signal to obtain denoised features. These denoised features are then modulated and identified to obtain the final classification result. The classification module includes a one-dimensional patch embedding block, a Mamba block, and a classification head, connected in sequence.

[0077] Specifically, the one-dimensional patch embedding block is used to transform the input feature map of the classification module into a set of embedding forms and add positional encoding to obtain the embedded feature map; the Mamba block is used to extract global information from the embedded feature map to obtain classification features; the classification head is used to perform layer normalization processing on the classification features, flatten the last two dimensions through the flatten operation, and then input it into the linear layer to map the number of channels to the same number of categories to obtain the final classification result.

[0078] Furthermore, the loss function in the recognition model is a denoising-classification joint loss function, the expression of which is:

[0079] loss DC =loss rec +loss lae +loss cla

[0080] In the formula, loss rec Represents the reconstruction loss, loss lae The loss represents the log-absolute error loss. cla This represents the cross-entropy loss.

[0081] The aforementioned recognition model not only avoids overfitting caused by insufficient training data, but also boasts high recognition accuracy and classification ability.

[0082] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings and specific embodiments. The described embodiments are only some embodiments of the present invention, and not all embodiments.

[0083] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations are intended to cover non-exclusive inclusion, such that an article or apparatus comprising a list of elements includes not only those elements but also other elements not expressly listed. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or apparatus that includes said element.

[0084] Example 1

[0085] like Figure 1-6 As shown, this embodiment provides an electromagnetic signal identification method based on U-Mamba, and the specific implementation method is as follows.

[0086] A recognition model is constructed based on the U-Mamba algorithm and trained using an electromagnetic signal dataset, which includes both distorted and undisturbed signals.

[0087] Please see Figure 1 and Figure 2 The recognition model includes an encoding module 101, a multi-scale learning module 102, a decoding module 103, and a classification module 104. The encoding module 101 is used to input the features of the interfered signal, extract the features of the input signal through Mamba blocks and attention mechanisms, and use downsampling to reduce the feature size to capture deep encoded features. The multi-scale learning module 102 uses the features of the encoding module to further obtain features at different scales and supplements multi-scale information during the decoding process. The decoding module 103 is used to further extract global features from the input features of the decoding module through Mamba blocks, calculate similarity by combining convolution outputs to fuse local information, and increase the feature size through upsampling. The classification module 104 is used to extract information from the decoded output features to obtain the final classification result.

[0088] In this embodiment, the original electromagnetic signal data input to the recognition model during training is obtained from the RML series electromagnetic signal dataset, with a uniform size of 2×128, where 2 represents the number of channels and 128 represents the signal length. The generation of the interfered signal involves randomly selecting a position within the signal length of all 18dB signal data in the electromagnetic signal dataset, and then taking 20 consecutive data points from that position to simulate adding corresponding noise perturbations, including Gaussian noise, Rayleigh noise, and periodic noise. Therefore, the electromagnetic signal after noise perturbation is the interfered signal, while the undisturbed signal (or clean signal) is the original data in the dataset. This results in interfered signal-clean signal data pairs used to train the recognition model, with the classification category corresponding to the original category in the dataset.

[0089] Denoising is essentially the recovery and reconstruction of interfered signal data. Based on the premise of the additive noise model, the interfered signal can be represented as the sum of the clean signal and the noise signal. The task of denoising is to remove the noise from the interfered signal and obtain the clean signal data. Directly learning the mapping from the interfered signal to the clean signal is difficult. This means that the model not only needs to learn the complex features of the interfered signal and identify the noise, but also needs to recover the clean signal based on the identified noise distribution. The task that the model needs to complete is more complex and requires more computational resources. Therefore, in the denoising process, the model only needs to learn the noise distribution in the interfered signal, which usually has certain statistical characteristics. This allows the model to converge faster and improves model performance.

[0090] like Figure 3As shown, the encoding module 101 in this embodiment includes a channel attention block, a Mamba block, a concatenation block, and a fusion downsampling block. The channel attention block performs channel attention calculation on the input features of the encoding module to obtain a channel attention map; the Mamba block extracts global features from the input features of the encoding module to obtain a global encoding feature map; the concatenation block concatenates the channel attention feature map and the global encoding feature map to obtain a concatenated encoding feature map; and the fusion downsampling block fuses the concatenated encoding feature map and the input features of the encoding module to obtain a fused encoding feature map, and performs downsampling on the fused encoding feature map to obtain a deep encoding feature map.

[0091] It should be noted that the input distorted signal is first subjected to a 3×3 one-dimensional convolution to increase the number of channels to 8. The channel attention block obtains channel weights through adaptive max pooling and average pooling, concatenates the two obtained channel weights, and increases the number of channels to 16. Then, a 1×1 one-dimensional convolution is used to reduce the number of channels to one-eighth of the input number of channels, i.e., 1. After calculation, another 1×1 one-dimensional convolution is used to increase the dimensionality to the input number of 8 channels. Finally, the contribution of each channel is calculated using the sigmoid function. The input of the encoding module is multiplied by the channel weights, and the importance of each channel is combined to enhance the model's focus on useful information.

[0092] Similarly, the number of channels in the encoded global feature map extracted by the Mamba block remains 8. After fusion by the concatenation block, the number of channels in the encoded concatenated feature map becomes 16. Then, a 1×1 one-dimensional convolution is set to readjust the number of channels to 8, and finally, a fusion downsampling block is applied.

[0093] Furthermore, such as Figure 3 As shown, the fusion downsampling block in this embodiment is specifically used for: adding the encoded spliced ​​feature map and the input features of the encoding module to perform feature fusion to obtain a first processing result; performing downsampling processing on the first processing result, first performing 3×3 one-dimensional convolution to increase the dimensionality, making the number of channels twice the original, then performing ReLU operation and max pooling to halve the size of the feature map, and finally performing batch normalization layer to obtain the encoded deep feature map output by the encoding module.

[0094] The processing procedure of this encoding module can also be described as follows:

[0095] f1 = CA(Y)

[0096] f2 = Mamba(Y)

[0097] f = DS(Conv(Concat(f1,f2)) + Y)

[0098] Where Y represents the input features of the encoding module, CA(Y) represents the channel attention operation on the input features of the encoding module, f1 represents the channel attention feature map; Mamba(Y) represents the global feature extraction operation on the input features of the encoding module, f2 represents the encoding global feature map; Concat represents the concatenation operation, Conv represents the 1×1 one-dimensional convolution, DS represents the downsampling operation, and f represents the encoding deep feature map output by the encoding module.

[0099] The encoding module processes the input feature map using channel attention blocks, Mamba blocks, concatenation blocks, and fusion downsampling blocks. Channel attention blocks enhance the interaction between channels within the encoding module, allowing the model to pay more attention to and learn from channels that contribute significantly. They are also fused with global features extracted by Mamba blocks, resulting in richer spatial global features and improving the model's feature representation capabilities. Furthermore, the downsampling operation reduces computational memory by shrinking the feature map size.

[0100] In this embodiment, the multi-scale learning module 102 performs a 1×1 one-dimensional dilated convolution operation on the deep features output by the encoding module, setting the dilation factors to fixed values ​​of 1, 2, and 4 to obtain feature maps {f1, f2, f3} at different scales; the feature maps {f1, f2, f3} at different scales are concatenated to obtain a concatenated multi-scale feature map, and then the concatenated multi-scale feature map is processed by a 1×1 one-dimensional convolution operation to finally obtain a multi-scale learning feature map F.

[0101] Specifically, in this embodiment, a set of feature maps obtained through three dilated convolutions is represented as {f1, f2, f3}. Experience shows that fusing features of different resolutions is better than using only single-scale features. This embodiment uses a simple concatenation operation to complete the fusion operation. The multi-scale learning module 102 can also be represented as:

[0102] f1 = Dilate_Conv1(f)

[0103] f2 = Dilate_Conv2(f)

[0104] f3 = Dilate_Conv4(f)

[0105] F = Conv(Concat(f1,f2,f3))

[0106] Where f represents the output of the encoding module, Dilate_Conv(·) represents dilated convolution operation, the subscript of Dilate_Conv represents the dilation factor, Concat represents the concatenation operation, and Conv(·) represents 1×1 one-dimensional convolution.

[0107] Subsequently, the decoding module 103 further extracts global and local features from the input features, and obtains a deep feature map through upsampling. For example... Figure 4 As shown. The decoding module 103 is specifically used to: process the multi-scale learning feature map F output by the multi-scale learning module 102 to obtain the input features, further extract global features from the input features through Mamba blocks, calculate similarity by combining convolution output to fuse local information, and finally increase the feature size through upsampling operation to increase detail information.

[0108] Specifically, the decoding module 103 subtracts the multi-scale learning feature map F output by the multi-scale learning module 102 from the output of the previous layer decoding module to obtain the input features of the next layer decoding block; it further extracts global features from the input features through the Mamba block to obtain the decoded global features G, further enriching the global information in the features; it simultaneously enhances the local information in the features by applying two 1×1 one-dimensional convolutions to the input features to obtain the decoded local features L; it multiplies the decoded local features L with the transpose of the decoded global features G to obtain a similarity map; the output of the similarity map after linear layer processing is added to the original similarity map by residual addition to obtain the final fused similarity map; and the final fused similarity map is multiplied with the decoded local features L to obtain the enhanced decoded deep feature map.

[0109] The decoding module 103 improves the model's performance and representational capability by fusing global and local features. The fusion method is based on similarity graph computation, considering the information interaction between global and local features. We believe that local features contain more local characteristics, while global features mainly focus on the overall attributes of the features; fusing the two can better extract important information from the signal features.

[0110] The calculation formula for decoding module 103 is as follows:

[0111] L = Conv(Conv(FO))

[0112] G = Mamba(FO)

[0113] sim0 = L × G T

[0114] sim = linear(sim0) + sim0

[0115] DF = sim × L

[0116] Where F represents the multi-scale learned feature map, O represents the output of the previous decoding module, Conv() represents a 1×1 one-dimensional convolution, Mamba represents a Mamba block, L represents the decoded local feature, G represents the decoded global feature, sim represents the similarity, and DF represents the decoded deep feature map output by the decoding module.

[0117] Before the input classification module, the deep feature map of the last decoding module is subtracted from the features of the input distorted signal to obtain a denoised feature map. The classification module 104 uses one-dimensional patch embedding operation, Mamba block and classification head to extract features from the denoised feature map for modulation recognition. After the features are flattened, the number of channels is mapped to the number of categories by a linear layer to obtain the final classification result.

[0118] Furthermore, the goal of this embodiment is to perform denoising and classification tasks jointly. Therefore, a joint denoising-classification loss function is introduced. This loss function helps the model optimize for both denoising and classification tasks simultaneously, thus reducing the possibility that denoising results may be detrimental to classification due to separate denoising and classification. This allows the model to consider the impact on classification while denoising, improving recognition accuracy. Denoising and classification complement each other, contributing to the improvement of model performance.

[0119] Specifically, considering that the reconstruction loss uniformly processes noise disturbances across all channels of the interfered signal, but suffers from uneven noise distribution, the log-absolute error loss is used in conjunction with the reconstruction loss to address the denoising task unevenly. In this embodiment, the denoising-classification joint loss function minimizes the reconstruction loss, log-absolute error loss, and classification loss (cross-entropy loss).

[0120] The expression for the overall denoising-classification joint loss function is as follows:

[0121] loss DC =Loss rec +Loss lae +Loss cla

[0122] Where, loss rec Represents the reconstruction loss, loss lae The loss represents the log-absolute error loss. cla This represents the cross-entropy loss.

[0123] Specifically, the expression for the reconstruction loss is:

[0124]

[0125] Among them, y i Let f(x) represent the i-th clean signal data. i ) represents the i-th input feature of the classification module, and N represents the number of signal data.

[0126] The formula for the log-absolute error loss is:

[0127]

[0128] Among them, y n,ij f(x) represents the j-th pixel value of the i-th channel of the n-th clean signal data. n,ij ) represents the j-th pixel value of the i-th channel of the n-th input feature of the classification module, N represents the number of signal data, C represents the number of channels of the signal data, L represents the length of the signal data, log represents the logarithmic operation, and S represents the Sigmoid function.

[0129] The formula for classification loss is:

[0130]

[0131] Among them, Y ik P is a sign function; its value is 1 if the true class of the i-th sample belongs to the k-th class, and 0 otherwise. ik Let represent the probability that the i-th sample belongs to the k-th class, log represents the logarithmic operation, N represents the number of signal data, and K represents the number of classes.

[0132] The embodiment proposes four feature processing modules: an encoding module, a multi-scale learning module (Multi-L), a decoding module, and a classification module. Both the encoding and decoding modules utilize Mamba blocks instead of CNNs and transformers for feature extraction, offering powerful global feature extraction capabilities with low computational complexity and efficient training and inference. In the encoding module, the combination of channel attention blocks and Mamba blocks enhances the model's ability to extract and focus on important information within features. Downsampling operations further reduce computational memory, prevent overfitting, and promote the model's learning of deep features. In the multi-scale learning module, multi-scale feature extraction is performed on encoded deep features, enriching the model's understanding of feature information at different scales and contributing to improved recognition accuracy. In the decoding module, the calculation of similarity maps allows global and local features to mutually guide each other, enhancing the model's feature representation capabilities and improving performance. Finally, the classification module also leverages the advantages of Mamba in processing sequence data, transforming denoised features into embedded forms, which are then further extracted by Mamba blocks to obtain the classification results.

[0133] To better verify the practicality of this method, this embodiment also compares the results with those of other existing deep learning network models, such as... Figure 6 As shown, recognition models are constructed based on CGDNet, DCNNPF, sigformer, transformer, Mamba algorithms, and the algorithm in this embodiment, respectively, and the interference signals are modulated and recognized based on the recognition models. Figure 6The horizontal axis represents the perturbation rate, and the vertical axis represents the recognition accuracy. The perturbation rate represents the intensity of the noise perturbation added to the original 18dB dataset of RML2016.10a. The specific calculation process is as follows: First, calculate the root mean square amplitude of all test data used. Based on the signal-to-noise ratio, control the root mean square amplitude of the noise perturbation to be 1-10 times the root mean square amplitude of the original test data. Then, randomly select 20 consecutive points at a random location from 500 samples in the 18dB data to add perturbation to simulate the signal interfered with during transmission. Figure 6 The results reflect the trend of the recognition accuracy of the model as the disturbance intensity changes from 1 to 10 times. The results show that the interference signal recognition system of the present invention can more accurately and efficiently identify the modulation of the signal. Compared with the comparison model, the recognition curve of the present invention is above other curves.

[0134] Furthermore, Table 1 shows the experimental results of different recognition methods on the RML2016.10a dataset for modulation recognition. The evaluation metrics are various types of accuracy, and OA represents the overall accuracy. As can be seen from Table 1, the interference signal recognition system of this embodiment achieves the best results to date on the RML2016.10a dataset used for modulation recognition, and the interference signal recognition system of this embodiment did not incorporate any data augmentation methods during testing.

[0135] Table 1: Results of different models with a perturbation rate of 2 on the RML2016.10a dataset

[0136]

[0137]

[0138] As shown in Table 1, the interference signal identification method of this invention can simultaneously denoise and classify the interference signal, enabling the model to efficiently and quickly identify the modulation scheme of the interference signal. This invention achieves accurate and rapid identification while minimizing computational cost.

[0139] In summary, jammed signal identification can be applied to military and civilian fields such as electronic reconnaissance, spectrum monitoring, and signal reconstruction. By detecting and identifying electromagnetic signals, we can better understand whether there is unauthorized occupation or interference, providing technical protection for communication security. Secondly, jammed signal identification can be used in military countermeasures, analyzing enemy jamming equipment by identifying electromagnetic signals to facilitate our reconnaissance, providing practical support for military security. Overall, jammed signal identification has broad application prospects and can provide important technical protection and practical support for various fields. With the continuous development and improvement of identification technology, the application scope and accuracy of jammed signal identification models will continue to improve.

[0140] Example 2

[0141] This embodiment provides an electromagnetic signal identification system based on U-Mamba, used to implement the U-Mamba electromagnetic signal identification method described in Embodiment 1, mainly including:

[0142] The signal recognition module acquires the electromagnetic signal data to be identified, inputs it into a pre-constructed recognition model, modulates and identifies interference signals in the electromagnetic signal data, and obtains the final classification result. The recognition model in the signal recognition module includes an encoding module, a multi-scale learning module, a decoding module, and a classification module connected in sequence.

[0143] The encoding module extracts features from the input interference signal and performs downsampling to obtain deep encoding features. The multi-scale learning module processes the deep encoding feature map to obtain decoded input features at different scales. The decoding module subtracts the deep encoding features at symmetrical positions from the decoded input feature map at different scales, further extracts local-global features from the subtracted features, and performs upsampling to obtain decoded output features. The classification module subtracts the decoded output features from the input interference signal to obtain denoised features, performs modulation recognition on the denoised features, and obtains the final classification result.

[0144] The recognition model construction module is used to construct the recognition model based on the U-Mamba algorithm. The recognition model is a model trained using an electromagnetic signal dataset, which includes both interfered and undisturbed signals.

[0145] Each module, along with its corresponding principles and implementation steps, has been described in detail in Example 1 regarding a U-Mamba electromagnetic signal identification method, and will not be repeated here.

[0146] In the several embodiments provided by this invention, it should be understood that the apparatus and methods disclosed in this invention can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For example, the division of modules is merely a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0147] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or in the form of hardware plus software functional modules.

[0148] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. An electromagnetic signal identification method based on U-Mamba, characterized in that, include: The electromagnetic signal data to be identified is acquired, and the electromagnetic signal data to be identified is input into the constructed identification model. The interference signal in the electromagnetic signal data is modulated and identified to obtain the final classification result. The recognition model is constructed based on the U-Mamba algorithm. The recognition model is a model trained using an electromagnetic signal dataset, which includes both interfered and undisturbed signals. The U-Mamba algorithm is used to construct the model based on the network structure of Unet using the Mamba algorithm. The recognition model includes an encoding module, a multi-scale learning module, a decoding module, and a classification module connected in sequence. The encoding module is used to extract the features of the input interference signal and perform downsampling to obtain deep encoded features; The multi-scale learning module is used to process the encoded deep feature map to obtain decoded input features at different scales; The decoding module is used to process the multi-scale learning feature map output by the multi-scale learning module to obtain the input features. It extracts global features from the input features through Mamba blocks, calculates similarity by combining convolution output to fuse local information, and obtains the decoded deep feature map through upsampling. The classification module is used to subtract the features of the decoded output feature from the features of the input interference signal to obtain the denoised feature, and to perform modulation recognition on the denoised feature to obtain the final classification result. The decoding module includes a subtraction block, a similarity calculation block, and an upsampling block; The subtraction block subtracts the multi-scale learning feature map output by the multi-scale learning module from the output of the previous layer decoding module to obtain the subtracted feature map. The similarity calculation block extracts local and global decoding features from the subtracted feature map by passing it through two one-dimensional convolutions and a Mamba block, respectively. The local decoding features are multiplied by the transpose of the global high-level features to obtain the decoding similarity map. The decoding similarity map is further processed by a linear layer. The output of the linear layer is added to the original decoding similarity map as a residual. The result is multiplied by the output of the two one-dimensional convolutions to obtain the enhanced global decoding features. The upsampling block processes the enhanced global decoding features sequentially through one-dimensional deconvolution, ReLU operation, and batch normalization layer to obtain the deep decoding feature map.

2. The electromagnetic signal identification method based on U-Mamba according to claim 1, characterized in that, The encoding module includes a channel attention block, a Mamba block, a splicing block, and a fusion downsampling block connected in sequence: The channel attention module is used to perform channel attention processing on the features of the input interfered signal to obtain a channel attention feature map; The Mamba block is used to extract features from the input distorted signal to obtain a coded global feature map; The splicing block is used to splice the channel attention feature map and the encoded global feature map to obtain the encoded spliced ​​feature map; The fusion downsampling block is used to fuse the coded spliced ​​feature map and the input interference signal features to obtain a coded fused feature map. Then, a downsampling operation is performed on the coded fused feature map to obtain a reduced-size coded deep feature.

3. The electromagnetic signal identification method based on U-Mamba according to claim 2, characterized in that, The channel attention block is specifically used for: The features of the interfered signal are processed by adaptive average pooling to obtain the first processing result; The features of the interfered signal are subjected to adaptive max pooling to obtain a second processing result; The first processing result and the second processing result are concatenated, and then the concatenated output is processed sequentially by one-dimensional convolution, ReLU operation, one-dimensional convolution and Sigmoid function to obtain the third processing result; The third processing result is multiplied by the features of the interfered signal to obtain the channel attention feature map.

4. The electromagnetic signal identification method based on U-Mamba according to claim 2, characterized in that, The convolution block is also included between the splicing block and the fusion downsampling block. The convolution block is used to perform convolution processing on the encoded splicing feature map so that the number of channels of the encoded splicing feature map is the same as the input interference signal features.

5. The electromagnetic signal identification method based on U-Mamba according to claim 1, characterized in that, The specific steps of the multi-scale learning module include: The encoded deep features are processed by one-dimensional dilated convolution, and the dilation factor is set to a fixed value to obtain feature maps of different scales. The feature maps of different scales are concatenated to obtain a concatenated multi-scale feature map. After the concatenated multi-scale feature map is processed by a one-dimensional convolution operation, a multi-scale learning feature map is finally obtained.

6. The electromagnetic signal identification method based on U-Mamba according to claim 1, characterized in that, The classification module consists of a one-dimensional patch embedding block, a Mamba block, and a classification header connected in sequence. The one-dimensional patch embedding block is used to transform the input feature map of the classification module into a set of embedding forms and add position encoding to obtain the embedded feature map; The Mamba block is used to extract global information from the embedded feature map to obtain classification features; The classification head is used to perform layer normalization on the classification features. The last two dimensions are flattened by the flatten operation and then input into the linear layer to map the number of channels to the same number of categories, thus obtaining the final classification result.

7. The electromagnetic signal identification method based on U-Mamba according to claim 1, characterized in that, The loss function in the recognition model is a denoising-classification joint loss function, and the expression of the denoising-classification joint loss function is as follows: In the formula, Indicates the reconstruction loss. This represents the logarithmic absolute error loss. This represents the cross-entropy loss.

8. An electromagnetic signal identification system based on U-Mamba, used to implement the U-Mamba electromagnetic signal identification method according to any one of claims 1-7, characterized in that, include: The signal recognition module is used to acquire electromagnetic signal data to be identified, input the electromagnetic signal data to be identified into the constructed recognition model, modulate and identify the interference signals in the electromagnetic signal data, and obtain the final classification result. The recognition model construction module is used to construct the recognition model based on the U-Mamba algorithm. The recognition model is a model trained using an electromagnetic signal dataset, which includes both interfered and undisturbed signals. The U-Mamba algorithm is used to construct the model based on the network structure of Unet using the Mamba algorithm.

Images