Low-orbit satellite communication signal identification method based on time-frequency Transform model

Through the deep learning method based on the time-frequency Transformer model, the time-frequency characteristics of the OFDM signal are extracted and the attention mechanism is used to solve the problem of low accuracy of OFDM signal recognition in low-orbit satellite communication, and efficient signal recognition and adaptability are achieved. Improves the flexibility and adaptability of the communication system.

CN119025964AActive Publication Date: 2024-11-26JILIN UNIVERSITY

Patent Information

Application Number
CN202411121730.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-15
Publication Date
2024-11-26
Estimated Expiration
2044-08-15

AI Technical Summary

Technical Problem

The OFDM signal modulation method in the prior art medium and low-orbit satellite communication has low accuracy and low efficiency, making it difficult to adapt to complex communication environments and variable channel conditions.

Method used

Using a deep learning method based on the time-frequency Transformer model, a new signal recognizer is designed through time-frequency feature extraction and attention mechanism, which can enrich the time-domain features without changing the frequency-domain characteristics and capture it through the attention mechanism. The global and local characteristics of the signal enable accurate identification of signals of various modulation types.

Benefits of technology

It improves the accuracy and efficiency of signal recognition, enhances the robustness and generalization capabilities of the model, and can adapt to complex communication environments and diversified modulation methods, achieving 93% recognition accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119025964A_ABST
    Figure CN119025964A_ABST
Patent Text Reader

Abstract

The invention discloses a low-orbit satellite communication signal identification method based on a time-frequency Transform model, relates to a low-orbit satellite communication signal identification method, and belongs to the field of wireless communication, satellite communication and deep learning. The invention aims to solve the problems of low identification accuracy and low efficiency of the modulation mode of the modulated OFDM signal in the existing method. The method comprises the following steps: obtaining modulated OFDM signal data with a label, and taking the modulated OFDM signal data as a training set; constructing a deep learning model, taking the training set as the input of the deep learning model, taking the classification result as the output of the deep learning model, training the deep learning model until convergence, and obtaining a trained deep learning model; and collecting modulated OFDM signal data to be detected, inputting the modulated OFDM signal data to be detected into the trained deep learning model, and outputting a classification result by the trained deep learning model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a method for identifying low-orbit satellite communication signals, and belongs to the fields of wireless communication, satellite communication, and deep learning. Background Art

[0002] Low Earth Orbit (LEO) communication satellites play an increasingly important role in the field of modern communications. With the rapid growth of Internet demand and the acceleration of the global networking trend, LEO communication satellites have become an important part of building a global communication network with their advantages of low latency, high bandwidth and global coverage. Compared with ground communication infrastructure, LEO satellites can provide efficient and reliable communication services in areas with complex terrain or insufficient infrastructure, which shows a wide range of application prospects in the fields of Internet of Things (IoT), emergency communications, remote sensing and scientific research. The complexity and high dynamics of LEO satellite communication systems bring challenges to signal monitoring and analysis. Due to their high-speed operation and frequent switching, LEO satellites require high-precision tracking and real-time signal monitoring. The transmission process of signals is affected by many factors, including atmospheric attenuation, multipath effects, frequency drift and interference. Therefore, how to accurately monitor and analyze the communication signals of LEO satellites to ensure the stability of communication links and the reliability of data transmission has become an important topic of current research.

[0003] Orthogonal frequency division multiplexing (OFDM) and signal modulation identification technology are the core technologies in low-orbit satellite communications. As a multi-carrier modulation technology, OFDM effectively combats multipath fading and frequency selective fading by dividing the data stream into multiple parallel low-speed subcarriers for transmission. Its anti-interference ability and high spectrum efficiency in the high-speed mobile environment of low-orbit satellites make it an ideal choice for satellite communications. However, OFDM signals are susceptible to multipath effects and frequency drift during transmission, and how to monitor and optimize OFDM signals in real time has become a research focus. Time domain and frequency domain features are important features of signal data, and each plays an important role in signal identification, providing rich information for comprehensive understanding and accurate identification of signals. First, time domain information describes the change of the signal over time, including the waveform, amplitude, phase, etc. of the signal. Learning the time domain features can help the model identify the time characteristics of the signal. The frequency domain information shows the characteristics of the signal at different frequencies and describes the distribution of the signal at frequency, including the spectrum distribution and frequency components of the signal. We can use frequency domain features to identify the frequency characteristics of the signal, which plays an important role in identifying different types of signals. Therefore, enriching the time domain and frequency domain features of data is meaningful for the recognition of OFDM signals.

[0004] Since satellite communication systems may use a variety of modulation modes (such as PSK, QAM, ASK, etc.), accurately identifying the modulation mode of the received signal is a prerequisite for correct demodulation and further processing. We can simulate signals of various modulation types and use automatic modulation recognition technology to identify signals of different modulation types through spectrum analysis and feature extraction to improve the flexibility and adaptability of the communication system. This is particularly important for dealing with complex satellite communication environments and changing channel conditions. Modulation mode recognition refers to identifying the modulation mode of an unknown signal received under limited or no prior information. Due to the rapid development of modern communication technology and the continuous improvement of signal processing complexity, deep learning has broad application prospects in the field of channel modulation recognition. Signal modulation is a key link in wireless communication systems. It involves signal coding, modulation and demodulation processes, which has an important impact on communication quality and efficiency. Traditional signal modulation mode recognition mainly relies on manual feature extraction by experts, which often has problems such as low accuracy and low efficiency. Summary of the invention

[0005] The purpose of the present invention is to solve the problem of low accuracy and low efficiency in identifying the modulation mode of modulated OFDM signals in existing methods, and to propose a low-orbit satellite communication signal recognition method based on a time-frequency Transformer model.

[0006] The specific process of the low-orbit satellite communication signal recognition method based on the time-frequency Transformer model is as follows:

[0007] Step 1: Obtain OFDM signal data, modulate the OFDM signal data, obtain modulated OFDM signal data with labels, and use the modulated OFDM signal data as a training set;

[0008] Step 2: Build a deep learning model, use the training set as the input of the deep learning model, and the classification result as the output of the deep learning model. Train the deep learning model until convergence to obtain a trained deep learning model. The specific process is as follows:

[0009] Step 21: Split the modulated OFDM signal data in the time domain and frequency domain to obtain the time domain matrix E t and the frequency domain matrix E f ;

[0010] Step 2: Time domain matrix E t and the frequency domain matrix E f Add position codes respectively to get matrix Z t and the matrix Z f ;

[0011] Step 23: Transform the matrix Z tInput to M layers of Transformer encoder in sequence, and the last layer of Transformer encoder outputs matrix R t ;

[0012] The matrix Z f Input to Transformerencoder, Transformerencoder output matrix R f ;

[0013] Step 24: Matrix R t and R f Perform feature fusion to obtain a set of features R containing time-frequency characteristics;

[0014] Step 25: Input the feature R containing time-frequency characteristics into flatten, Dropout, and MLP in sequence. The MLP output feature is input into the Dense layer. The Dense layer outputs the category probability to obtain the classification result.

[0015] Step 26: Repeat steps 21 to 25 until convergence to obtain a trained deep learning model;

[0016] Step 3: Collect the modulated OFDM signal data to be tested, input the modulated OFDM signal data to be tested into the trained deep learning model, and the trained deep learning model outputs the classification result.

[0017] The beneficial effects of the present invention are:

[0018] With the continuous advancement of communication technology, signal modulation methods have become more and more complex and diverse, and traditional recognition methods can no longer meet actual needs. As a powerful machine learning technology, deep learning has the ability to automatically learn and extract features, can handle complex nonlinear relationships, and extract useful information from large amounts of data. Therefore, the application of deep learning in the field of signal modulation has gradually attracted attention. By applying deep learning algorithms to signal modulation recognition, the feature representation of the signal can be automatically learned and accurate classification of different modulation methods can be achieved, reducing a large number of errors caused by human participation. This method not only improves the accuracy and efficiency of recognition, but also can adapt to various complex communication environments and modulation changes. The Transformer model has outperformed convolutional neural networks in natural language processing (NLP) and computer vision (CV), so we consider applying it to the field of signal modulation recognition. Compared with the way that convolutional neural networks use fixed-size filters to capture local information, the attention mechanism can dynamically adjust and integrate local receptive fields and global information at different levels, so that the model can better focus on task-related areas, thereby improving the discriminability of features. In addition, the attention mechanism can adaptively adjust the attention distribution according to the characteristics of the input data, so that the network can better adapt to different input situations. This adaptive property enables the network to better process image data with complex structures and diversity. By weighting the importance of different regions, the attention mechanism can improve the network's grasp and utilization of key information, thereby improving the performance of the network.

[0019] This paper aims to explore the monitoring and analysis technology of low-orbit communication satellite signals, focusing on the processing of OFDM signals in satellite communications and automatic modulation recognition technology. By designing the transmission characteristics of OFDM signals and designing efficient signal identifiers based on the attention mechanism, the reliability and performance of signal recognition are improved. This will not only help meet the growing global communication needs, but also provide a solid theoretical foundation and technical support for future communication technology innovations.

[0020] The purpose of this invention is to design an OFDM signal identifier based on the time-frequency Transformer model for the monitoring and analysis tasks of low-orbit satellite communication signals. By simulating the OFDM signal of low-orbit satellite communication, the simulated signal is trained using the Transformer network to capture the time-frequency characteristics of the data. Finally, the modulation recognition task of various modulation type signals is realized. It belongs to the fields of wireless communication, satellite communication, and deep learning.

[0021] The present invention designs a new data construction form by combining the time domain and frequency domain characteristics of OFDM signal data, so as to enrich the time domain characteristics of the data without affecting the frequency domain characteristics of the data. In addition, we build an OFDM data generator to complete the production of the data set.

[0022] The present invention utilizes the excellent ability of the attention mechanism to capture the global and local features of the data, and designs a Transformer model based on the time-frequency attention mechanism based on the attention mechanism for the time domain and frequency domain features of the signal data, so as to realize the synchronous extraction of the time domain and frequency domain features of the data by the model, so that the model can more comprehensively capture the complex characteristics of the signal, and improve the robustness, generalization ability and recognition performance of the model.

[0023] In terms of data construction, we splice signal data from different periods on the time domain axis to capture the temporal changes of the signal. This construction has the following advantages for the model to learn data:

[0024] 1. Without changing the frequency domain characteristics of each OFDM signal itself, the time domain characteristics of the data can be displayed sequentially on the time axis, thereby enriching the time domain characteristics of the data.

[0025] 2. The spliced ​​data contains the frequency domain information of the signal at different time periods, and the model can more effectively capture the changes of the signal over time. This helps the model identify the dynamic characteristics of the modulated signal, such as the change pattern of frequency and the change trend of amplitude.

[0026] 3. This splicing method can improve the flexibility of feature representation. The spliced ​​data forms a two-dimensional time-frequency graph, which not only provides information in the time dimension, but also combines the features of the frequency domain. The attention mechanism can operate flexibly in these two dimensions. By simultaneously analyzing the time and frequency dimensions, the time-varying characteristics of the signal can be more comprehensively captured, thereby generating richer and more detailed feature representations. This multi-dimensional feature representation is very helpful for the recognition of complex signals.

[0027] 4. Through splicing, the changes in signal characteristics can be more clearly displayed. For example, the intensity of certain frequency components may change in different time periods, and these changes will be more significant in the spliced ​​spectrum, making it easier to capture these time-varying characteristics.

[0028] In short, splicing the frequency domain signal data of different time periods on the time axis can obtain more complete and detailed spectrum information, help capture the time-varying characteristics of the signal, extract a variety of features, and significantly improve the effect and accuracy of signal modulation recognition.

[0029] Complete the design of the time-frequency Transformer signal recognition model, complete the extraction of time-frequency features, and finally realize the recognition of signal data

[0030] In terms of signal identifiers, we improved the time-frequency characteristics of signal data based on the attention mechanism and made an OFDM signal identifier based on the time-frequency Transformer model. The attention mechanism can pay attention to both the global information and local details of the data. For example, it can consider the spectral characteristics of the entire signal when identifying the modulation type, while also noting the changes in frequency components at certain key moments. This integration of global and local information helps improve the accuracy of recognition. In addition, the attention mechanism automatically pays attention to time periods with significant frequency changes and those frequency components with identification characteristics, so as to accurately identify the modulation type of the signal and improve the accuracy and robustness of signal recognition. For each attention head, the input data is linearly transformed to generate query, key, and value vectors, which are used to calculate the attention weight. The query, key, and value vectors play an important role in capturing the characteristics of frequency domain data. Through the dot product operation of the query vector and the key vector, the attention mechanism calculates the attention weight, which represents the importance of the frequency domain data of a certain time period in the overall recognition task. This matching process helps identify time periods and frequency components that contribute significantly to the modulation type.

[0031] We design the model network based on the time-frequency characteristics of the signal data. After the spliced ​​data is input into the model, in order to better extract the time domain and frequency domain characteristics of the data, we will design the model network from the time domain and frequency domain to process the data in parallel. The specific steps are as follows:

[0032] 1. First, in the design of the Patch encoder, we perform two parallel Patch segmentation methods on the same set of data. In the branch that focuses on extracting frequency domain features, each Patch contains data from all time periods but only a few frequency domains. This method retains the changes in specific frequency components at all time points, and each patch represents the changes in a specific frequency across the entire time axis; in the branch that focuses on extracting time domain features of data, there are only data from the same time period inside the Patch. This method retains multiple sets of frequency component information within the same time period, and each patch represents multiple sets of frequency domain distributions for a specific time period. The two groups of branches segmented in this way have different emphases on time domain characteristics and frequency domain characteristics, which helps the attention mechanism to focus on learning the two features of the data. The segmented patch groups will be mapped to a high-dimensional space to capture the basic characteristics of the signal.

[0033] 2. Secondly, we will add position encoding (Position embedding) to the segmented time domain patch group and frequency domain patch group respectively. The position encoding provides position information, so that the model can perceive the position and order of elements in the input sequence. For the time domain patch group, position encoding is added. Each patch contains the frequency domain data and time position information in a specific time period, which helps the model understand the change of the signal over time. For the frequency domain patch group, position encoding is added. Each patch contains the time series data of a specific frequency component and its frequency position information, which helps the model understand the change of the signal at different frequencies.

[0034] 3. We send the two sets of time and frequency data with position encoding to the Transformer encoder module for training, and use the multi-head attention mechanism to parallel process and multi-head segment the input patch group to effectively learn the characteristics of the data. In the self-attention layer, the correlation between each position in the input signal and all other positions is calculated, which can highlight the important features of the input signal while suppressing the unimportant parts and generating weighted feature representations. The output of each input position is the result of the weighted average of all input positions, where the weight is determined by the similarity (correlation) of the input signal. The self-attention layer deeply extracts the features of the input signal, recognizes complex signal patterns, and enhances the model's recognition ability for complex signals through nonlinear activation functions.

[0035] By learning the two groups of time and frequency patches, we can obtain two sets of features learned by the multi-head attention mechanism for the time domain characteristics and frequency domain characteristics of the data respectively.

[0036] 4. We fuse the two sets of features obtained, and perform operations including linear transformation, activation function, normalization, etc. on the fused features, and finally complete the design of the model to realize the recognition of the modulation mode; the recognition accuracy and efficiency of the modulation mode of the modulated OFDM signal are improved.

[0037] 5. We use the model to train the produced data set. Through training, our model can achieve an accuracy of 93% for this kind of time-frequency data. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0039] Specific implementation method 1: Combination Figure 1 This embodiment is described. The specific process of the low-orbit satellite communication signal recognition method based on the time-frequency Transformer model in this embodiment is as follows:

[0040] Step 1: Obtain OFDM signal data, modulate the OFDM signal data (modulation mode such as PSK, QAM, ASK, etc.), obtain the modulated OFDM signal data with labels, and use the modulated OFDM signal data as a training set;

[0041] Step 2: Build a deep learning model, use the training set as the input of the deep learning model, and the classification result as the output of the deep learning model. Train the deep learning model until convergence to obtain a trained deep learning model. The specific process is as follows:

[0042] Step 21: Split the modulated OFDM signal data in the time domain and frequency domain to obtain the time domain matrix E t and the frequency domain matrix E f ;

[0043] Step 2: Time domain matrix E t and the frequency domain matrix E f Add position codes respectively to get matrix Z t and the matrix Z f ;

[0044] Step 23: Transform the matrix Z t Input to M layers of Transformer encoder in sequence, and the last layer of Transformer encoder outputs matrix R t ;

[0045] The matrix Z f Input to Transformerencoder, Transformerencoder output matrix R f ;

[0046] Step 24: Matrix R t and R f Perform feature fusion to obtain a set of features R containing time-frequency characteristics;

[0047] Step 25: Input the feature R containing time-frequency characteristics into flatten, Dropout, and MLP in sequence. The MLP output feature is input into the Dense layer. The Dense layer outputs the category probability to obtain the classification result.

[0048] Step 26: Repeat steps 21 to 25 until convergence to obtain a trained deep learning model;

[0049] Step 3: Collect the modulated OFDM signal data to be tested, input the modulated OFDM signal data to be tested into the trained deep learning model, and the trained deep learning model outputs the classification result.

[0050] The LayerNormalization layer performs normalization once for each signal block, performing N independent normalizations in total. The normalization formula is as follows:

[0051]

[0052] Where γ and β are the scale parameter and bias parameter of the LN layer, respectively, both of which are learnable parameters. μ and σ2 are the mean and variance of the data in each signal block, and ε is a 10 -6 A constant is used to prevent the variance from becoming zero.

[0053] We will get the Z with position encoding added t and Z f The input is sent to the Transformer encoder, which mainly includes the multi-head self-attention mechanism and the feedforward neural network (MLP) module, as well as the layer normalization and skip connection between them. Before applying the multi-head self-attention mechanism, the layer normalization will be applied to Z t and Z f Normalization is performed to speed up training and stabilize the model training process. The processed data will then be input into the multi-head attention module>The input data is linearly transformed and mapped to generate query (Q), key (K) and value (V) vectors, and the query, key and value vectors are divided into multiple heads to parallelly calculate the attention output of different subspaces.

[0054] Q=XW Q

[0055] K=XW K

[0056] V=XW V

[0057] For each attention head, the dot product between the query matrix Q and the key matrix K is calculated, and the attention weight is calculated by the softmax function. The calculation for each head is as follows:

[0058]

[0059] Among them, QK T is the dot product of the query matrix and the key matrix. is a scaling factor to prevent the dot product value from being too large and causing the gradient to disappear. The softmax function converts the similarity score into a probability distribution.

[0060] Then, the outputs of all heads are concatenated and a linear transformation is performed to obtain the final multi-head self-attention output:

[0061] MultiHead(Q,K,V)=Concat(head1,...,head h )W O

[0062] Where W O ∈R D×D is the output learnable weight matrix. The attention weight matrix represents the degree of attention of each element in the input sequence to all other elements. Through weighted summation, the attention mechanism generates new feature representations that integrate important time and frequency domain information and improve the expressiveness of the model. Through multiple attention heads, the model can process input data from different angles and extract rich feature representations in the time and frequency domains, greatly enhancing the model's ability to extract and express signal features. After that, the output of the multi-head self-attention will be layer normalized and skipped to ensure smooth information flow and reduce the problem of gradient disappearance. Finally, two sets of features are obtained, which we use R and t and R f To express.

[0063] Dropout is a regularization technique used to prevent overfitting. By randomly setting a certain proportion (here 50%) of neuron outputs to zero, it can enhance the generalization ability of the model, especially when there is less training data or the model is more complex. MLP is a multi-layer perceptron that uses two linear and GELU activation functions to further process and extract high-level features. We input the extracted features into the Dense layer, set the number of its output units to the number of categories, map the features extracted by MLP to the final classification results, and finally calculate the final category probability and output the classification results.

[0064] Specific implementation method 2: The difference between this implementation method and specific implementation method 1 is that in step 21, the modulated OFDM signal data is divided in the time domain and the frequency domain respectively to obtain the time domain matrix E t and the frequency domain matrix E f ;

[0065] The specific process is:

[0066] Step 2: Segment the acquired modulated OFDM signal data in the time domain to obtain the time domain matrix E t ; The specific process is:

[0067] Step 2: Divide the modulated OFDM signal data X in the time domain, and each small block after division is x t , The number of small blocks is N;

[0068] Where X∈R H×W×C , H is the height of the modulated OFDM signal data X, W is the length of the modulated OFDM signal data X, C is the number of channels of the modulated OFDM signal data X, H t is the height of the modulated OFDM signal data X corresponding to each patch, W t The length of the modulated OFDM signal data X corresponding to each block patch;

[0069] Step 2: Flatten each patch into an H t ×W t ×C vector;

[0070] Step 2-13: The expanded vectors corresponding to each patch are subjected to a linear transformation (fully connected layer), and the vectors output by the linear transformation form the time domain matrix E t ; Map the vector to a fixed-dimensional feature space, so that we get the time domain embedding matrix E t ;

[0071] Step 2: Split the acquired modulated OFDM signal data in the frequency domain to obtain the frequency domain matrix E. f ; The specific process is:

[0072] Step 2121: Split the acquired modulated OFDM signal data in the frequency domain, and each small block after segmentation is x f , The number of small blocks is N;

[0073] Where X∈R H×W×C , H is the height of the modulated OFDM signal data X, W is the length of the modulated OFDM signal data X, C is the number of channels of the modulated OFDM signal data X, H f is the height of the modulated OFDM signal data X corresponding to each patch, W f The length of the modulated OFDM signal data X corresponding to each block patch;

[0074] Step 2: Flatten each patch into an H f ×W f ×C vector;

[0075] Step 2123: The expanded vectors corresponding to each patch are subjected to a linear transformation (fully connected layer), and the vectors output by the linear transformation form the frequency domain matrix E f ; Map the vector to a fixed-dimensional feature space, so that we get the time domain embedding matrix Ef ;

[0076] In the PatchEmbedding layer, our data will be split into two parts: time domain and frequency domain. The number of small blocks in the two parts is the same, which is N.

[0077] The other steps and parameters are the same as those in the first embodiment.

[0078] Specific implementation method three: This implementation method is different from specific implementation methods one or two in that the expression for the number of small blocks N divided in step 2-11-1 is:

[0079]

[0080] The other steps and parameters are the same as those in the first or second embodiment.

[0081] Specific implementation method 4: This implementation method is different from the specific implementation methods 1 to 3 in that the expression of the number of small blocks N divided in step 2121 is:

[0082]

[0083] The other steps and parameters are the same as those in Specific Embodiments 1 to 3.

[0084] Specific implementation method 5: This implementation method is different from the specific implementation methods 1 to 4 in that the time domain matrix E in step 22 is t and the frequency domain matrix E f Add position codes respectively to get matrix Z t and the matrix Z f ; The specific process is:

[0085] PositionalEmbedding

[0086] In order to allow the model to capture the order and spatial position information of patches, we will add position encoding to the time domain patch group and the frequency domain patch group respectively. In this model design, we choose to use learnable position encoding, which allows the model to automatically adjust the position encoding according to the data during training, so that it may better adapt to signal recognition tasks.

[0087] Step 221: Initialize a learnable position encoding matrix P for the time domain matrix t ;

[0088] The time domain matrix E t With the position encoding matrix P t Add together to get the new input matrix Z t ;

[0089] Step 222: Initialize a learnable position encoding matrix P for the frequency domain matrix f ;

[0090] The time domain matrix E f With the position encoding matrix P f Add together to get the new input matrix Z f ;

[0091] This ensures that the model can understand and utilize sequential information when processing input data, while allowing the position encoding vector to be optimized and adjusted according to the data during training so that the model can subsequently extract the time domain features and frequency domain features of the data respectively.

[0092] The other steps and parameters are the same as those in Specific Embodiments 1 to 4.

[0093] Specific implementation method 6: This implementation method is different from the specific implementation methods 1 to 5 in that in step 221, a learnable position encoding matrix P for the time domain matrix is ​​initialized. t ;

[0094] The time domain matrix E t With the position encoding matrix P t Add together to get the new input matrix Z t ;

[0095] The specific process is:

[0096] Initialize a learnable position encoding matrix P for the time domain matrix t ;

[0097] Matrix P t The size is N×D, where N is the sequence length (num_patches) and D is the embedding dimension; each element of this matrix is ​​a model parameter and can be updated during training;

[0098] The time domain matrix E t With the position encoding matrix P t Add together to get the new input matrix Z t ;

[0099] Z t =E t +P t .

[0100] The other steps and parameters are the same as those in Specific Implementation Methods 1 to 5.

[0101] Specific implementation method 7: This implementation method is different from the specific implementation methods 1 to 6 in that in step 222, a learnable position coding matrix P for the frequency domain matrix is ​​initialized.f ;

[0102] The time domain matrix E f With the position encoding matrix P f Add together to get the new input matrix Z f ;

[0103] The specific process is:

[0104] Initialize a learnable position encoding matrix P for the frequency domain matrix f ;

[0105] Matrix P f The size is N×D, where N is the sequence length (num_patches) and D is the embedding dimension; each element of this matrix is ​​a model parameter and can be updated during training;

[0106] The time domain matrix E f With the position encoding matrix P f Add together to get the new input matrix Z f ;

[0107] Z f =E f +P f .

[0108] The other steps and parameters are the same as those in Specific Embodiments 1 to 6.

[0109] Specific implementation eight: This implementation differs from the specific implementations one to seven in that the matrix R in step 24 is t and R f Perform feature fusion to obtain a set of features R containing time-frequency characteristics; the expression is:

[0110] R = Concat(R t ,R f )

[0111] Among them, Concat(,) means concatenation; R means features including time-frequency characteristics.

[0112] The other steps and parameters are the same as those in Specific Embodiments 1 to 6.

[0113] Specific implementation method 9: This implementation method is different from any one of specific implementation methods 1 to 8 in that, in the M-layer Transformer encoder, 3≤M≤7.

[0114] The other steps and parameters are the same as those in Specific Embodiments 1 to 8.

[0115] Specific implementation method ten: This implementation method is different from any one of specific implementation methods one to nine in that, in the M-layer Transformer encoder, M=4.

[0116] The other steps and parameters are the same as those in Specific Embodiments 1 to 9.

[0117] The present invention may also have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. A low-orbit satellite communication signal recognition method based on a time-frequency Transformer model, characterized by: The specific process of the method is: Step 1: Obtain OFDM signal data, modulate the OFDM signal data, obtain modulated OFDM signal data with labels, and use the modulated OFDM signal data as a training set; Step 2: Build a deep learning model, use the training set as the input of the deep learning model, and the classification result as the output of the deep learning model. Train the deep learning model until convergence to obtain a trained deep learning model. The specific process is as follows: Step 21: Split the modulated OFDM signal data in the time domain and frequency domain respectively to obtain the time domain matrix E t and the frequency domain matrix E f ; Step 2: Time domain matrix E t and the frequency domain matrix E f Add position codes respectively to get matrix Z t and the matrix Z f ; Step 23: Transform the matrix Z t Input to M layers of Transformer encoder in sequence, and the last layer of Transformer encoder outputs matrix R t ; The matrix Z f Input to Transformerencoder, Transformerencoder output matrix R f ; Step 24: Matrix R t and R f Perform feature fusion to obtain a set of features R containing time-frequency characteristics; Step 25: Input the feature R containing time-frequency characteristics into flatten, Dropout, and MLP in sequence. The MLP output feature is input into the Dense layer. The Dense layer outputs the category probability to obtain the classification result. Step 26: Repeat steps 21 to 25 until convergence to obtain a trained deep learning model; Step 3: Collect the modulated OFDM signal data to be tested, input the modulated OFDM signal data to be tested into the trained deep learning model, and the trained deep learning model outputs the classification result.

2. The low-orbit satellite communication signal recognition method based on the time-frequency Transformer model according to claim 1 is characterized in that: In step 21, the modulated OFDM signal data is divided in the time domain and the frequency domain respectively to obtain the time domain matrix E t and the frequency domain matrix E f ; The specific process is: Step 2: Segment the acquired modulated OFDM signal data in the time domain to obtain the time domain matrix E t ; The specific process is: Step 2: Segment the modulated OFDM signal data X in the time domain. After segmentation, each small block is The number of small blocks is N; Where X∈R H×W×C , H is the height of the modulated OFDM signal data X, W is the length of the modulated OFDM signal data X, C is the number of channels of the modulated OFDM signal data X, H t is the height of the modulated OFDM signal data X corresponding to each block, W t is the length of the modulated OFDM signal data X corresponding to each block; Step 2: Flatten each block into an H t ×W t ×C vector; Step 2-13: The expanded vectors corresponding to each block are subjected to a linear transformation, and the vectors output by the linear transformation form the time domain matrix E t ; Step 2: Split the acquired modulated OFDM signal data in the frequency domain to obtain the frequency domain matrix E. f ; The specific process is: Step 2121: Split the acquired modulated OFDM signal data in the frequency domain, and each small block after segmentation is x f , The number of small blocks is N; Where X∈R H×W×C , H is the height of the modulated OFDM signal data X, W is the length of the modulated OFDM signal data X, C is the number of channels of the modulated OFDM signal data X, H f is the height of the modulated OFDM signal data X corresponding to each block, W f is the length of the modulated OFDM signal data X corresponding to each block; Step 2: Flatten each block into an H f ×W f ×C vector; Step 2123: The expanded vectors corresponding to each block are subjected to a linear transformation, and the vectors output by the linear transformation form the frequency domain matrix E f .

3. The low-orbit satellite communication signal recognition method based on the time-frequency Transformer model according to claim 2 is characterized in that: The expression of the number of small blocks N divided in step 2-1-1 is:

4. The low-orbit satellite communication signal recognition method based on the time-frequency Transformer model according to claim 3 is characterized in that: The expression for the number of small blocks N divided in step 2121 is:

5. The method for identifying low-orbit satellite communication signals based on a time-frequency Transformer model according to claim 4, characterized in that: In step 22, the time domain matrix E t and the frequency domain matrix E f Add position codes respectively to get matrix Z t and the matrix Z f ; The specific process is: Step 221: Initialize a learnable position encoding matrix P for the time domain matrix t ; The time domain matrix E t With the position encoding matrix P t Add together to get the new input matrix Z t ; Step 222: Initialize a learnable position encoding matrix P for the frequency domain matrix f ; The time domain matrix E f With the position encoding matrix P f Add together to get the new input matrix Z f .

6. The method for identifying low-orbit satellite communication signals based on a time-frequency Transformer model according to claim 5, characterized in that: In step 221, a learnable position encoding matrix P for the time domain matrix is ​​initialized. t ; The time domain matrix E t With the position encoding matrix P t Add together to get the new input matrix Z t ; The specific process is: Initialize a learnable position encoding matrix P for the time domain matrix t ; Matrix P t The size is N×D, where N is the sequence length and D is the embedding dimension; The time domain matrix E t With the position encoding matrix P t Add together to get the new input matrix Z t ; With t =E t +P t 。 7. The method for identifying low-orbit satellite communication signals based on a time-frequency Transformer model according to claim 6, characterized in that: In step 222, a learnable position encoding matrix P for the frequency domain matrix is ​​initialized. f ; The time domain matrix E f With the position encoding matrix P f Add together to get the new input matrix Z f ; The specific process is: Initialize a learnable position encoding matrix P for the frequency domain matrix f ; Matrix P f The size is N×D, where N is the sequence length and D is the embedding dimension; The time domain matrix E f With the position encoding matrix P f Add together to get the new input matrix Z f ; With f =E f +P f 。 8. The method for identifying low-orbit satellite communication signals based on a time-frequency Transformer model according to claim 7, characterized in that: In step 24, the matrix R t and R f Perform feature fusion to obtain a set of features R containing time-frequency characteristics; the expression is: R=Concat(R t ,R f ) Among them, Concat(,) means concatenation; R means features including time-frequency characteristics.

9. The method for identifying low-orbit satellite communication signals based on a time-frequency Transformer model according to claim 8, characterized in that: In the M-layer Transformer encoder, 3≤M≤7.

10. The method for identifying low-orbit satellite communication signals based on a time-frequency Transformer model according to claim 9, characterized in that: In the M-layer Transformer encoder, M=4.

Citation Information

Patent Citations

  • Conventional communication signal modulation identification method based on deep learning

    CN116720115A

  • Transform-based OFDM (Orthogonal Frequency Division Multiplexing) receiving method

    CN117176534A

  • OFDM based orbital angular momentum system and method

    EP3066809A1

  • Automatic modulation classification method based on deep learning network fusion

    US20240112037A1

  • Composite interference signal recognition method and system

    WO2023092923A1

Cited By

  • Specific system OFDM signal detection and identification method based on multi-domain feature fusion

    CN121056285A

  • A specific type of OFDM signal detection and identification method based on multi-domain feature fusion

    CN121056285B