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

Abstract

The 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 present application aims to solve the problems of low identification accuracy and low efficiency of the existing method for the modulation mode of the modulated OFDM signal.The process is as follows: obtaining labeled modulated OFDM signal data, 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 the trained deep learning model; collecting the modulated OFDM signal data to be tested, inputting the modulated OFDM signal data to be tested into the trained deep learning model, and outputting the classification result of the trained deep learning model.

Description

Technical Field

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

[0002] Low Earth Orbit (LEO) satellites are playing an increasingly important role in modern communications. With the rapid growth in internet demand and the accelerating trend of global networking, LEO satellites, with their advantages of low latency, high bandwidth, and global coverage, have become a crucial component in building global communication networks. Compared to terrestrial communication infrastructure, LEO satellites can provide efficient and reliable communication services in areas with complex terrain or insufficient infrastructure, demonstrating broad application prospects in fields such as the Internet of Things (IoT), emergency communications, remote sensing, and scientific research. The complexity and high dynamism of LEO satellite communication systems present challenges for signal monitoring and analysis. Due to their high-speed operation and frequent switching, LEO satellites require high-precision tracking and real-time signal monitoring. Signal transmission is affected by various factors, including atmospheric attenuation, multipath effects, frequency drift, and interference. Therefore, accurately monitoring and analyzing LEO satellite communication signals to ensure the stability of communication links and the reliability of data transmission has become an important research topic.

[0003] Orthogonal Frequency Division Multiplexing (OFDM) and signal modulation identification technology are core technologies in low-Earth orbit (LEO) satellite communication. OFDM, as a multi-carrier modulation technique, effectively combats multipath fading and frequency-selective fading by dividing the data stream into multiple parallel low-speed subcarriers for transmission. Its anti-interference capability and high spectral efficiency in the high-speed moving environment of LEO satellites make it an ideal choice for satellite communication. However, OFDM signals are susceptible to multipath effects and frequency drift during transmission, making real-time monitoring and optimization of OFDM signals a key research focus. Time-domain and frequency-domain characteristics are important features of signal data, each playing a crucial role in signal identification, providing rich information for comprehensive understanding and accurate identification of signals. First, time-domain information describes how a signal changes over time, including its waveform, amplitude, and phase. Learning time-domain features helps models identify the temporal characteristics of signals. Frequency-domain information displays the characteristics of a signal at different frequencies, describing its frequency distribution, including spectral distribution and frequency components. We can utilize frequency-domain features to identify the frequency characteristics of signals; these features play a vital role in identifying different types of signals. Therefore, the time-domain and frequency-domain characteristics of rich data are meaningful for the identification of OFDM signals.

[0004] Since satellite communication systems may employ various modulation schemes (such as PSK, QAM, ASK, etc.), accurately identifying the modulation scheme of the received signal is a prerequisite for correct demodulation and further processing. We can simulate signals of various modulation types and utilize automatic modulation identification technology through spectrum analysis and feature extraction to identify signals of different modulation types, thereby improving the flexibility and adaptability of the communication system. This is particularly important for dealing with complex satellite communication environments and variable channel conditions. Modulation scheme identification refers to identifying the modulation scheme of a received, unknown signal under conditions of limited or no prior information. Due to the rapid development of modern communication technology and the continuous increase in signal processing complexity, deep learning has broad application prospects in the field of channel modulation identification. Signal modulation is a crucial link in wireless communication systems, involving signal encoding, modulation, and demodulation processes, and has a significant impact on communication quality and efficiency. Traditional signal modulation scheme identification mainly relies on manual feature extraction by experts, a method that often suffers from low accuracy and low efficiency. Summary of the Invention

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

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

[0007] Step 1: Acquire OFDM signal data, modulate the OFDM signal data to obtain tagged modulated OFDM signal data, and use the modulated OFDM signal data as the training set.

[0008] Step 2: Construct a deep learning model. Use the training set as input and the classification results as output. Train the deep learning model until it converges to obtain a trained deep learning model. The specific process is as follows:

[0009] Step 2: First, divide the modulated OFDM signal data into the time domain and frequency domain respectively to obtain the time domain matrix E. t and frequency domain matrix E f ;

[0010] Step 22: For the time domain matrix E t and frequency domain matrix E f By adding positional encodings, we obtain matrix Z. t And matrix Z f ;

[0011] Steps two and three: Calculate matrix Z tThe inputs are sequentially fed into M layers of Transformer encoders, and the output matrix R of the last Transformer encoder layer is... t ;

[0012] matrix Z f The input is fed into a Transformer encoder, and the Transformer encoder outputs a matrix R. f ;

[0013] Step 24: For matrix R t and R f Feature fusion is performed to obtain a set of features R that include time-frequency characteristics;

[0014] Step 25: Input the features R containing time-frequency characteristics into flatten, dropout, and MLP in sequence. Input the MLP output features into the Dense layer. The Dense layer outputs the class probabilities to obtain the classification results.

[0015] Step 26: Repeat steps 21 to 25 until convergence, and obtain the 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 this invention are as follows:

[0018] With the continuous advancement of communication technology, signal modulation methods have become increasingly complex and diverse, making traditional recognition methods insufficient to meet practical needs. Deep learning, as a powerful machine learning technique, possesses the ability to automatically learn and extract features, handling complex nonlinear relationships and extracting useful information from massive 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 signals can be automatically learned, achieving accurate classification of different modulation methods and reducing numerous errors caused by human intervention. This method not only improves the accuracy and efficiency of recognition but also adapts to various complex communication environments and modulation method variations. The Transformer model has already demonstrated superior performance compared to 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 to convolutional neural networks using fixed-size filters to capture local information, the attention mechanism can dynamically adjust and integrate local receptive fields and global information at different levels, enabling the model to better focus on task-related regions, thereby improving feature discrimination. Furthermore, the attention mechanism can adaptively adjust the attention distribution according to the characteristics of the input data, allowing the network to better adapt to different input conditions. This adaptive characteristic enables the network to better handle image data with complex structures and diversity. By weighting the importance of different regions, the attention mechanism can improve the network's ability to capture and utilize key information, thereby enhancing network performance.

[0019] This invention aims to explore monitoring and analysis technologies for low-Earth orbit (LEO) communication satellite signals, focusing on OFDM signal processing and automatic modulation identification (EMI) techniques in satellite communication. By designing an efficient signal identifier based on the transmission characteristics of OFDM signals and utilizing an attention mechanism, the reliability and performance of signal identification are improved. This not only helps meet the growing global communication demands but also provides 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 a time-frequency Transformer model for monitoring and analyzing low-Earth orbit (LEO) satellite communication signals. By simulating LEO satellite communication OFDM signals, a Transformer network is used to train the simulated signals and capture the time-frequency characteristics of the data. Ultimately, this achieves modulation identification of signals with various modulation types. This invention belongs to the fields of wireless communication, satellite communication, and deep learning.

[0021] This invention combines the time-domain and frequency-domain characteristics of OFDM signal data to design a new data construction form, which enriches the time-domain features of the data without affecting the frequency-domain characteristics. In addition, we built an OFDM data generator to complete the production of the dataset.

[0022] This invention utilizes the excellent ability of the attention mechanism to capture both global and local features of data. Based on the attention mechanism, a Transformer model based on the time-frequency attention mechanism is designed to target the time-domain and frequency-domain features of signal data. This enables the model to simultaneously extract the time-domain and frequency-domain features of the data, allowing the model to capture the complex characteristics of the signal more comprehensively and improving the model's robustness, generalization ability, and recognition performance.

[0023] In terms of data construction, we stitch signal data from different time periods together on the time-domain axis to capture the changes in the signal over time. This construction has the following advantages for the model's learning of the data:

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

[0025] 2. The stitched data contains frequency domain information of the signal at different time periods, allowing the model to more effectively capture signal changes over time. This helps the model identify the dynamic characteristics of the modulated signal, such as frequency variation patterns and amplitude trends.

[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 frequency domain features. The attention mechanism can flexibly operate in both dimensions. By analyzing the time and frequency dimensions simultaneously, the time-varying characteristics of the signal can be captured more comprehensively, thereby generating richer and more detailed feature representations. This multi-dimensional feature representation is very helpful for the recognition of complex signals.

[0027] 4. By splicing, changes in signal characteristics can be displayed more clearly. For example, the intensity of certain frequency components may change over 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 summary, stitching frequency domain signal data from different time periods together on the time axis can provide a more complete and detailed spectral information, which helps to capture the time-varying characteristics of signals, extract diverse features, and significantly improve the effectiveness and accuracy of signal modulation recognition.

[0029] The design of a time-frequency Transformer signal recognition model was completed, time-frequency features were extracted, and ultimately, the recognition of signal data was achieved.

[0030] In terms of signal recognition, we improved upon the attention mechanism to address the time-frequency characteristics of signal data, creating an OFDM signal recognizer based on a time-frequency Transformer model. The attention mechanism can simultaneously consider both global information and local details of the data. For example, it can consider the spectral characteristics of the entire signal when identifying modulation types, while also noting frequency component changes at certain key moments. This integration of global and local information helps improve recognition accuracy. Furthermore, the attention mechanism automatically focuses on time periods with significant frequency changes and those frequency components with recognizable characteristics, thereby accurately identifying the modulation type of the signal and improving the accuracy and robustness of signal recognition. For each attention head, the input data is transformed linearly to generate query, key, and value vectors, which are used to calculate attention weights. Query, key, and value vectors play a crucial role in capturing the features of frequency domain data. By performing a dot product operation between the query vector and the key vector, the attention mechanism calculates the attention weights, representing the importance of frequency domain data within a certain time period in the overall recognition task. This matching process helps identify time periods and frequency components that significantly contribute to modulation type.

[0031] To address the time-frequency characteristics of the signal data, we designed a model network. After inputting the concatenated data into the model, to better extract the time-domain and frequency-domain features, we designed the model network to process the data in parallel from both the time and frequency domains. The specific steps are as follows:

[0032] 1. First, in the design of the patch encoder, we perform parallel patch segmentation on the same set of data in two ways. On the branch focusing on extracting frequency domain features, each patch contains data from all time periods, but only a few frequency domain entries. This method preserves the variation of specific frequency components across all time points, with each patch representing the variation of a specific frequency along the entire time axis. On the branch focusing on extracting time domain features, each patch contains only data from the same time period. This method preserves multiple sets of frequency component information within the same time period, and each patch represents multiple frequency domain distributions for a specific time period. These two branches, segmented in this way, have different emphases on time domain and frequency domain characteristics, which helps the attention mechanism to learn the two types of features of the data with different focuses. The segmented patch groups will be mapped to a high-dimensional space to capture the basic features of the signal.

[0033] 2. Secondly, for the segmented time-domain and frequency-domain patch groups, we will add position embeddings to them respectively. Utilizing the positional information provided by position embeddings, the model can perceive the position and order of elements in the input sequence. Adding position embeddings to the time-domain patch groups means that each patch contains frequency-domain data and temporal position information within a specific time period, helping the model understand signal changes over time. Adding position embeddings to the frequency-domain patch groups means that each patch contains time-series data of specific frequency components and their frequency position information, helping the model understand signal changes at different frequencies.

[0034] 3. We fed the two sets of time-frequency data with added positional encoding into the Transformer encoder module for training. We utilized a multi-head attention mechanism to perform parallel processing and multi-head segmentation of the input patch groups, effectively learning the data features. In the self-attention layer, we calculated the correlation between each position in the input signal and all other positions, highlighting important features while suppressing unimportant parts, generating a weighted feature representation. The output at each input position is the weighted average of all input positions, where the weights are determined by the similarity (correlation) of the input signals. The self-attention layer deeply extracts features from the input signal, identifies complex signal patterns, and enhances the model's ability to recognize complex signals through a non-linear activation function.

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

[0036] 4. We fuse the two sets of features obtained, and then perform operations such as linear transformation, activation function, and normalization on the fused features to finally complete the model design and realize the identification of modulation mode; this improves the accuracy and efficiency of the identification of modulation mode of the modulated OFDM signal.

[0037] 5. We use the model to train the generated dataset. Through training, our model can achieve an accuracy of 93% for this type of time-frequency data. Attached Figure Description

[0038] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0039] Specific implementation method one: Combining Figure 1 This embodiment describes the specific process of the low-Earth orbit satellite communication signal identification method based on the time-frequency Transformer model:

[0040] Step 1: Acquire OFDM signal data, modulate the OFDM signal data (modulation method such as PSK, QAM, ASK, etc.), obtain tagged modulated OFDM signal data, and use the modulated OFDM signal data as the training set.

[0041] Step 2: Construct a deep learning model. Use the training set as input and the classification results as output. Train the deep learning model until it converges to obtain a trained deep learning model. The specific process is as follows:

[0042] Step 2: First, divide the modulated OFDM signal data into the time domain and frequency domain respectively to obtain the time domain matrix E. t and frequency domain matrix E f ;

[0043] Step 22: For the time domain matrix E t and frequency domain matrix E f By adding positional encodings, we obtain matrix Z. t And matrix Z f ;

[0044] Steps two and three: Calculate matrix Z t The inputs are sequentially fed into M layers of Transformer encoders, and the last Transformer encoder outputs matrix R. t ;

[0045] matrix Z f The input is fed into a Transformer encoder, and the Transformer encoder outputs a matrix R. f ;

[0046] Step 24: For matrix R t and R f Feature fusion is performed to obtain a set of features R that include time-frequency characteristics;

[0047] Step 25: Input the features R containing time-frequency characteristics into flatten, dropout, and MLP in sequence. Input the MLP output features into the Dense layer. The Dense layer outputs the class probabilities to obtain the classification results.

[0048] Step 26: Repeat steps 21 to 25 until convergence, and obtain the 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, and performs a total of N independent normalizations. The normalization formula is as follows:

[0051]

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

[0053] We will obtain the Z with added position encoding t and Z f The input is fed into the Transformer encoder, which mainly includes a multi-head self-attention mechanism and a feedforward neural network (MLP) module, as well as layer normalization and skip connections between them. Before applying the multi-head self-attention mechanism, layer normalization will affect Z... t and Z f Normalization is performed to accelerate and stabilize the training process of the model. The processed data is then input into a multi-head attention module. The input data undergoes a linear transformation to generate query (Q), key (K), and value (V) vectors. These vectors are then divided into multiple heads to compute the attention outputs for different subspaces in parallel.

[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 weights are computed using a softmax function. The calculation for each head is as follows:

[0058]

[0059] Among them, QK T It is the dot product of the query matrix and the key matrix. This is a scaling factor to prevent the gradient from vanishing due to excessively large dot product values. The softmax function converts similarity scores into a probability distribution.

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

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

[0062] Among them W O ∈R D×D This is the output learnable weight matrix. The attention weight matrix represents the degree of attention each element in the input sequence receives from all other elements. Through weighted summation, the attention mechanism generates new feature representations that integrate important time and frequency domain information, improving the model's expressive power. With multiple attention heads, the model can process input data from different perspectives, extracting rich time and frequency domain feature representations, greatly enhancing the model's ability to extract and express signal features. The outputs of the multi-head self-attention are then sequentially passed through layer normalization and skip connections to ensure smooth information flow and reduce the gradient vanishing problem. Finally, two sets of features are obtained, which we represent using R... t and R f To express.

[0063] Dropout is a regularization technique used to prevent overfitting. By randomly setting a certain percentage (in this case, 50%) of neurons to output zero, it can enhance the model's generalization ability, especially when training data is limited or the model is complex. MLP (Multilayer Perceptron) uses two linear layers and its activation function employs GELU to further process and extract high-level features. We input the extracted features into a Dense layer, setting its output unit number to the number of classes, mapping the MLP-extracted features to the final classification result, calculating the final class probabilities, and finally outputting the classification result.

[0064] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that, in step two-one, the modulated OFDM signal data is divided into the time domain and frequency domain respectively to obtain the time domain matrix E. t and frequency domain matrix E f ;

[0065] The specific process is as follows:

[0066] Step 211: Divide the acquired modulated OFDM signal data in the time domain to obtain the time-domain matrix E. t The specific process is as follows:

[0067] Step 2: Divide the acquired modulated OFDM signal data X in the time domain. Each segment is x. t , The number of small pieces is N;

[0068] In the formula, 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, and H t The height W of the modulated OFDM signal data X corresponding to each patch block. t The length of the modulated OFDM signal data X corresponding to each patch;

[0069] Step 2112: Flatten each patch block into an H-shape. t ×W t A vector of ×C;

[0070] Steps 2-3: For each patch, the expanded vector is subjected to a linear transformation (fully connected layer). The vectors output by the linear transformation form the time-domain matrix E. t Mapping the vectors to a fixed-dimensional feature space yields the temporal embedding matrix E. t ;

[0071] Step 2.12: Divide the acquired modulated OFDM signal data in the frequency domain to obtain the frequency domain matrix E. f The specific process is as follows:

[0072] Step 21: Divide the acquired modulated OFDM signal data in the frequency domain. Each segment is x. f , The number of small pieces is N;

[0073] In the formula, 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, and H f The height W of the modulated OFDM signal data X corresponding to each patch block. f The length of the modulated OFDM signal data X corresponding to each patch;

[0074] Step 2122: Flatten each patch block into an H-shape. f ×W f A vector of ×C;

[0075] Steps 2-3: The expanded vectors corresponding to each patch are subjected to a linear transformation (fully connected layer). The vectors output by the linear transformation form the frequency domain matrix E. f Mapping the vectors to a fixed-dimensional feature space yields the temporal embedding matrix E.f ;

[0076] In the PatchEmbedding layer, our data will be divided into two partitioning methods: time domain and frequency domain. The number of small blocks in both partitions is the same, N.

[0077] The other steps and parameters are the same as in Specific Implementation Method 1.

[0078] Specific Implementation Method Three: This implementation method differs from Specific Implementation Method One or Two in that the expression for the number of small pieces N in step two is as follows:

[0079]

[0080] Other steps and parameters are the same as in specific implementation method one or two.

[0081] Specific Implementation Method Four: This implementation method differs from Specific Implementation Methods One to Three in that the expression for the number of small pieces N in step Two to Two is as follows:

[0082]

[0083] The other steps and parameters are the same as those in one of the specific implementation methods one to three.

[0084] Specific Implementation Method Five: This implementation method differs from Specific Implementation Methods One to Four in that, in step two-two, the time-domain matrix E... t and frequency domain matrix E f By adding positional encodings, we obtain matrix Z. t And matrix Z f The specific process is as follows:

[0085] Positional Embedding

[0086] To enable the model to capture the order and spatial location information of patches, we will add positional encoding to both the time-domain and frequency-domain patch groups. In this model design, we have chosen to use learnable positional encoding. This method allows the model to automatically adjust the positional encoding based on data during training, potentially leading to better adaptation 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 position encoding matrix P t Adding them together yields a new input matrix Z. t ;

[0089] Step 2. Initialize a learnable position coding matrix P for the frequency domain matrix. f ;

[0090] The time domain matrix E f With position encoding matrix P f Adding them together yields a new input matrix Z. f ;

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

[0092] The other steps and parameters are the same as those in one of the specific implementation methods one to four.

[0093] Specific Implementation Method Six: This implementation method differs from Specific Implementation Methods One to Five in that, in step two-two-one, a learnable position encoding matrix P for the time-domain matrix is ​​initialized. t ;

[0094] The time domain matrix E t With position encoding matrix P t Adding them together yields a new input matrix Z. t ;

[0095] The specific process is as follows:

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

[0097] Matrix P t The matrix has a size of 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 that can be updated during training.

[0098] The time domain matrix E t With position encoding matrix P t Adding them together yields a new input matrix Z. t ;

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

[0100] The other steps and parameters are the same as those in one of the specific implementation methods one to five.

[0101] Specific Implementation Method Seven: This implementation method differs from Specific Implementation Methods One through Six in that, in step two, a learnable position coding matrix P for the frequency domain matrix is ​​initialized.f ;

[0102] The time domain matrix E f With position encoding matrix P f Adding them together yields a new input matrix Z. f ;

[0103] The specific process is as follows:

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

[0105] Matrix P f The matrix has a size of 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 that can be updated during training.

[0106] The time domain matrix E f With position encoding matrix P f Adding them together yields a new input matrix Z. f ;

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

[0108] The other steps and parameters are the same as those in one of the specific implementation methods one to six.

[0109] Specific Implementation Method Eight: This implementation method differs from Specific Implementation Methods One to Seven in that, in step two-four, the matrix R... t and R f Feature fusion is performed to obtain a set of features R containing time-frequency characteristics; the expression is:

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

[0111] Where Concat(,) represents concatenation; R represents features containing time-frequency characteristics.

[0112] The other steps and parameters are the same as those in one of the specific implementation methods one to six.

[0113] Specific Implementation Method Nine: This implementation method differs from Specific Implementation Methods One to Eight in that 3 ≤ M ≤ 7 in the M-layer Transformer encoder.

[0114] The other steps and parameters are the same as those in one of the specific implementation methods one to eight.

[0115] Specific Implementation Method 10: This implementation method differs from Specific Implementation Methods 1 to 9 in that M = 4 in the M-layer Transformerencoder.

[0116] The other steps and parameters are the same as those in any of the specific implementation methods one to nine.

[0117] This invention may have other embodiments. Without departing from the spirit and essence of this invention, those skilled in the art can make various corresponding changes and modifications according to this invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. A method for identifying low-Earth orbit satellite communication signals based on a time-frequency Transformer model, characterized in that: The specific process of the method is as follows: Step 1: Acquire OFDM signal data, modulate the OFDM signal data to obtain tagged modulated OFDM signal data, and use the modulated OFDM signal data as the training set. Step 2: Construct a deep learning model. Use the training set as input and the classification results as output. Train the deep learning model until it converges to obtain a trained deep learning model. The specific process is as follows: Step 2: First, divide the modulated OFDM signal data into the time domain and frequency domain respectively to obtain the time domain matrix. and frequency domain matrix ; Step 22: For the time domain matrix and frequency domain matrix By adding positional encodings, we obtain matrices. sum matrix ; The specific process is as follows: Step 221: Initialize a learnable position encoding matrix for the time-domain matrix. ; Time domain matrix With position encoding matrix Add them together to get a new input matrix. ; Step 2. Initialize a learnable position encoding matrix for the frequency domain matrix. ; Time domain matrix With position encoding matrix Add them together to get a new input matrix. ; Steps two and three: Transform the matrix Enter in sequence The last Transformer encoder layer outputs a matrix. ; matrix The input is fed into the Transformer encoder, and the Transformer encoder outputs a matrix. ; Step 2.4: For the matrix and Feature fusion is performed to obtain a set of features that include time-frequency characteristics. The expression is: in, Indicates splicing; This indicates a feature that includes time-frequency characteristics; Step 25: Incorporate features containing time-frequency characteristics The flatten, dropout, and MLP layers are input sequentially. The MLP outputs features, which are then input into the Dense layer. The Dense layer outputs class probabilities, and the classification result is obtained. Step 26: Repeat steps 21 to 25 until convergence, and obtain the 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 method for identifying low-Earth orbit satellite communication signals based on the time-frequency Transformer model according to claim 1, characterized in that: In step two, the modulated OFDM signal data is divided into time and frequency domains respectively to obtain a time-domain matrix. and frequency domain matrix ; The specific process is as follows: Step 211: Divide the acquired modulated OFDM signal data in the time domain to obtain the time-domain matrix. The specific process is as follows: Step 2: Process the acquired modulated OFDM signal data. The time domain is divided, and each small block after division is... , The number of small pieces is ; In the formula, , For modulated OFDM signal data height, For modulated OFDM signal data Length, For modulated OFDM signal data The number of channels, Modulated OFDM signal data corresponding to each block height, Modulated OFDM signal data corresponding to each block Length; Step 2112: Flatten each block into a single piece. ; Step 2-3: Perform a linear transformation on the expanded vector corresponding to each block. The vectors output by the linear transformation form a time-domain matrix. ; Step 2.12: Divide the acquired modulated OFDM signal data in the frequency domain to obtain the frequency domain matrix. The specific process is as follows: Step 21: Divide the acquired modulated OFDM signal data in the frequency domain. Each segment is then divided into smaller blocks. , The number of small pieces is ; In the formula, , For modulated OFDM signal data height, For modulated OFDM signal data Length, For modulated OFDM signal data The number of channels, Modulated OFDM signal data corresponding to each block height, Modulated OFDM signal data corresponding to each block Length; Step 2: Flatten each block into a single piece. ; Steps 2-3: Perform a linear transformation on the expanded vectors corresponding to each block. The vectors output by the linear transformation form a frequency domain matrix. .

3. The method for identifying low-Earth orbit satellite communication signals based on the time-frequency Transformer model according to claim 2, characterized in that: The number of small pieces in step two-one The expression is: 。 4. The method for identifying low-Earth orbit satellite communication signals based on the time-frequency Transformer model according to claim 3, characterized in that: The number of small pieces in step 2121 The expression is: 。 5. The method for identifying low-Earth orbit satellite communication signals based on the time-frequency Transformer model according to claim 4, characterized in that: In step 221, a learnable position encoding matrix for the time-domain matrix is ​​initialized. ; Time domain matrix With position encoding matrix Add them together to get a new input matrix. ; The specific process is as follows: Initialize a learnable position encoding matrix for the time-domain matrix. ; matrix Size is ,in It is the sequence length. It is the embedded dimension; Time domain matrix With position encoding matrix Add them together to get a new input matrix. ; 。 6. The method for identifying low-Earth orbit satellite communication signals based on the time-frequency Transformer model according to claim 5, characterized in that: In step two, a learnable position encoding matrix for the frequency domain matrix is ​​initialized. ; Time domain matrix With position encoding matrix Add them together to get a new input matrix. ; The specific process is as follows: Initialize a learnable position encoding matrix for the frequency domain matrix. ; matrix Size is ,in It is the sequence length. It is the embedded dimension; Time domain matrix With position encoding matrix Add them together to get a new input matrix. ; 。 7. The method for identifying low-Earth orbit satellite communication signals based on the time-frequency Transformer model according to claim 6, characterized in that: The In the layer Transformer encoder .

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

Images