Satellite signal modulation identification method based on enhanced multi-scale feature attention network

Abstract

The invention discloses a satellite signal modulation recognition method based on an enhanced multi-scale feature attention network, and the method comprises the steps: obtaining a to-be-recognized satellite signal, inputting the to-be-recognized satellite signal into a trained enhanced multi-scale feature attention network, and obtaining a satellite signal modulation recognition result. The training process of the enhanced multi-scale feature attention network comprises the following steps: acquiring a satellite communication data set, and performing sample expansion to obtain an augmented sample; performing noise removal processing and feature extraction on the augmented sample to obtain denoised data features; inputting the de-noised data features into a global context sensing module, extracting modulation features of different frequencies and time scales, and performing feature fusion on the modulation features and the de-noised data features to obtain fusion features; and performing feature reconstruction and channel recovery on the fusion features, inputting the fusion features into a Softmax classifier, performing classification prediction, calculating classification loss, and optimizing network parameters through the classification loss. Under a complete blind recognition scene, the accuracy of a satellite signal modulation recognition result can be obviously improved.

Description

Technical Field

[0001] The present invention belongs to the field of satellite communication technology, and in particular relates to a satellite signal modulation recognition method based on an enhanced multi-scale feature attention network. Background Art

[0002] In recent years, the National Aeronautics and Space Administration (NASA) has actively promoted research on future-oriented communication architectures, focusing on cognitive communication technologies. This research aims to enhance the data processing capabilities and scientific information return efficiency of next-generation satellite communication systems, while optimizing transmission quality and simplifying operational processes. In this technical system, blind detection of satellite signals is the process of detecting, identifying, and analyzing satellite signals in the absence of prior information. The main objectives include signal presence detection, parameter estimation, and modulation mode identification. The present invention focuses on modulation mode identification. Automatic Modulation Recognition (AMR) technology can select the most suitable scheme for the current transmission environment from a preset modulation mode library based on specific communication requirements and channel conditions, achieving the optimal balance between reliability and rate. In non-cooperative communication systems, since the receiver does not know the sender's signal parameters, modulation mode, etc. in advance, it is impossible to directly demodulate the signal to obtain information. Blind modulation recognition becomes a key link in non-cooperative communication systems. In the civilian field, radio management departments can use modulation identification technology to regulate the spectrum and prevent illegal users from occupying scarce spectrum resources. In the military field, modulation identification technology can help to crack and interfere with enemy communication systems, helping to seize the initiative on the battlefield.

[0003] With the continuous advancement of aerospace and satellite communication technologies, the modulation methods and parameters of satellite signals have become increasingly complex and diverse. In addition, the increasing complexity of communication channels has brought greater challenges to modulation method identification and put forward higher requirements for classifiers.

[0004] Furthermore, practical applications must consider factors such as device size, weight, and power consumption. Due to environmental influences such as space radiation and extreme temperatures, the performance of onboard computers significantly lags behind that of ground-based equipment. Therefore, modulation scheme recognition methods for on-orbit satellite applications must be lightweight and computationally efficient to adapt to the harsh space environment and limited computing resources. Summary of the Invention

[0005] The purpose of the present invention is to overcome the shortcomings of the existing technology, especially when processing low signal-to-noise ratio signals of satellite signals, where it is difficult to extract the unique characteristics of the modulation scheme due to noise interference, thereby affecting the learning ability of the automatic modulation classification model.

[0006] To solve the problems of the existing technology, the present invention proposes a satellite signal modulation recognition method based on an enhanced multi-scale feature attention network, which can significantly improve the recognition accuracy under different signal-to-noise ratio conditions while achieving a substantial reduction in the number of model parameters.

[0007] The technical ideas for achieving the purpose of the present invention are: by constructing an enhanced multi-scale feature attention network, applying soft thresholding and efficient channel attention mechanism, selectively suppressing noise and retaining key signal features; using multi-scale dilated convolution and spatial pyramid pooling, it is possible to simultaneously extract global and local contextual information of different time scales and frequency features; by introducing the query and key construction method to construct contextual information, the ability to understand the overall signal is enhanced.

[0008] In a completely blind recognition scenario, the satellite signal modulation recognition method based on the enhanced multi-scale feature attention network includes:

[0009] The satellite signal to be identified is obtained and input into the trained enhanced multi-scale feature attention network to obtain the satellite signal modulation recognition result.

[0010] The training process of the enhanced multi-scale feature attention network includes:

[0011] S1: Obtain satellite communication data sets and perform sample expansion using phase rotation on the complex plane to obtain augmented samples;

[0012] S2: Remove noise and extract features from the augmented samples to obtain denoised data features;

[0013] S3: Input the denoised data features into the global context perception module to extract the modulation features of different frequencies and time scales, and fuse the extracted modulation features of different frequencies and time scales with the denoised data features to obtain fused features;

[0014] S4: After feature reconstruction and channel restoration of the fused features, the fused features are input into the Softmax classifier for classification prediction, classification loss is calculated, and the network parameters are optimized through the classification loss.

[0015] Furthermore, the network structure of the enhanced multi-scale feature attention network includes: input layer, data enhancement module, denoising convolution module, global context perception module, global average pooling layer, fully connected layer, Softmax classifier and output layer, where:

[0016] Input layer, used to obtain input data;

[0017] The data enhancement module is used to expand the samples of the input data to obtain augmented data;

[0018] The denoising convolution module is used to denoise the augmented data and extract features to obtain the denoised data features;

[0019] The global context perception module is used to extract the modulation features of different frequencies and time scales from the denoised data features, and perform feature fusion on the denoised data features to obtain fused features;

[0020] The first global average pooling layer is used to compress the fused features and extract global information;

[0021] The fully connected layer is used to reconstruct the compressed features and restore the channels to obtain the reconstructed features;

[0022] Softmax classifier, used to classify and predict the reconstructed features to obtain classification prediction results;

[0023] The output layer is used to output the final classification prediction result, which is the satellite signal modulation recognition result.

[0024] Furthermore, the denoising convolution module includes a convolution layer, a batch normalization module, a soft threshold denoising layer and a channel attention layer, and the global context perception module includes a spatial pyramid pooling layer and a multi-scale dilated convolution layer.

[0025] The beneficial effects of the present invention are as follows: the comprehensive performance of satellite communication signal modulation recognition is significantly improved in a completely blind recognition scenario. First, the present invention effectively suppresses noise interference and autonomously extracts key signal features by fusing soft threshold denoising and channel attention mechanism without relying on prior information such as carrier synchronization and signal timing, thereby solving the feature ambiguity problem caused by low signal-to-noise ratio and phase uncertainty in traditional methods. Secondly, the present invention combines multi-scale dilated convolution with spatial pyramid pooling for collaborative design, and realizes global perception and hierarchical analysis of signal time-frequency characteristics without artificial prior guidance, accurately capturing cross-scale modulation laws from transient phase jumps to long-term statistical distributions. Thirdly, the network model of the present invention adopts a lightweight structure optimization strategy to meet the stringent constraints of the satellite environment on resource efficiency, and replaces the traditional manual parameter calibration process with an end-to-end adaptive learning mechanism to achieve autonomous reasoning of unknown signal modulation types. Finally, based on phase rotation augmentation and dynamic channel interference simulation technology, the model of the present invention has strong robustness against Doppler frequency shift, co-channel interference and heterogeneous channel damage in a completely blind state, forming a full-chain solution from front-end signal blind processing to autonomous determination of modulation characteristics, providing core technical support for signal perception and cognitive decision-making in non-cooperative satellite-to-ground communication scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 2 is a schematic diagram of the network structure of the enhanced multi-scale feature attention network in an embodiment of the present invention;

[0027] Figure 2 is a flow chart of the training steps for enhancing the multi-scale feature attention network in an embodiment of the present invention;

[0028] Figure 3 (a) is a recognition accuracy curve corresponding to different signal-to-noise ratios on a simulated satellite signal dataset according to an embodiment of the present invention;

[0029] Figure 3 (b) is a confusion matrix for identifying different modulated signals on a simulated satellite signal dataset according to an embodiment of the present invention;

[0030] Figure 4 It is the accuracy curve of the embodiment of the present invention and different existing models under different signal-to-noise ratios during simulation verification. DETAILED DESCRIPTION

[0031] The terms "first," "second," "third," "fourth," and so on, in the specification and claims of this application and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate and are merely used to describe the manner in which objects with the same attributes are described in the embodiments of this application.

[0032] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0033] An embodiment of the present invention proposes a satellite signal modulation recognition method based on an enhanced multi-scale feature attention network, the method comprising:

[0034] The satellite signal to be identified is obtained and input into the trained enhanced multi-scale feature attention network to obtain the satellite signal modulation recognition result.

[0035] Figure 1 2 is a schematic diagram of the network structure of the enhanced multi-scale feature attention network in an embodiment of the present invention. Figure 1 In the diagram, “Add” represents addition, “multiply” represents multiplication, “Concatenation” represents concatenation, “Channel” represents channel, “Average” represents averaging, and “Dense” represents a fully connected layer.

[0036] Reference Figure 1As shown in the figure, the network structure of the enhanced multi-scale feature attention network includes: input layer, data augmentation module, denoising convolution module, global context perception module, global average pooling (GAP) layer, fully connected layer, softmax classifier and output layer. The denoising convolution module includes convolution layer, batch normalization module, soft threshold denoising layer and channel attention layer. The soft threshold denoising layer includes absolute value (Abs) layer, global average pooling (GAP) layer, flattening layer and activation layer. The global context perception module includes spatial pyramid pooling layer and multi-scale dilated convolution layer. Among them:

[0037] The input layer is used to obtain input data, which is satellite communication signal data.

[0038] The data augmentation module is used to perform sample expansion on the input data to obtain augmented data. Specifically, the input data is subjected to phase rotation on the complex plane to perform sample expansion and phase invariance modeling.

[0039] The denoising convolution module is used to denoise and extract features from the augmented data, eliminating the influence of specific range characteristics to attenuate noise features and obtain denoised data features. This module combines a deep residual architecture with a soft threshold denoising layer to filter and retain effective information. The channel attention layer focuses on key features that contribute to modulation recognition, suppressing narrowband interference and irrelevant features.

[0040] The global context-aware module performs global context learning on the denoised data, preserving both the short-term mutations and long-term statistical components of the signal envelope and extracting modulation features at different frequencies and time scales. In this module, pyramid pooling layers are stacked with maximum pooling layers with pooling windows of 2 and 4, respectively, for hierarchical pooling and concatenation to extract the global statistical characteristics of the signal. Dilated convolutional layers are stacked with dilation rates of 1 and 3, respectively, to expand the receptive field without sacrificing resolution and cover different inter-symbol interference periods.

[0041] The global average pooling (GAP) layer is used to compress the extracted modulation features of different frequencies and time scales and extract global information.

[0042] It should be noted that GAP is a parameter-free pooling operation that compresses the spatial dimensions of a feature map into a single value while preserving the channel dimension. Specifically, it directly converts the three-dimensional feature map (height × width × channels) into a one-dimensional vector (number of channels) by calculating the global mean of each channel, thereby replacing the traditional fully connected layer (Dense layer) to complete feature aggregation and dimensionality transformation.

[0043] The fully connected layer is used to reconstruct the compressed features and restore the channels to obtain the reconstructed features.

[0044] The Softmax classifier is used to perform classification prediction on the reconstructed features to obtain classification prediction results.

[0045] The output layer is used to output the final classification prediction results, that is, the satellite signal modulation recognition results.

[0046] Figure 2 Flowchart of the training steps for enhancing the multi-scale feature attention network in an embodiment of the present invention.

[0047] Reference Figure 1 、 2 As shown, the training process of the enhanced multi-scale feature attention network includes:

[0048] S1: Obtain a satellite communication dataset and perform sample expansion using phase rotation on the complex plane to obtain augmented samples.

[0049] Specifically, the acquired data set is subjected to sample expansion and phase invariance modeling using phase rotation on the complex plane. The calculation formula is:

[0050]

[0051] Satellite signals are usually represented in complex baseband form, and the two-dimensional real vectors corresponding to their complex envelopes are R(·) and L(·). Denotes the augmented sample, and X denotes the original sample. Based on modulation symmetry, θ∈{0,π / 2,π,3π / 2} is set to avoid constellation topology disruption and account for phase switching characteristics such as QPSK's 90° phase jumps. This approach not only simulates the phase uncertainty caused by carrier synchronization errors, Doppler residual phase, and transient ionospheric disturbances in satellite communications, but also ensures that all symbols of the same signal rotate synchronously, conforming to the time-domain continuity characteristics of true phase errors. Furthermore, this approach quadruples the augmented sample size to the original sample size.

[0052] S2: Remove noise and extract features from the augmented samples to obtain denoised data features.

[0053] In the illustrated embodiment, a denoising convolution module is used to remove noise from the augmented samples and extract features, resulting in denoised data features. A soft threshold denoising layer is built in conjunction with a deep residual architecture. By introducing frequency band focusing in the channel attention layer, a denoising convolution module is constructed. This module removes noise from the augmented samples and extracts sparse features, resulting in denoised data features.

[0054] Reference Figure 1As shown in the figure, the denoising convolution module includes a convolution layer, a batch normalization module, a soft threshold denoising layer and a channel attention layer. The soft threshold denoising layer includes an absolute value (Abs) layer, a global average pooling (GAP) layer, a flattening layer and an activation layer. Specifically:

[0055] The convolution layer is a one-dimensional convolution layer with a convolution kernel size of 3, a number of convolution kernels of 32, a convolution kernel step size of 1, and an activation function of ReLU.

[0056] The batch normalization module is used to perform batch normalization processing on the output of the convolutional layer.

[0057] The soft threshold denoising layer includes an absolute value (Abs) layer, a global average pooling (GAP) layer, a flattening layer, and an activation layer. It attenuates noise features by eliminating the influence of specific range characteristics. The soft threshold function is shown in formula (2):

[0058]

[0059] Among them, x represents the input feature, y represents the output feature, and τ is the value parameter.

[0060] Soft thresholding converts features close to zero to zero while retaining useful negative features; its derivative is either 0 or 1, which can effectively prevent the problems of gradient disappearance and gradient explosion, as shown in formula (3):

[0061]

[0062] Specifically, combined with the deep residual architecture, a hierarchical soft threshold operation is designed: for the input feature map X∈R W×H×C , calculate the soft threshold channel by channel:

[0063]

[0064] Here, α represents the dynamic threshold, which is initialized to 0.5. This threshold dynamically adjusts the feature shrinkage strength to avoid the traditional fixed threshold ignoring the local dynamic range of the signal.

[0065] The process of feature denoising using the soft threshold denoising layer is:

[0066] X denoise =X+f(SoftThreshold(X,τ)) (5)

[0067] Where f(·) is a 1×1 convolution feature map, and the SoftThreshold function is defined as:

[0068] SoftThreshold(x,τ)=sign(x)·max(|x|-τ,0) (6)

[0069] Specifically, the feature map is first subjected to absolute value processing and global average pooling, followed by linear projection to obtain scaling parameters. These scaling parameters are then processed by a sigmoid function and compressed to the range [0, 1]. Finally, the scaling parameters are multiplied by the absolute average of the feature map to derive the soft threshold. This structure ensures that the soft threshold is positive and limits its magnitude to within the maximum value of the feature map, thereby effectively filtering and retaining information. By soft-thresholding the original input data, the primary attributes are successfully retained while attributes considered noise are eliminated, thereby enhancing the saliency of high-level features.

[0070] The channel attention layer proposes a channel attention layer with adaptive kernel size based on the energy distribution of satellite signal frequency bands. Its working principle is to first perform global average pooling on each channel of the input feature map to obtain a channel-dimensional description vector, thereby extracting the global spatial information of each channel. Unlike the traditional compressed excitation attention mechanism that uses a fully connected layer to learn the dependencies between channels, the channel attention layer uses a one-dimensional convolution instead of a fully connected layer. This not only reduces the number of parameters but also can efficiently capture the local importance relationship between channels through the convolution kernel. In order to adaptively adjust the receptive field in networks of different sizes, it can adaptively select the one-dimensional convolution kernel size according to the number of channels:

[0071]

[0072] Among them, γ and b are adjustable hyperparameters, The convolution kernel size is forced to be an odd number to ensure the symmetry of the convolution calculation and the focusing of the frequency band energy, avoiding the problem of asymmetric padding. This adaptive adjustment balances efficiency and effect. Finally, the learned channel weights are applied to the original feature map through element-by-element multiplication.

[0073] S3: The denoised data features are input into the global context perception module to extract the modulation features of different frequencies and time scales. The extracted modulation features of different frequencies and time scales are fused with the denoised data features to obtain fused features.

[0074] In the illustrated embodiment, reference is made to Figure 1 As shown, the global context perception module includes a spatial pyramid pooling layer and a multi-scale dilated convolution layer. The multi-scale dilated convolution layer is stacked to set convolution layers with different dilation rates to capture multi-scale features in the time domain. The spatial pyramid pooling layer is stacked to set pooling layers of different scales to perform hierarchical pooling and splicing to extract the global statistical characteristics of the input data features.

[0075] Specifically, the global context perception module combines a multi-scale dilated convolution layer with a spatial pyramid pooling layer. The former captures the inter-symbol interference period through convolution kernels with different dilation rates (for example, the dilation rate is 1 or 3), and the latter uses hierarchical pooling to extract global statistical characteristics, enabling the module to simultaneously utilize global context information and local fine features, retaining the short-term mutations and long-term statistical components of the signal envelope, greatly enhancing the ability to capture modulation features of different frequencies and time scales, improving confusion problems such as BPSK and QPSK, and significantly optimizing computational efficiency.

[0076] The multi-scale dilated convolution layer embeds a local processing mechanism in a non-local operation framework and uses convolution kernels with multiple groups of dilation rates to capture multi-scale features in the time domain, as shown in formula (8):

[0077]

[0078] Where W d (k) is the convolution kernel weight with dilation rate d, K is the kernel length, X(t) is the time domain input signal, and F d (t,c) is the feature output corresponding to the expansion rate d, and the expansion rate d, f s is the sampling rate, T s is the symbol period length, which can cover the inter-symbol interference period.

[0079] By stacking convolutional layers with different expansion rates, it is possible to jointly model short-term mutations and long-term statistical components.

[0080] The spatial pyramid pooling layer is used to perform hierarchical pooling and splicing to extract the global statistical characteristics of the signal, as shown in formula (9):

[0081]

[0082] Where l is the pooling window length, E2 is the input feature, P l is the feature after pooling.

[0083] By performing pooling operations on windows of different sizes l (for example, l = 2 or 4), multi-scale feature representations are generated. This allows the module to simultaneously focus on local details of the signal (such as phase jumps) and global features (such as overall spectral characteristics), thereby more comprehensively understanding the characteristics of different modulation modes. Among them, smaller pooling windows can capture rapidly changing modulation features, while larger windows help identify slowly changing or periodic modulation patterns.

[0084] The input of the global context-aware module (i.e., denoised data) is represented as X, X∈R L×C , L represents the sequence length, and C represents the number of channels.

[0085] In the illustrated embodiment, reference is made to Figure 1 、 2 As shown in Figure 1, the global context-aware module is used to perform global context-aware learning on the denoised data to extract modulation features of different frequencies and time scales. The specific process includes:

[0086] Step a: Use two 1×1 convolution kernels W q and W k Map the input feature X to obtain different feature embeddings φ and in:

[0087]

[0088] Among them, C q Represents the convolution kernel W q The number of channels, C k Represents the convolution kernel W k The number of channels, That is, the number of channels is 0.5 times the original input X.

[0089] Step b: Construct the Query matrix, Key matrix and Value matrix. Specifically:

[0090] Flatten the embedded feature φ to L×C q The size of , recorded as Query; for embedded features Use spatial pyramid pooling and multi-scale dilated convolution to capture global and local context information respectively. Then concatenate features of different scales to obtain a new embedding key with the size of:

[0091]

[0092] Among them, L SPP Indicates the number of feature points obtained by SPP sampling;

[0093] Generate Value: Copy the Key once to get the Value, which is represented as follows:

[0094]

[0095] Step c: Calculate the similarity matrix between Query and Key, and use SoftMax normalization to obtain the attention weight, which is specifically expressed as:

[0096]

[0097] Among them, Q represents the Query matrix, K represents the Key matrix, (.) T Indicates the matrix transpose.

[0098] Step d: Use the attention weight A to perform a weighted summation of the values to obtain the context vector Attention(Q, K, V). This context vector Attention(Q, K, V) represents the extracted modulation features at different frequency and time scales. This multi-scale feature extraction capability is crucial for distinguishing between different modulation modes, such as AM, FM, PSK, and QAM.

[0099] The different frequency and time scale modulation features Attention (Q, K, V) are jump-connected with the denoised data features to obtain fusion features.

[0100] S4: After feature reconstruction and channel restoration of the fused features, the fused features are input into the Softmax classifier for classification prediction, classification loss is calculated, and the network parameters are optimized through the classification loss.

[0101] Perform output reconstruction and channel recovery on the fusion features, specifically:

[0102] The attention output is adjusted to the size of L×C, and the global average pooling layer GAP is used to compress the fusion features and extract global information. Then, the number of channels is restored to C through the fully connected layer Dense (W0 represents a 1×1 convolution kernel) to obtain the reconstructed features. The specific calculation formula is:

[0103] X′=W o (Reshape(Attention(Q,K,V)))+X (14)

[0104] Where X' represents the reconstructed features, X represents the denoised data features, W0(.) represents the convolution operation, Reshape(.) represents the feature reconstruction, and Attention(Q,K,V) represents the context vector, that is, the extracted modulation features of different frequencies and time scales.

[0105] The classification loss is the cross entropy loss between the predicted result and the true label, and its classification loss function is The calculation formula is:

[0106]

[0107] Among them, Y represents the set of true labels, represents the set of predicted labels; Indicates that the true label of the ath sample in the training sample is the bth label, Indicates that the predicted label of the ath sample in the training sample is the bth label; a∈[1,m], m is the number of training samples, b∈[1,c], c is the number of label categories.

[0108] Based on the Adam optimization algorithm and classification loss function, the enhanced multi-scale feature attention network is optimized. The optimization process includes:

[0109] Step 1: Initialize the Adam optimizer and set the learning rate to 0.001 and the batch size to 400. Use the categorical cross entropy loss function as the loss function for the lightweight deep learning network. Set the maximum number of training rounds to 1000.

[0110] Step 2: Input the training set samples into the network model, use the current model parameters for forward propagation, generate the model's prediction results; calculate the cross entropy loss value between the prediction results and the true labels.

[0111] Step 3: After forward propagation calculates the loss value, apply the backpropagation algorithm to calculate the gradient of each model parameter based on the loss value; use the Adam optimizer to update the parameters of the network model based on this gradient information to reduce the value of the loss function; and use the custom callback function SafeModelCheckpoint to save the updated model parameters to the specified file.

[0112] Step 4: Repeat steps 2 and 3. After each round of training, the model will be validated using the validation set and the loss value of the validation set will be calculated. The learning effect of the model will be judged by monitoring the validation loss. The learning rate will be adjusted through the callback mechanism to prevent overfitting during training.

[0113] After each training round, check the changes in validation loss:

[0114] If the validation loss does not decrease within 5 epochs: the current learning rate is multiplied by 0.5 for decay; this process continues until the learning rate drops to a minimum of 0.000001; if this lower limit is reached, the model parameters are readjusted and return to step 2 to continue training;

[0115] If the validation loss does not decrease within 50 epochs, the training process is stopped and the model weights that performed best during the training process are saved to obtain the trained lightweight deep learning neural network.

[0116] Simulation experiment

[0117] To verify the effectiveness of the method provided by the present invention, the technical personnel of the present invention developed an experimental plan and conducted simulation experiments on the invention scheme. The dynamic satellite channel model was constructed using the GNU Radio open source framework, strictly following the ITU-RM.2083 recommendation and the DVB-S2 communication standard. The key configuration parameters are as follows: the RF parameters adopt the typical configuration of C-band satellite communication, the center frequency f c =12GHz, symbol rate R s=25MBaud. Each signal sample is shaped by a root raised cosine filter with a roll-off factor of 0.35 and a filter order of 64, and unit energy normalization is performed to ensure that the training set, validation set, and test set meet the power consistency constraint. To simulate the high-speed motion scenario of low-orbit satellites, the maximum Doppler frequency deviation is configured to Δf = ±5.6kHz, whose value is determined by the orbital velocity v = 7.8km / s through Δf / f c =v / c (c is the speed of light). Depending on the settings, the received signal model can be expanded to the following formula:

[0118]

[0119] Where K is the Rice factor, which is used to describe the energy ratio of the direct path to the multipath component, and N p is the number of multipaths, α i ~CN(0,1) is the complex gain of the i-th path, T i is the time delay, n(t)~CN(0,σ n 2 ) is AWGN, and its power spectrum density N=σ n 2 / B, where B is the signal bandwidth. The channel impairment model integrates typical effects of satellite-to-ground links: the Rice factor K = 10dB simulates the direct path-dominated scenario, the three-path fading model describes multipath propagation, and the phase noise power spectral density N = 10 -8.5 f -2 rad2 / Hz simulates the unstable characteristics of the frequency source. For the single-user uplink model, additional co-channel interference was introduced during the test phase, with its power increased by 15dB compared to the main signal to evaluate the model's ability to resist interference from neighboring satellites. Considering the typical satellite signal types, the modulation type set can be set as follows:

[0120] Ω={8PSK,BPSK,CPFSK,GFSK,PAM4,16QAM,64QAM,QPSK,AM-DSB,AM-SSB,WBFM}(17)

[0121] Each modulated signal frame structure contains 1024 I / Q sampling points, corresponding to 40 symbol periods, and the symbol interval is T s = 40ns. A total of 220,000 samples are generated, divided into training samples, validation samples, and test samples in a ratio of 7:2:1, and the signal generation time of different subsets is strictly isolated within the 12-hour orbit period to avoid information leakage.

[0122] The experiment is set up as a completely blind recognition scenario, that is, the receiver has no prior information such as carrier frequency, symbol timing, phase synchronization, etc. Asynchronous communication characteristics are forced to be introduced in data generation: the starting position of the time slot is between 0 and 10T. sThe signals are uniformly and randomly distributed, and the signal frame structures of different modulation types are heterogeneous. All signals were verified using an Agilent N9020A RF analyzer, and their error vector magnitude (EVM) indicators were all better than -25dB, complying with the 3GPP TR 38.811 satellite-to-ground channel calibration specification.

[0123] Simulation conditions: NVIDIA GeForce RTX2080Ti GPU is used, the network model is implemented by Python software, based on the Tensorflow2.5.0 computing framework and accelerated by CUDA11.2. The calculation formula is:

[0124]

[0125] Among them, n represents the total number of samples, Y d represents the true label of the d-th sample, represents the predicted label of the d-th sample.

[0126] Simulation content and result analysis:

[0127] Simulation 1: The present invention is used to identify the signal modulation type under different signal-to-noise ratio conditions of different data sets, and its accuracy is calculated to obtain the curve of the change of recognition accuracy with signal-to-noise ratio. Figure 3 (a) and Figure 3 (b). Figure 3 (a) shows the recognition accuracy curve corresponding to different signal-to-noise ratios on the simulated satellite signal dataset. The horizontal axis represents the signal-to-noise ratio, and the vertical axis represents the recognition accuracy. Figure 3 (b) is the confusion matrix for identifying different modulation signals.

[0128] from Figure 3 It can be seen that the present invention can achieve an accuracy of nearly 90% when distinguishing high-order modulation modes such as QAM16 and QAM64 when the signal-to-noise ratio is 0dB or above, which reflects its ability to sensitively capture slight differences in signals. At the same time, for phase modulation modes such as BPSK and QPSK, the model also shows a high recognition accuracy rate, and can achieve good performance at a lower signal-to-noise ratio. Its misclassification is mainly manifested in the confusion of slight offsets of adjacent phase orders, such as 8PSK misjudged as QPSK, accounting for 3.8%. In particular, for multi-level modulation such as PAM4, the model can achieve a recognition accuracy rate of nearly 98% at a -5dB signal-to-noise ratio, showing its advantage in complex signal pattern recognition. For the amplitude modulation pairs of AM-DSB and AM-SSB, the two-way misclassification rate is 3.2%, which is directly related to the geometric similarity of the two in the envelope feature space.

[0129] In simulation 2, the present invention and the existing CGDNet, MCNet, IC-AMCNet, MCLDNN, PET-CGDNN and ResNet models are used to identify the signal modulation type. The total training time, single cycle training time, test set loss value, test set accuracy and maximum accuracy are calculated, and the parameter quantity is statistically analyzed. The results are shown in Table 1 and Figure 4 As shown:

[0130] Table 1 Comparison of simulation test results of the present invention and existing models

[0131]

[0132] From Table 1 and Figure 4 It can be seen that the present invention outperforms existing mainstream models in multiple key performance indicators, and has significant advantages and wide application potential. Its lightweight design greatly reduces the number of parameters, reduces storage requirements and computing resource consumption, and is suitable for deployment on resource-constrained devices such as mobile terminals and embedded systems. Compared with complex models such as ResNet and IC-AMCNet, the present invention improves computing efficiency and deployment flexibility while maintaining high performance. In terms of accuracy and classification capabilities, it has stronger feature extraction and data classification capabilities, surpassing traditional models such as CGDNet and MCNet. In addition, the present invention has better convergence and generalization capabilities, and maintains stable performance under different data distributions and application scenarios. In terms of training efficiency, the single-cycle training time is significantly shortened, which is conducive to rapid iteration and updating. The present invention also demonstrates high robustness and stability in complex environments, and can accurately identify targets in environments with large noise or changing conditions.

[0133] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing the relevant hardware through a program, and the program can be stored in a computer-readable storage medium, which may include: ROM, RAM, disk or CD, etc.

[0134] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. Satellite signal modulation recognition method based on enhanced multi-scale feature attention network, characterized by: include: Obtain the satellite signal to be identified and input it into the trained enhanced multi-scale feature attention network to obtain the satellite signal modulation recognition result; The training process of the enhanced multi-scale feature attention network includes: Obtain satellite communication data sets, perform sample expansion using phase rotation on the complex plane, and obtain enhanced samples; De-noise and extract features from the augmented samples to obtain denoised data features; The denoised data features are input into the global context perception module to extract the modulation features of different frequencies and time scales. The extracted modulation features of different frequencies and time scales are fused with the denoised data features to obtain fused features. After feature reconstruction and channel restoration of the fused features, the fused features are input into the Softmax classifier for classification prediction, classification loss is calculated, and the network parameters are optimized through the classification loss.

2. The satellite signal modulation recognition method based on enhanced multi-scale feature attention network according to claim 1 is characterized in that: The network structure of the enhanced multi-scale feature attention network includes: an input layer, a data enhancement module, a denoising convolution module, a global context perception module, a global average pooling layer, a fully connected layer, a Softmax classifier and an output layer, wherein: Input layer, used to obtain input data; The data enhancement module is used to expand the samples of the input data to obtain augmented data; The denoising convolution module is used to denoise the augmented data and extract features to obtain the denoised data features; The global context perception module is used to extract the modulation features of different frequencies and time scales from the denoised data features, and perform feature fusion on the denoised data features to obtain fused features; The first global average pooling layer is used to compress the fused features and extract global information; The fully connected layer is used to reconstruct the compressed features and restore the channels to obtain the reconstructed features. The Softmax classifier is used to classify and predict the reconstructed features to obtain the classification prediction results. The output layer is used to output the final classification prediction result, which is the satellite signal modulation recognition result.

3. The satellite signal modulation recognition method based on enhanced multi-scale feature attention network according to claim 2 is characterized in that: The denoising convolution module includes a convolution layer, a batch normalization module, a soft threshold denoising layer and a channel attention layer. The soft threshold denoising layer includes an absolute value layer, a second global average pooling layer, a flattening layer and an activation layer.

4. The satellite signal modulation recognition method based on enhanced multi-scale feature attention network according to claim 2 is characterized in that: The global context perception module includes a spatial pyramid pooling layer and a multi-scale dilated convolution layer. The multi-scale dilated convolution layer is stacked with convolution layers of different dilation rates to capture multi-scale features in the time domain. The spatial pyramid pooling layer is stacked with pooling layers of different pooling window sizes to perform hierarchical pooling and splicing to extract the global statistical characteristics of the input data features.

5. The satellite signal modulation recognition method based on enhanced multi-scale feature attention network according to claim 1 or 2 is characterized in that: The global context-aware module is used to perform global context-aware learning on the denoised data features to extract modulation features of different frequencies and time scales. The specific process includes: performing feature-feature embedding conversion on the input data to obtain the first feature embedding and the second feature embedding; Construct Query matrix, Key matrix and Value matrix; Calculate the similarity matrix between Query and Key, and use Softmax normalization to obtain the attention weight A; Use the attention weight A to perform weighted summation on Value to obtain the modulation features Attention(Q, K, V) of different frequencies and time scales.

6. The satellite signal modulation recognition method based on enhanced multi-scale feature attention network according to claim 2 is characterized in that: The reconstructed feature is expressed as: X′=W o (Reshape(Attention(Q,K,V)))+X Where X' represents the reconstructed features, X represents the denoised data features, W0(.) represents 1×1 convolution, Reshape(.) represents feature reconstruction, and Attention(Q,K,V) represents the modulation features at different frequencies and time scales.

7. The satellite signal modulation recognition method based on enhanced multi-scale feature attention network according to claim 1 is characterized in that: The classification loss is the cross entropy loss between the predicted result and the true label, and its classification loss function is The calculation formula is: Among them, Y represents the set of true labels, represents the set of predicted labels; Indicates that the true label of the ath sample in the training sample is the bth label, Indicates that the predicted label of the ath sample in the training sample is the bth label; a∈[1,m], m is the number of training samples, b∈[1,c], c is the number of label categories.

Images