Satellite signal identification method and model based on manifold spatial features
By using the wavelet convolution encoder WCAE network in satellite signal recognition to extract manifold space features and perform feature fusion, the problems of low accuracy and high computational complexity in complex signal environments are solved, and end-to-end automatic recognition and accurate recognition of satellite signals are realized.
Patent Information
- Application Number
- CN202510177994.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-18
- Publication Date
- 2025-06-10
AI Technical Summary
Traditional satellite signal recognition methods have low accuracy in complex and variable signal environments, are susceptible to noise interference, have high computational complexity, and require manual extraction of signal characteristics, making it difficult to promote and apply.
Using a satellite signal recognition method based on manifold space characteristics, multi-scale manifold space features are extracted through the WCAE network of the wavelet convolution encoder, and feature fusion and classification recognition are performed to realize end-to-end automatic recognition of the signal.
It improves the robustness and accuracy of satellite signal recognition, reduces the dependence on professional knowledge of signal characteristics, realizes automatic recognition without human intervention, and is suitable for time-varying satellite signal recognition in complex scenarios.
Smart Images
Figure CN120123866A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of satellite signal carrier monitoring research, and particularly relates to a satellite signal recognition method and model based on manifold space features in this field. Background Art
[0002] Due to the complex and variable channel environment in the satellite communication process, there are problems such as inconsistent spatio-temporal benchmarks and data discreteness in multi-source electromagnetic data such as space-based spectrum sensing data and electromagnetic signal feature data obtained in a complex electromagnetic environment. At the same time, satellite signals are easily affected by the group delay characteristic, resulting in more serious inter-symbol interference than terrestrial communication, making the signal-to-noise ratio of the received signal low and there being serious interference. Due to facing the above problems, the existing satellite signal recognition methods inevitably face the following difficulties:
[0003] The result accuracy of traditional satellite signal recognition methods is easily affected by the signal feature estimation accuracy. Traditional satellite signal recognition methods are based on instantaneous features such as the amplitude, frequency, and phase of the signal, and are easily affected by noise interference, resulting in large fluctuations in the eigenvalue. Although methods based on features such as high-order cumulants and cyclic spectra are less sensitive to noise, they have disadvantages such as high computational complexity and high requirements for the cumulants and quality of signal data. In a complex satellite communication environment, the quality of the obtained satellite signals is poor and the accuracy of feature estimation is low, resulting in a sharp decline in the recognition performance of the signals.
[0004] The limitation of traditional satellite signal recognition methods lies in the need to manually extract signal features. The existing methods rely on manual participation to extract signal features and require prior knowledge of signal features to ensure that the extracted features can clearly distinguish different signal types. On the premise of fully extracting signal features, traditional methods can achieve good recognition effects. However, for different recognition tasks, different signal features need to be extracted, which poses certain challenges for the popularization and application of such methods. Summary of the Invention
[0005] The technical problem to be solved by the present invention is to provide a satellite signal recognition method and model based on manifold space features to achieve end-to-end automatic recognition of signals in view of the disadvantages of traditional recognition algorithms in dealing with complex and variable satellite signal tasks.
[0006] The present invention adopts the following technical solutions:
[0007] A satellite signal recognition method based on manifold space features, the improvement lies in that it includes the following steps:
[0008] Step 1, input the IQ sample data of the original satellite signal into the encoding network of the wavelet convolutional encoder WCAE network;
[0009] Step 2, in the encoding network, through layer-by-layer downsampling and batch normalization operations, extract the manifold space features in the sample data:
[0010] The extraction formula for the manifold space features is:
[0011]
[0012] In the above formula, S is the original IQ sample data, F i is the manifold space feature of the i-th layer, Conv is the convolutional layer, DWT is the discrete wavelet transform layer, and ReLU is the activation layer;
[0013] Step 3, transfer the manifold space features extracted by the last group of downsampling layers to the decoding network, and through continuous upsampling layers, realize the reconstruction of the original satellite signal data;
[0014] Step 4, use the loss function to continuously optimize the WCAE network parameters. The loss function selects the binary cross-entropy function BCE and the consistency enhancement loss function CEL:
[0015] Loss = L bce (p, g) + L cel (p, g)
[0016] When the prediction result p is exactly the same as the original sample data g, the loss value is 0; when the two are completely different, the maximum penalty is given and the loss value is 1;
[0017] Step 5, after the training of the feature extraction and data reconstruction part is completed, retain the encoding network of the WCAE network, remove the decoding network, and use the multi-scale manifold space features extracted by the encoding network for subsequent signal recognition;
[0018] Step 6, fusion of multi-scale manifold space features:
[0019] Fuse the multi-scale manifold space features F i extracted by different downsampling layers in the encoding network to obtain the fused feature F = [F 1 , F 2 ... F i ...];
[0020] Step 7, input the fused manifold space features into the fully connected layer to map the manifold space features of different scales to the same dimension;
[0021] Step 8, use the softmax classifier to classify the fused features, continuously optimize the network parameters through supervised learning, compare the output signal type with the actual label to ensure that the contribution degree of different scale features reaches the optimal during the process of optimizing the network parameters, and output the recognition result O of the satellite signalk , and its formula is as follows:
[0022]
[0023] In the above formula, u k represents the output features of the last dense layer, and K represents the total number of signal types;
[0024] The output is written as:
[0025] O = [O 1 , O 2 ,..., O K T
[0026] Take the maximum value of O as the type of this signal.
[0027] Furthermore, in step 2, the downsampling layer consists of a convolutional layer, a discrete wavelet transform layer, and a ReLU activation layer.
[0028] Furthermore, in step 3, the upsampling layer consists of an inverse discrete wavelet transform layer, a transposed convolutional layer, and a ReLU activation layer.
[0029] Furthermore, in step 8, the initial learning rate is set to 1e-3, and the "Poly" strategy is adopted to dynamically adjust the learning rate. The specific formula is as follows:
[0030]
[0031] In the above formula, lr represents the learning rate of the current epoch, base_lr is the initial learning rate, epoch is the current iteration number, num_epoch is the maximum number of iterations, and Power is set to 1.1.
[0032] Furthermore, in step 8, the stochastic gradient descent algorithm is used to optimize and train the network parameters. The weight decay is set to 5e-4, the momentum is set to 0.9, the training batch size is set to 8, and the number of iterations is set to 50.
[0033] A satellite signal recognition model based on manifold space features, the improvement of which lies in: it consists of an encoding network, a decoding network, a feature fusion layer, and an identification output layer; the encoding network reduces the dimension of the original IQ sample data and extracts the manifold space features in the data; the decoding network reconstructs the original data using the manifold space features to verify the accuracy and robustness of the manifold space features; the feature fusion layer integrates the extracted multi-scale manifold space features, and the identification output layer realizes the identification of satellite signals based on the fused comprehensive features.
[0034] A satellite signal recognition model based on manifold space features, characterized in that: the satellite signal recognition model based on manifold space features consists of an encoding network, a decoding network, a feature fusion layer and an identification output layer; the encoding network performs dimensionality reduction processing on the original IQ sample data to extract the manifold space features in the data; the decoding network uses the manifold space features to reconstruct the original data to verify the accuracy and robustness of the manifold space features; the feature fusion layer integrates the extracted multi-scale manifold space features, and the identification output layer realizes the recognition of satellite signals based on the integrated comprehensive features.
[0035] A computer-readable storage medium, characterized in that: the computer-readable storage medium stores a computer program, and the computer program is configured to implement the steps of the above-mentioned satellite signal recognition method based on manifold space features when called by a processor.
[0036] A satellite signal recognition device based on manifold space features, the device includes at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the above-mentioned satellite signal recognition method based on manifold space features to realize the recognition of satellite signals with manifold space features.
[0037] The beneficial effects of the present invention are:
[0038] The present invention effectively solves the drawbacks of traditional recognition algorithms in dealing with complex and changeable satellite signal tasks, and realizes end-to-end automatic recognition of signals. Technical points: Input the IQ sample data of the original satellite signal into the encoding network of the wavelet convolutional encoder WCAE network; in the encoding network, extract the manifold space features in the sample data through layer-by-layer downsampling and batch normalization operations; transfer the manifold space features extracted by the last group of downsampling layers to the decoding network, and realize the reconstruction of the original satellite signal data through continuous upsampling layers; use the loss function to continuously optimize the WCAE network parameters, and the loss function selects the binary cross-entropy function and the consistency enhancement loss function; use the multi-scale manifold space features extracted by the encoding network for subsequent signal recognition; fusion of multi-scale manifold space features; map the manifold space features of different scales to the same dimension; use the softmax classifier to classify the fused features, and continuously optimize the network parameters through supervised learning. The method of the present invention utilizes the parallel structure and the ability of autonomous learning, can autonomously learn and gradually optimize the ability of the model, and analyze the feature laws in different signal samples.
[0039] The method disclosed in the present invention has stronger robustness in dealing with the recognition task of time-varying satellite signals in complex scenarios. It can learn the characteristics of input IQ signal samples under the condition of time-varying signal-to-noise ratio. Even in the face of challenges such as heavy noise interference or large signal strength fluctuations, it can still efficiently extract and utilize the manifold space characteristics of the signal to achieve signal reconstruction and accurate recognition of the signal type.
[0040] The method disclosed in the present invention gets rid of the reliance on signal feature expertise and signal modulation prior information in traditional methods. Driven by a large number of data sets, the method of the present invention utilizes parallel structures and autonomous learning capabilities to autonomously learn and gradually optimize the model's capabilities and analyze the characteristic laws in different signal samples. Without human intervention, the deep learning model can automatically reduce the dimensionality of signal data and mine more layers of more representative data features. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] Figure 1 It is a schematic diagram of the structure of the autoencoder;
[0042] Figure 2 It is a schematic diagram of the downsampling layer of the WCAE network;
[0043] Figure 3 It is a schematic diagram of the sampling layer on the WCAE network;
[0044] Figure 4 It is a schematic diagram of the signal feature extraction and signal reconstruction network;
[0045] Figure 5 is a schematic diagram of the signal type recognition network;
[0046] Figure 6 It is a schematic diagram of the learning rate decrease curve. DETAILED DESCRIPTION
[0047] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the following Figure 1-6 It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0048] Manifold space features refer to representative features extracted from data, which represent sample data in manifold space. Manifold space features can mine nonlinear relationships and intrinsic structures between samples, which helps improve the generalization ability and classification performance of classification models, thereby helping the model to classify and predict unknown data more accurately.
[0049] The method of the present invention is Figure 1Based on the shown Autoencoder (AE), a deep learning network architecture integrating a Wavelet Convolution Autoencoder (WCAE) was constructed, aiming to complete the tasks of satellite signal feature extraction and recognition. In the feature extraction stage, through convolution operations, the coherence of the signal between adjacent domains is maintained, and the time-frequency feature retention ability of the wavelet function is utilized, enabling this architecture to effectively extract signal features from the original IQ data, realize the mapping of satellite signal data from a high-dimensional sample space to a low-dimensional feature space, mine the manifold space features contained in the signal, and ensure the high-quality reconstruction of satellite signal data. In the fused feature stage, the manifold space features of different scales are fused after being processed by the fully connected layer, compensating for the feature loss caused by data dimensionality reduction. The fused features are output to the softmax layer to achieve end-to-end signal recognition, avoiding the process of human intervention, simplifying the signal recognition process, and improving the accuracy of signal recognition.
[0050] Since the autoencoder performs well in removing signal redundant information and extracting abstract features, and the fused wavelet convolution autoencoder adds a wavelet function and a convolution layer on the basis of the autoencoder to jointly form a wavelet hidden layer. WCAE combines the advantages of the traditional autoencoder's unsupervised learning, the good time-frequency localization of the wavelet function, and the convolution layer's ability to maintain the coherence of signal adjacent domain spatial information, enhancing the ability to extract manifold space features.
[0051] WCAE consists of two parts: an encoding network and a decoding network. The encoding network is composed of three downsampling blocks, and a Batch Normalization (BN) layer is added between adjacent downsampling blocks to accelerate the model convergence speed, improve the model stability, and reduce the dependence on parameter initialization. In addition, a certain proportion of Dropout is introduced to enhance the generalization ability of the model, reduce the mutual dependence between neurons, and avoid the occurrence of overfitting. The encoding network can map the satellite signal from a high-dimensional input sample to a low-dimensional abstract representation, realizing the compression and dimensionality reduction of the sample data. The decoding network is composed of three upsampling blocks, which are used to restore the manifold space features extracted by the encoding network to an output close to the original signal to realize the reconstruction of satellite signal samples.
[0052] The method of the present invention is divided into two paths: 1) extraction of the manifold space features of the original data stream of satellite signals and reconstruction of the data; 2) fusion of multi-scale manifold space features and classification recognition of satellite signals.
[0053] Embodiment 1, this embodiment discloses a satellite signal recognition method based on manifold space features, which directly receives the raw data of satellite signals at the input end, and uses a neural network instead of manual work to complete the task of feature extraction and signal recognition. Without human intervention, the neural network can automatically mine the manifold space features from the raw data, and can still ensure the stability of the features even under low signal-to-noise ratio conditions. This feature can help the neural network make effective recognition decisions, thereby realizing the recognition of satellite signals. In terms of feature extraction, by integrating wavelet transform and convolutional network into the feature extraction process of satellite signals, a deep autoencoder is constructed to achieve dimensionality reduction and reconstruction of signal samples, retaining the structural characteristics and time-frequency characteristics of signal samples. This process does not require human intervention or prior information of the signal, and significantly improves the efficiency of feature extraction. In terms of signal recognition, the optimization model parameter strategy adopts multi-scale feature fusion rather than traditional decision fusion, thereby compensating for the feature loss caused by data dimensionality reduction, ensuring the maximum contribution of different features in recognition decisions, and avoiding the previous inefficient method of simple feature accumulation. It includes the following steps:
[0054] Step 1, input the IQ sample data of the original satellite signal into the encoding network of the wavelet convolution encoder WCAE network;
[0055] Step 2, such as Figure 4 As shown in the figure, in the coding network, through layer-by-layer downsampling and batch normalization operations, the manifold space features in the sample data are extracted, the dimension of the signal data is reduced, the representation ability of the features is improved, and the time-frequency feature information in the signal is mined. Figure 2 As shown in Figure 1, the downsampling layer of the wavelet autoencoder consists of a convolutional layer, a discrete wavelet transform layer (DWT) and a ReLU activation layer.
[0056] The formula for extracting manifold space features is:
[0057]
[0058] In the above formula, S is the original IQ sample data, F i is the i-th layer manifold space feature, Conv is the convolution layer, DWT is the discrete wavelet transform layer, and ReLU is the activation layer;
[0059] Step 3: Pass the manifold spatial features extracted by the last set of downsampling layers to the decoding network, and reconstruct the low-resolution features into the original high-resolution information through continuous upsampling layers to achieve the reconstruction of the original satellite signal data; Figure 3As shown in the figure, the upsampling layer consists of an Inverse Discrete Wavelet Transform (IDWT) layer, a ConvTranspose layer, and a ReLU activation layer.
[0060] Step 4: Continuously optimize the WCAE network parameters using the loss function to ensure that the similarity between the reconstructed data and the original data reaches the optimal value, and guarantee the representativeness of the manifold space features and the quality of the reconstructed signal. The loss function selects the Binary Cross Entropy (BCE) function and the Consistency-Enhanced Loss (CEL) function, which are widely used in prediction and classification tasks:
[0061] Loss = L bce (p, g) + L cel (p, g)
[0062] During the training process, the BCE loss function calculates and accumulates the losses of each independent sample point, but ignores the structural relationship between the sample points. Its formula is as follows:
[0063] L bce = -∑[G(r, c)log(S(r, c)) + (1 - G(r, c))log(1 - S(r, c))]
[0064] In this embodiment, CEL is introduced to comprehensively consider the correlation and structural information between samples and enhance the model's ability to understand the signal data structure. Its formula is as follows:
[0065]
[0066] When the prediction result p is exactly the same as the original sample data g, the loss value is 0; when the two are completely different, the maximum penalty is given and the loss value is 1.
[0067] Step 5: After the training of the feature extraction and data reconstruction part is completed, retain the encoding network of the WCAE network, remove the decoding network, and use the multi-scale manifold space features extracted by the encoding network for subsequent signal recognition.
[0068] Step 6: Fusion of multi-scale manifold space features:
[0069] Fuse the multi-scale manifold space features F i extracted from different downsampling layers in the encoding network to obtain the fused feature F = [F 1 , F 2 ... F i ...];
[0070] Step 7: Input the fused manifold space features into the fully connected layer to map the manifold space features of different scales to the same dimension, eliminating the interference caused by dimensional differences to the fusion result;
[0071] Step 8: As Figure 5 shown, use a softmax classifier to classify the fused features, continuously optimize the network parameters through supervised learning, compare the output signal type with the actual label to ensure that the contribution degree of different scale features reaches the optimal during the process of optimizing the network parameters, and output the recognition result O k of this satellite signal, and its formula is as follows:
[0072]
[0073] In the above formula, u k represents the output feature of the last dense layer, and K represents the total number of signal types;
[0074] The output is written as:
[0075] O = [O 1 , O 2 ,..., O K T
[0076] Take the maximum value of O as the type of this signal.
[0077] The method of the present invention will be further described in detail below, including the following steps:
[0078] Step 1: Obtain the IQ data of satellite signal samples. For the acquisition method of this data, the data acquisition method can be flexibly adjusted according to actual needs and goals. For example, satellite signals with a set signal size and signal duration are collected at a certain sampling frequency, the data set is preprocessed, and the data set is divided into a training set and a test set. In this embodiment, the ratio of 8:2 is adopted for division to ensure that the model has sufficient data for training and improve the generalization ability and accuracy of the model;
[0079] Step 2: After completing the classification preprocessing of the data set, construct a satellite signal recognition model based on manifold space features. The model architecture consists of an encoding network, a decoding network, a feature fusion layer, and an identification output layer. The encoding network performs dimensionality reduction processing on the original IQ data to extract the manifold space features in the signal. The decoding network reconstructs the original signal using the manifold space features to verify the accuracy and robustness of the manifold space features. The feature fusion layer integrates the extracted multi-scale manifold space features, and the identification output layer realizes the recognition of satellite signals based on the fused comprehensive features;
[0080] Step 3: Input the training set into the model and train the feature extraction and reconstruction, and signal fusion and recognition networks in sequence. When the reconstruction effect of the signal reaches the optimal, remove the decoding network and only retain the encoding network;
[0081] Step 4: Use the multi-scale manifold space features extracted by different downsampling blocks of the encoding network for feature fusion to make up for the data loss caused by the signal during continuous downsampling. Through the fusion of multi-scale features, the model can maintain the multi-scale manifold space feature information with fewer parameters;
[0082] Step 5: Train the signal recognition network to ensure that the model can achieve the optimal recognition effect;
[0083] Step 6: Use the softmax classifier to classify the data set, input the data set into the softmax classifier, and accurately obtain the recognition result of the satellite signal.
[0084] In this example, the Stochastic Gradient Descent (SGD) algorithm is used to optimize and train the network parameters. The weight decay is set to 5e-4, the momentum is set to 0.9, the batch size is set to 8, and the epoch is set to 50. Through training on the training set, the recognition accuracy and efficiency of the model are improved;
[0085] In this example, the initial learning rate is set to 1e-3, and the "Poly" strategy is used to dynamically adjust the learning rate, so that the learning rate continuously decays during the iteration process of the model. The specific formula is as follows:
[0086]
[0087] In the above formula, lr represents the learning rate of the current epoch, base_lr is the initial learning rate, epoch is the current number of iterations, and num_epoch is the maximum number of iterations. As Figure 6 shown, Power controls the shape of the learning rate change curve and the decay speed of the learning rate. If power > 1, the decay speed of the learning rate accelerates, enabling the learning rate to quickly decrease from the initial value to zero, causing the model to converge rapidly in the early stage of training; if power < 1, the decay speed of the learning rate changes from slow to fast, helping the model to more carefully explore the parameter space and facilitating finding the local minimum or global minimum. In this example, power is set to 1.1;
[0088] Step 7: Use the trained network model architecture to test the test set and output the signal type recognition results of the test set.
[0089] It has been verified that the method proposed by the present invention solves the technical problems proposed by the present invention. The method of the present invention has been practically applied, verifying the practicability of the present invention and being able to achieve the technical problems proposed in the specification of the present application and reach the technical effects recorded in the specification of the present application.
[0090] The method of the present invention has been verified by simulation experiments and practical applications to have the technical effects claimed by the present invention.
[0091] The algorithm (method) proposed by the present invention is the underlying technical core of the present invention. Based on the algorithm, various products can be derived, such as developing corresponding software systems.
[0092] It should be understood that various forms of processes shown above can be used, reordering, adding or deleting steps. For example, the steps recorded in the present application can be executed in parallel, sequentially or in a different order, as long as the desired results of the technical solutions disclosed in the present application can be achieved, all within the protection scope of the present invention.
Claims
1. A satellite signal recognition method based on manifold spatial features, characterized in that: The steps include: Step 1, input the IQ sample data of the original satellite signal into the encoding network of the wavelet convolution encoder WCAE network; Step 2: In the encoding network, the manifold space features in the sample data are extracted through layer-by-layer downsampling and batch normalization operations: The formula for extracting manifold space features is: In the above formula, S is the original IQ sample data, F i is the i-th layer manifold space feature, Conv is the convolution layer, DWT is the discrete wavelet transform layer, and ReLU is the activation layer; Step 3: Pass the manifold spatial features extracted by the last set of downsampling layers to the decoding network, and reconstruct the original satellite signal data through continuous upsampling layers; Step 4: Use the loss function to continuously optimize the WCAE network parameters. The loss function selects the binary cross entropy function BCE and the consistent enhancement loss function CEL: Loss=L bce (p,g)+L cel (p,g) When the predicted result p is exactly the same as the original sample data g, the loss value is 0; when the two are completely different, the maximum penalty is given and the loss value is 1; Step 5: After the feature extraction and data reconstruction training is completed, the encoding network of the WCAE network is retained, the decoding network is removed, and the multi-scale manifold space features extracted by the encoding network are used for subsequent signal recognition; Step 6: Fusion of multi-scale manifold spatial features: The multi-scale manifold spatial features F extracted from different downsampling layers in the encoding network are i Perform feature fusion to obtain fusion feature F = [F1, F2...F i ...]; Step 7: Input the fused manifold space features into the fully connected layer to map the manifold space features of different scales to the same dimension; Step 8: Use the softmax classifier to classify the fused features, continuously optimize the network parameters through supervised learning, compare the output signal type with the actual label, ensure that the contribution of different scale features in the process of optimizing network parameters is optimal, and output the recognition result of the satellite signal. k , the formula is as follows: In the above formula, u k represents the output features of the last dense layer, and K represents the total number of signal types; The output is written as: O=[O1,O2,...,O K ] T The maximum value of O is the type of the signal.
2. The satellite signal recognition method based on manifold spatial features according to claim 1 is characterized in that: In step 2, the downsampling layer consists of a convolutional layer, a discrete wavelet transform layer, and a ReLU activation layer.
3. The satellite signal recognition method based on manifold spatial features according to claim 1 is characterized in that: In step 3, the upsampling layer consists of an inverse discrete wavelet transform layer, a transposed convolution layer, and a ReLU activation layer.
4. The satellite signal recognition method based on manifold spatial features according to claim 1 is characterized in that: In step 8, the initial learning rate is set to 1e-3, and the "Poly" strategy is used to dynamically adjust the learning rate. The specific formula is as follows: In the above formula, lr represents the learning rate of the current epoch, base_lr is the initial learning rate, epoch is the current number of iterations, num_epoch is the maximum number of iterations, and Power is set to 1.
1.
5. The satellite signal recognition method based on manifold spatial features according to claim 1 is characterized in that: In step 8, the stochastic gradient descent algorithm is used to optimize the network parameters for training, with weight decay set to 5e-4, momentum set to 0.9, training batch set to 8, and iteration set to 50.
6. A satellite signal recognition model based on manifold spatial features, characterized by: The satellite signal recognition model based on manifold space features is composed of a coding network, a decoding network, a feature fusion layer and a recognition output layer. The coding network performs dimensionality reduction processing on the original IQ sample data to extract the manifold space features in the data. The decoding network uses the manifold space features to reconstruct the original data to verify the accuracy and robustness of the manifold space features; The feature fusion layer integrates the extracted multi-scale manifold spatial features, and the recognition output layer recognizes satellite signals based on the fused comprehensive features.
7. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and the computer program is configured to implement the steps of the satellite signal identification method based on manifold spatial features according to any one of claims 1 to 5 when called by a processor.
8. A satellite signal recognition device based on manifold spatial features, the device comprising at least one processor and a memory in communication with the at least one processor, wherein: The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the above-mentioned satellite signal recognition method based on manifold spatial features to realize the recognition of satellite signals with manifold spatial features.