Satellite signal individual identification method based on multi-modal feature fusion and deep learning

Abstract

The application discloses a satellite signal individual identification method based on multi-modal feature fusion and deep learning, and the implementation steps are as follows: a module of time domain, frequency domain and statistical domain feature extraction branch combination is built, a global feature extraction module is built, a satellite individual identification network connected by the two modules is built, the network is trained by using a generated training set, and the satellite individual class of a to-be-identified downlink signal is identified. The application extracts signal features in the time domain, the frequency domain and the statistical domain, overcomes the problem that time-frequency information is not completely reserved, and enriches the signal features used for individual identification. By globally fusing the features extracted by different branches, the dependence on prior information is reduced, the long-term correlation in the features is enhanced, and the identification precision of the satellite individual identification network in a non-cooperative environment with incomplete information is improved.

Description

Technical Field

[0001] This invention belongs to the field of radar communication technology, and further relates to a satellite downlink signal individual identification method based on multimodal feature fusion and deep learning in the field of electronic countermeasures technology. This invention can be used to effectively identify individual satellite platforms receiving signals under conditions where parameters such as received signal modulation and communication methods are unknown, and the channel noise environment is unknown. Background Technology

[0002] Traditional satellite identification typically employs high-resolution radar detection technology, combined with orbital information, to confirm the identity of collaborating satellites. However, high-resolution radar equipment is generally expensive, has a limited deployment capacity, and its ability to track the ever-increasing number of space targets is also limited. On the other hand, by passively receiving downlink signals from satellites, such as telemetry and beacon signals, the electromagnetic characteristics of the platform carried by these signals can be captured for satellite identification. Furthermore, by combining Doppler frequency offset and orbital data, satellite tracking and anomaly behavior analysis can be achieved.

[0003] Hainan University proposed a method for identifying individual satellite communication radiation sources in a marine environment based on noise reduction and reconstruction in its patent application, "A Satellite Communication Signal Noise Reduction System and Method in a Marine Environment" (Patent Application No.: 202310891894.7, Publication No.: CN 116781181 A). The method's implementation steps are as follows: using two encoders to acquire the original image features of the acquired communication signal and the image features retaining equipment noise, respectively; processing and encoding the two types of features, then fusing them through an embedding layer; decoding and reconstructing the fused feature vector to achieve denoising of the satellite communication signal while retaining equipment noise, followed by individual satellite communication radiation source identification. This method can identify radiation sources for marine satellite communication equipment while significantly reducing signal noise, removing parts unrelated to the satellite communication signal, and reducing noise interference in the identification of individual radiation sources for marine satellite communication equipment. However, this method still has shortcomings: it requires specific information about environmental noise to be obtained before signal denoising can be completed. Using only simple statistical features such as the mean and standard deviation of signal data fails to effectively extract the time-frequency features of the signal. Furthermore, using only neural networks based on convolutional neural units causes the extracted features to lose long-term correlations with the signals, making it difficult to extend this method to applications beyond the ocean.

[0004] The University of Science and Technology of China (USTC) disclosed a method for identifying individual satellite communication radiation sources based on feature fusion in its patent application, "A Method for Individual Identification of Satellite Communication Radiation Sources Based on Phase Modulation Signal Feature Fusion" (Patent Application No.: 202310479126.0, Publication No.: CN 116431991 A). The method's implementation steps are as follows: receiving and demodulating communication data from a satellite transponder to obtain the original phase-modulated satellite communication signal dataset; extracting fingerprint feature mapping matrices with different mapping forms using wavelet decomposition, higher-order spectrum, Hausdorff dimension, and fractal box dimension algorithms; fusing and reducing the dimensionality of the extracted feature matrices using canonical correlation analysis (CCA) and kinematic association analysis (KPCA); and feeding the individual identification feature matrices extracted from different individuals into a classifier for training to achieve the identification and classification of individual satellite communication radiation sources. However, this method still has shortcomings, namely, the need to receive and demodulate communication data from the satellite transponder, making it difficult to handle situations where parameters such as the received signal modulation and communication mode are unknown. It mainly relies on traditional features extracted manually, such as wavelet decomposition and higher-order spectra, and is heavily dependent on expert experience. Moreover, these statistical features are only applicable to signal radiation sources that exist in the database and have standards for comparison.

[0005] In summary, due to the limited features extracted from the received signals, the time-frequency characteristics of the signals are not well preserved, and a large amount of prior information such as the environment and signal parameters is required, individual satellites cannot be accurately identified under non-cooperative conditions where prior information such as signal parameters is incomplete. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of the prior art by proposing a satellite downlink signal individual identification method based on multimodal feature fusion and deep learning. This method solves the problems of the lack of time-frequency features caused by the single feature extraction of received satellite downlink signals in the prior art, and the reduced adaptability due to the over-reliance on prior environmental and signal parameter information.

[0007] The specific approach to achieving the objective of this invention is as follows: This invention constructs a feature extraction module consisting of three parallel branches in the time domain, frequency domain, and statistical domain. This module extracts features from the time, frequency, and statistical domains of signal data, improving the diversity of feature extraction and making the signal information richer and more complete. This solves the problem of incomplete retention of signal time-frequency information in feature extraction under non-cooperative conditions in existing technologies. Furthermore, this invention constructs a global feature extraction LSTM module. This module can fuse local and global features of signals under different modalities. The LSTM used is a long-sequence analysis network model more suitable for signal analysis. The global feature extraction LSTM module can fully fuse the local features extracted from each branch, enhancing the long-term correlation in global features and improving the ability to suppress short-term noise. This solves the problem that existing technologies require a large amount of prior environmental information and have poor adaptability.

[0008] To achieve the above objectives, the technical solution adopted by the invention includes the following steps:

[0009] Step 1: Construct a feature extraction module consisting of three parallel branches: time-domain feature extraction, frequency-domain feature extraction, and statistical domain feature extraction.

[0010] Step 2: Construct a global feature extraction LSTM module consisting of a first LSTM layer, a second LSTM layer, a first fully connected layer, a second fully connected layer, and a Softmax layer connected in series, and set the parameters of the global feature extraction LSTM module.

[0011] Step 3: Construct a satellite individual recognition network consisting of a multimodal feature extraction module and a global feature extraction LSTM module connected in series;

[0012] Step 4: Train the satellite individual recognition network using the generated training set;

[0013] Step 5: Perform filtering and down-conversion preprocessing on the downlink signals of the satellites to be identified. Input the preprocessed signals to be identified into the trained satellite individual identification network and output the category of each identified satellite individual.

[0014] Compared with the prior art, the present invention has the following advantages:

[0015] First, this invention constructs a feature extraction module consisting of three parallel branches: time domain, frequency domain, and statistical domain. By extracting and fusing features from the time domain, frequency domain, and statistical domain of signal data, it overcomes the problem of incomplete retention of signal time-frequency information in existing technologies during feature extraction. By increasing the diversity of feature extraction, this invention improves the completeness of satellite signal feature extraction under non-cooperative conditions where prior information such as signal parameters is incomplete.

[0016] Secondly, this invention constructs a global feature extraction LSTM module, which enhances the long-term correlation of features by fusing local and global features of signals under different modalities. This overcomes the problem that existing technologies require a lot of prior environment and signal parameters. By making full use of the LSTM network, which is more suitable for signal analysis, the accuracy of satellite individual identification and the applicability of the proposed satellite individual identification method are improved. Attached Figure Description

[0017] Figure 1 This is a flowchart of the present invention;

[0018] Figure 2 This is a diagram of the satellite individual identification network structure of the present invention; wherein, Figure 2 (a) is a structural diagram of the residual convolution unit. Figure 2 (b) is a diagram of the overall structure of the satellite individual identification network;

[0019] Figure 3 These are simulation effect comparison diagrams of the present invention; wherein, Figure 3 (a) is a schematic diagram of the confusion matrix of individual satellite identification accuracy results. Figure 3 (b) Line graph comparing experimental results from different methods. Detailed Implementation

[0020] To illustrate the present invention more clearly, a further detailed description is provided below in conjunction with embodiments and accompanying drawings.

[0021] Reference Figure 1 The specific implementation steps of the embodiments of the present invention will be described in further detail below.

[0022] Step 1: Construct a feature extraction module consisting of three parallel branches: time domain feature extraction, frequency domain feature extraction, and statistical domain feature extraction, to extract multimodal features that are valuable for satellite identification.

[0023] The temporal feature extraction branch consists of three residual convolutional units connected in series. Each residual convolutional unit has the same structure, consisting of a group of convolutional layers and residual connections connected in parallel. The structure of the residual convolutional unit is as follows: Figure 2 As shown in (a). Figure 2 In (a), x represents the signal input to the residual convolutional unit, F(x) represents the convolutional layer group, H(x) represents the output of the residual convolutional unit, and CNN represents the convolutional layer. The residual connection skip structure of this invention allows the input signal of a unit to propagate directly to the output, which helps to solve the gradient vanishing problem in deep networks, enabling more efficient training of deeper networks, significantly improving the performance of deep neural networks, and thus enhancing the robustness of the model.

[0024] The convolutional layer group consists of a first convolutional layer, a first activation layer, a second convolutional layer, a second activation layer, a third convolutional layer, and a third activation layer connected in series. The convolutional kernels in each convolutional layer slide across the input data to acquire different features, simultaneously extracting both shallow and deep features. This invention uses this convolutional sliding method to optimize the number of weight parameters, thereby avoiding the high computational burden of existing fully connected networks. The number of convolutional kernels in the first to third convolutional layers is set to 2 each, the kernel sizes are set to 5, 1, and 3 respectively, and the convolutional strides are set to 4, 1, and 2 respectively. The ReLU function is used for all three activation layers.

[0025] The frequency domain feature extraction branch consists of a Fast Fourier Transform layer, a first residual convolutional unit, a second residual convolutional unit, and a third residual convolutional unit connected in series. Each residual convolutional unit has the same structure, consisting of a group of convolutional layers and residual connections connected in parallel. The group of convolutional layers consists of a first convolutional layer, a first activation layer, a second convolutional layer, a second activation layer, a third convolutional layer, and a third activation layer connected in series. The number of kernels in the first to third convolutional layers is set to 2, the kernel sizes are set to 5, 1, and 3 respectively, and the convolution stride is set to 4, 1, and 2 respectively. The ReLU function is used for all three activation layers. Compared to existing recurrent neural networks, the convolutional layer group of this invention can be computed in parallel, has excellent extraction capabilities for local features of data, and can more quickly map the original signal to a high-dimensional feature space.

[0026] The statistical domain feature extraction branch consists of a power spectrum feature extraction function layer, a higher-order cumulant feature extraction function layer, and a negative entropy feature extraction function layer connected in parallel. Due to the complex modulation and communication methods of the satellite downlink signals to be processed, the identification of individual satellites is extremely challenging. The statistical domain feature extraction branch employed in this invention has the advantage of requiring only a small amount of prior signal knowledge to extract the power spectrum, higher-order cumulants, and negative entropy features of the satellite downlink signal. These features contain rich characteristics of the signal and, compared to the signal's amplitude and frequency, possess stronger robustness and noise suppression capabilities.

[0027] The power spectrum feature extraction function layer in the statistical domain feature extraction branch outputs two statistical features. Since the ratio of the maximum to the second-largest power spectrum value varies significantly among different satellite individuals, the first statistical feature is set as the ratio of the maximum to the second-largest power spectrum value of the signal to be identified. Because the spectral sparsity differs among different satellite individuals within a given spectral line threshold, the second statistical feature is set as the number of spectral lines in the power spectrum of the signal to be identified that exceed the threshold.

[0028] The higher-order cumulant feature extraction function layer in the statistical domain feature extraction branch outputs five statistical features F3, F4, F5, F6, and F7 as follows:

[0029]

[0030]

[0031]

[0032]

[0033] F7 = log 10 (|C 80 |)

[0034] Among them, log10 (·) represents a base-10 logarithmic operation, |·| represents the absolute value operation, C 40 C represents the 4th-order cumulant characteristic under the 0th moment. 21 C represents the characteristic of the second-order cumulant under the first-order moment. 80 C represents the 8th-order cumulant characteristic under the 0th-order moment. 42 C represents the characteristic of the fourth-order cumulant under the second moment. 63 This represents the 6th-order cumulant feature under the 3rd-order moment. Since the cumulants above the 2nd order of Gaussian white noise are zero, the influence of noise on the signal can be ignored when using higher-order cumulant features to process noisy signals. Higher-order cumulant features also contain other statistical information about the signal, making them suitable as features to distinguish different individual signals.

[0035] The statistical feature F8 output by the negative entropy feature extraction function layer in the statistical domain feature extraction branch is as follows:

[0036]

[0037] Where log(·) denotes the logarithmic operation with the natural constant e as the base, δ 2 Let π represent the variance of the signal sequence input to the negative entropy feature extraction function layer, and e represent the natural constant. Since negative entropy is a non-linear statistical measure, it can be used to describe the non-Gaussianity of a signal, representing its complexity and randomness. Therefore, negative entropy can be used as a feature to distinguish different individual signals.

[0038] Step 2: Construct a global feature extraction LSTM module consisting of a first LSTM layer, a second LSTM layer, a first fully connected layer, a second fully connected layer, and a Softmax layer connected in series. Since the LSTM used in this invention is a type of efficient recurrent neural network specifically designed to capture long-term correlations in time-series data, it is well-suited for processing long-sequence inputs such as satellite downlink signals.

[0039] The parameters of the global feature extraction LSTM module are set as follows: the input dimension of the first LSTM layer is set to 2, and the hidden layer dimension is set to 5. The first LSTM layer is concatenated with the second LSTM layer, therefore the input dimension and hidden layer dimension of the second LSTM layer are both set to 5. The feature dimension input to the final Softmax layer needs to be consistent with the number of downlink signal categories for individual satellites in the training set; therefore, the number of fully connected neurons in the first and second fully connected layers is set to twice the number of signal categories and the number of categories, respectively.

[0040] Step 3: Construct a satellite individual recognition network consisting of a multimodal feature extraction module and a global feature extraction LSTM module connected in series, such as... Figure 2 As shown in (b).

[0041] Figure 2 (b) is a diagram of the overall structure of the satellite individual identification network. Figure 2 In (b), the CNN Block operation represents the residual convolutional unit mentioned above, the feature fusion operation represents the concatenation of the three sets of features extracted by the temporal feature extraction branch, the frequency domain feature extraction branch, and the statistical domain feature extraction branch into vectors, and the preprocessing operation represents the preprocessing of filtering and down-conversion. Figure 2 (b) The LSTM layer, by learning and memorizing the dynamic changes of the signal over time, can effectively integrate the outputs from different modal feature extraction branches into a comprehensive feature representation. Simultaneously, it reduces noise caused by random fluctuations in time-series data and maintains the integrity of long-term correlation information in the signal sequence. This structure of the present invention enhances the ability of individual recognition networks to process complex signals, ensuring that the extracted features contain complete time-frequency information of the signal, thus enhancing the robustness of the feature representation. Therefore, the present invention significantly enhances the system's adaptability to changes in time-series signals and its robustness in different scenarios.

[0042] Step 4: Train the satellite individual identification network using the generated training set.

[0043] The training set consists of downlink telemetry signals from 18 satellites operating in the S-band, with each signal category containing at least 200 downlink IQ signals from the same satellite. Each satellite's downlink IQ signal is filtered and down-converted to 70MHz. Then, k data segments are sampled at a sampling rate of 56Mbps, with each segment containing 16384 points. Each preprocessed IQ signal is labeled with a satellite category. All preprocessed signals and their corresponding labels form the training set. The first m segments of data are used for model training and adjustment, while the last (km) segments are used as the test set.

[0044] The training of the satellite individual identification network involves inputting the training set into the network. The time-domain IQ signal data used in this embodiment includes both In-phase (I) and Quadrature-phase (Q) branches, comprehensively representing the signal's all-around information. The frequency-domain Fast Fourier Transform sequence used in this invention reveals the signal's modulation characteristics and frequency composition. The power spectrum, higher-order cumulants, and negative entropy features in the statistical domain, primarily analyzed through statistical methods, also provide rich information. The loss between the predicted and true categories output by the satellite individual identification network is calculated. Gradient descent is used to iteratively update the network parameters until the network's loss function converges, resulting in the trained satellite individual identification network.

[0045] The loss function of the network is as follows:

[0046]

[0047] Where Loss represents the cross-entropy loss function, N represents the batch size in training, i represents the index of the training sample in the training batch, M represents the total number of downlink signal categories for individual satellites in the training set, c represents the index of the training sample in the downlink signal category for an individual satellite, and y ic Let y denote the sign function, where y represents the true class of the i-th sample if c is true. ic The value of p is 1 if it is not 1 otherwise it is 0. ic This represents the probability that the predicted category of the i-th sample is c after passing through the satellite individual identification network.

[0048] Because the cross-entropy loss function used in this invention is directly related to the probability output, it facilitates gradient optimization, the loss function is numerically stable, and intuitively represents the uncertainty of prediction well. This enables the model to effectively learn and improve its performance on classification tasks.

[0049] Step 5: Perform filtering and down-conversion preprocessing on the downlink signals of the satellites to be identified. Input the preprocessed signals to be identified into the trained satellite individual identification network and output the category of each identified satellite individual.

[0050] The effects of this invention will be further illustrated below with simulation experiments:

[0051] 1. Simulation experimental conditions:

[0052] The hardware platform for the simulation experiment of this invention is as follows: the processor is an i9-11900F CPU with a main frequency of 2.5GHz, the memory is 16GB, and the graphics card is an NVIDIA GeForce RTX 3090 GPU.

[0053] The software platform for the simulation experiment of this invention is: Windows 10 operating system, Python 3.8 and PyTorch 0.12.0.

[0054] The input data used in the simulation experiment of this invention is the downlink IQ signal of the S-band satellite. Each satellite downlink IQ signal is filtered and downconverted to 70MHz, and then sampled at a sampling rate of 56Mbps. After preprocessing, the signal sequence size in each sample data is 16384×2. Each preprocessed IQ signal is labeled with the category of individual satellites, which includes 18 categories of individual satellites. The data format is h5py.

[0055] 2. Simulation content and result analysis:

[0056] The simulation experiment of this invention uses this invention and three existing technologies (ResNet network recognition method, convolutional neural network classification method, and LSTM network classification method) to classify the preprocessed satellite downlink IQ signal, and the classification results are shown in the figure below. Figure 3 As shown.

[0057] The three existing technologies used in the simulation experiment are:

[0058] The existing ResNet network recognition method refers to the ResNet-based classification and recognition network method proposed by He, Kaiming et al. in "Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 770-778), 2016", and is referred to as the ResNet network recognition method.

[0059] The existing convolutional neural network recognition method refers to the classification and recognition method based on convolutional neural networks proposed by Alex Krizhevsk et al. in "Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems. 2012; 25.", which is referred to as the convolutional neural network recognition method.

[0060] The existing LSTM network identification method refers to the radiation source identification method based on LSTM proposed by Liu Kuoran in "LSTM-based radar radiation source identification technology [J]. Ship Electronic Engineering, 2019, 39(12):92-95.", which is abbreviated as LSTM network identification method.

[0061] The following is combined with Figure 3 The simulation diagrams further illustrate the effects of the present invention.

[0062] Figure 3 (a) A schematic diagram of the confusion matrix drawn from the simulation results using the method of the present invention. Figure 3 In (a), the columns of the confusion matrix represent the true categories, while the rows represent the categories predicted by the satellite individual identification network. The value in each cell represents the proportion of samples predicted as the true category of that column out of the total number of samples in the experiment. Figure 3In (a), the gray bars on the right side of the confusion matrix show the correspondence between cell values ​​and color depth. Darker colors correspond to values ​​closer to 1, while lighter colors correspond to values ​​closer to 0. The primary use of the confusion matrix is ​​to evaluate the performance of classification networks. Higher values ​​on the diagonal indicate better accuracy. Meanwhile, values ​​off-diagonal reveal which categories the classification network is prone to confusion with.

[0063] Figure 3 In (a), the values ​​on the diagonal of the confusion matrix are all close to or equal to 1, indicating that most satellite signals were correctly classified into their individual categories. Most cells outside the diagonal have values ​​close to or equal to 0. The presence of values ​​close to 0 in cells outside the diagonal indicates that some signals were misclassified. At the same time, the relatively small number of cells outside the diagonal with non-zero values ​​indicates that the number of misclassifications is relatively small.

[0064] Figure 3 (b) Line graphs showing the simulation results of the method of the present invention and three prior art techniques. Figure 3 (b) The horizontal axis of the line graph represents the number of different satellite individual categories, and the vertical axis represents the average identification accuracy of each category of individuals. Figure 3 (b) The circled lines are drawn from the simulation results of the method of the present invention, the lines with squares are drawn from the simulation results of the ResNet network recognition method of the prior art, the lines with triangles are drawn from the simulation results of the convolutional neural network recognition method of the prior art, and the lines with crosses are drawn from the simulation results of the LSTM network recognition method of the prior art. Figure 3 (b) The performance of the proposed satellite individual identification network is compared with existing technologies such as ResNet network identification method, convolutional neural network identification method and LSTM network identification method on 18 different satellite signal identification tasks.

[0065] from Figure 3As shown in (b), the accuracy of the proposed satellite individual identification network fluctuates between 0.85 and 1.00, demonstrating relatively stable and efficient identification capabilities. The accuracy of the existing ResNet network identification method fluctuates significantly across different signals, sometimes similar to the proposed method, and sometimes much lower. The accuracy of existing convolutional neural network identification methods is generally lower than that of the proposed method and the ResNet network identification method. The accuracy of existing LSTM network identification methods fluctuates the most, showing high accuracy for some signals but a significant decrease in accuracy for others. Therefore, the proposed satellite individual identification network exhibits high accuracy for most satellite signals, and the accuracy for different individual categories does not vary significantly, demonstrating its superior performance and stability in this task. While other existing network models also show high accuracy for individual signals, their overall accuracy is lower than that of the proposed network. This proves that the classification effect of the present invention is superior to the first three existing identification methods, and the identification effect is more ideal.

[0066] To quantitatively evaluate the simulation effect of this invention, the recognition results of the four methods are evaluated using the average individual recognition accuracy described below. The average individual recognition accuracy is calculated using the following formula:

[0067]

[0068] The calculation results of the average individual identification accuracy of the satellite individual identification network proposed in this invention and three existing technologies are presented in Table 1:

[0069] Table 1. Quantitative Analysis of the Classification Results of the Invention and Prior Art in Simulation Experiments

[0070] Satellite Individual Identification Network 92.68％ ResNet network recognition method 89.95％ Convolutional Neural Network Recognition Method 82.11％ LSTM network identification method 74.68％

[0071] As can be seen from Table 1, the average individual identification accuracy of the present invention is 92.68%, which is higher than that of the three existing technical methods, proving that the present invention can achieve higher satellite individual identification accuracy.

[0072] The simulation experiments above demonstrate that this invention, utilizing the constructed modal feature extraction modules in the time domain, frequency domain, and statistical domain, can extract complete time-frequency features of satellite downlink signals and combine them with statistical features. Furthermore, the constructed global feature extraction module performs global fusion of the features from these modalities and further extracts features for long-term correlation information in individual identification. This solves the problems in existing methods, such as the lack of signal time-frequency features due to the extraction of only single features from received satellite downlink signals, and the reduced adaptability and low accuracy caused by the over-reliance on prior environmental and signal parameter information in individual identification methods. Therefore, this invention represents a highly practical satellite individual identification method.

Claims

1. A satellite signal individual identification method based on multimodal feature fusion and deep learning, characterized in that, A feature extraction module with three parallel branches in the time domain, frequency domain, and statistical domain is constructed to extract complete time-frequency information. A global feature extraction LSTM module is constructed to fuse features from various modalities. The steps of this recognition method are as follows: Step 1: Construct a feature extraction module consisting of three parallel branches: time-domain feature extraction, frequency-domain feature extraction, and statistical domain feature extraction. The statistical domain feature extraction branch is composed of a power spectrum feature extraction function layer, a higher-order cumulant feature extraction function layer, and a negative entropy feature extraction function layer connected in parallel. The power spectrum feature extraction function layer in the statistical domain feature extraction branch outputs two statistical features: the first statistical feature is the ratio of the maximum value of the power spectrum of the signal to be identified to the second largest value of the power spectrum; the second statistical feature is the number of spectral lines in the power spectrum of the signal to be identified that are above the threshold. Step 2: Construct a global feature extraction LSTM module consisting of a first LSTM layer, a second LSTM layer, a first fully connected layer, a second fully connected layer, and a Softmax layer connected in series, and set the parameters of the global feature extraction LSTM module. Step 3: Construct a satellite individual recognition network consisting of a multimodal feature extraction module and a global feature extraction LSTM module connected in series; Step 4: Train the satellite individual recognition network using the generated training set; Step 5: Perform filtering and down-conversion preprocessing on the downlink signals of the satellites to be identified. Input the preprocessed signals to be identified into the trained satellite individual identification network and output the category of each identified satellite individual.

2. The satellite signal individual identification method based on multimodal feature fusion and deep learning according to claim 1, characterized in that, The temporal feature extraction branch described in step 1 consists of three residual convolutional units connected in series. Each residual convolutional unit has the same structure, consisting of a group of convolutional layers and residual connections connected in parallel. The group of convolutional layers consists of a first convolutional layer, a first activation layer, a second convolutional layer, a second activation layer, a third convolutional layer, and a third activation layer connected in series. The number of convolutional kernels in the first to third convolutional layers is set to 2, the kernel sizes are set to 5, 1, and 3 respectively, and the convolutional strides are set to 4, 1, and 2 respectively. The ReLU function is used for the first to third activation layers.

3. The satellite signal individual identification method based on multimodal feature fusion and deep learning according to claim 1, characterized in that, The frequency domain feature extraction branch described in step 1 consists of a Fast Fourier Transform layer, a first residual convolutional unit, a second residual convolutional unit, and a third residual convolutional unit connected in series. Each residual convolutional unit has the same structure, consisting of a group of convolutional layers and residual connections connected in parallel. The group of convolutional layers consists of a first convolutional layer, a first activation layer, a second convolutional layer, a second activation layer, a third convolutional layer, and a third activation layer connected in series. The number of convolutional kernels in the first to third convolutional layers is set to 2, the kernel sizes are set to 5, 1, and 3 respectively, and the convolution stride is set to 4, 1, and 2 respectively. The ReLU function is used for the first to third activation layers.

4. The satellite signal individual identification method based on multimodal feature fusion and deep learning according to claim 3, characterized in that, The higher-order cumulant feature extraction function layer in the statistical domain feature extraction branch outputs five statistical features. , , , , as follows: ， ， ， ， ， in, This represents a logarithmic operation with base 10. This indicates the absolute value operation. This represents the 4th-order cumulant characteristic under the 0th moment. This represents the characteristic of the second-order cumulant under the first-order moment. This represents the 8th-order cumulant characteristic under the 0th moment. This represents the characteristic of the fourth-order cumulant under the second moment. This represents the 6th cumulant characteristic under the 3rd moment.

5. The satellite signal individual identification method based on multimodal feature fusion and deep learning according to claim 4, characterized in that, The statistical features output by the negative entropy feature extraction function layer in the statistical domain feature extraction branch. as follows: ， in, Represented by natural constant Logarithmic operations with base 0. The sequence variance of the satellite downlink signal. Represents pi (π). Represents the natural constant.

6. The satellite signal individual identification method based on multimodal feature fusion and deep learning according to claim 1, characterized in that, In step 2, the parameters of the global feature extraction LSTM module are set as follows: the input dimension of the first LSTM layer is set to 2, the hidden layer dimension is set to 5, the input dimension and hidden layer dimension of the second LSTM layer are both set to 5, and the number of fully connected neurons in the first and second fully connected layers is set to twice the number of downlink signal categories of individual satellites in the training set and the number of categories, respectively.

7. The satellite signal individual identification method based on multimodal feature fusion and deep learning according to claim 1, characterized in that, The training set mentioned in step 4 is as follows: Select downlink signals containing at least 15 types of satellite individuals, with each type of signal containing at least 200 downlink IQ signals of the same satellite individual. Perform filtering and downconversion preprocessing on each satellite downlink IQ signal in sequence, label the satellite individual category for each preprocessed IQ signal, and form a training set by combining all preprocessed signals and their corresponding labels.

8. The satellite signal individual identification method based on multimodal feature fusion and deep learning according to claim 1, characterized in that, The training of the satellite individual identification network described in step 4 involves inputting the training set into the satellite individual identification network, calculating the loss between the predicted category and the true category of the network output, and iteratively updating the network parameters using the gradient descent method until the network's loss function converges, thereby obtaining the trained satellite individual identification network. The loss function of the network is as follows: ， in, This represents the cross-entropy loss function, where N represents the batch size during training. This indicates the sequence number of the training sample in the training batch, and M represents the total number of downlink signal categories for individual satellites in the training set. This indicates the sequence number of the training sample within the downlink signal category of the individual satellite. Represents a symbolic function, if the first The true class of each sample is ,but Its value is 1, otherwise its value is 0. Indicates the first The predicted category output by the satellite individual recognition network after processing each sample is: The probability of.

Images