A low-orbit satellite signal modulation recognition method and device based on radio frequency learning

Abstract

This invention provides a method and apparatus for low-Earth orbit (LEO) satellite signal modulation identification based on radio frequency (RF) learning. The method includes: calculating element values ​​based on the RF signal and constructing a three-dimensional spectral array; extracting frequency domain features, time-frequency features, and amplitude-phase features from the three-dimensional spectral array; determining the signal features for generating an RF semantic map by calculating the cross-entropy value of each feature; obtaining a frequency domain feature RF semantic map based on the frequency domain features, a time-frequency feature semantic map based on the time-frequency features, and an amplitude-phase feature RF semantic map based on the amplitude-phase features; constructing a LEO satellite RF semantic map identification and classification model; inputting the feature semantic map corresponding to the maximum cross-entropy value into the LEO satellite RF semantic map identification and classification model; classifying and identifying the feature semantic map; and outputting the modulation identification result. This invention can improve the detection accuracy of signal modulation methods in high-dynamic, high-bandwidth LEO satellite communication scenarios.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of signal recognition, and in particular to a low-orbit satellite signal modulation recognition method and device based on radio frequency learning. BACKGROUND

[0002] Signal modulation recognition is a key technical point in the field of signal sensing detection in cognitive radio technology. Signal sensing detection achieves rapid and sensitive response to business needs by obtaining important parameter information of radio frequency signals, and then makes targeted analysis and decision. Existing signal modulation recognition technologies are divided into two categories: one is a modulation recognition method based on feature extraction, and the other is a modulation recognition method based on statistical likelihood features. The former needs to pay attention to the specific features extracted, such as high-order cumulant features, cyclic statistical features, wavelet features, and signal spectrum correlation features. Only the features with strong pertinence and good quality can make the recognition accuracy meet the requirements. The latter relies on complete channel state prior information and is sensitive to model parameter estimation effect, and needs to calculate a large amount of prior data to support the algorithm performance. It has poor robustness to channels with non-ideal characteristics such as frequency offset, phase offset and time-varying fading. The two existing modulation recognition technologies are based on signal level data features, and the recognition accuracy and time overhead are highly dependent on the feature type and algorithm power. Since signal sensing detection is essentially blind detection of radio frequency received signals, signal level feature analysis needs to be performed on the unprocessed radio frequency received data. With the expansion of the communication bandwidth of low-orbit satellite communication scenes and the application of frequency polychromatic multiplexing, the difficulty of signal level analysis and processing of radio frequency received signals increases sharply with the complexity of radio frequency received data. Therefore, the signal level analysis and processing method is not suitable for recognizing the signal modulation method of the large bandwidth high dynamic low-orbit satellite communication scene. SUMMARY

[0003] In view of this, the embodiments of the present application provide a low-orbit satellite signal modulation recognition method and device based on radio frequency learning, to solve the problem that the signal level modulation recognition method of the prior art is not suitable for recognizing the low-orbit satellite signal modulation method.

[0004] One aspect of the present application provides a low-orbit satellite signal modulation recognition method based on radio frequency learning, which comprises the following steps:

[0005] The method comprises the following steps: acquiring radio frequency signals of a plurality of satellite terminal user points, performing discrete Fourier transform on the radio frequency signals of each satellite terminal user point to obtain the frequency spectrum of each satellite terminal user point at each time point; evolving and stacking the frequency spectrum of each satellite terminal user point at each time point into a planar graph, and performing image gray extraction processing on each element value in the planar graph to obtain a two-dimensional visual image; performing statistical analysis on the element values in each two-dimensional visual image and then performing sharpening and hiding operations, and stacking the two-dimensional visual images according to the serial numbers of the satellite terminal user points to obtain a three-dimensional spectral array graph; projecting the three-dimensional spectral array graph onto a two-dimensional plane in the time-frequency domain to obtain a two-dimensional visual frequency spectrum graph of the superposition of the satellite terminal user points, sampling the element values in the two-dimensional visual frequency spectrum graph, and dividing the sampled element values into a plurality of subsequences; sampling the two-dimensional visual frequency spectrum graph according to the positions of the elements in each subsequence to obtain a plurality of sub-images; calculating the Laplace operator of each sampled element value in each sub-image and combining the Laplace operators to obtain a Laplace matrix of all the sampled element values, wherein the Laplace operator in the Laplace matrix is used to represent the frequency domain feature of the radio frequency signal; cutting the three-dimensional spectral array graph along the element value gradient absolute value direction from the full time domain and the full frequency domain, calculating the influence values of a plurality of adjacent element values on any element value on the cutting surface, and performing noise interpolation, rendering the noise values and the element values according to the numerical size into a heat map to represent the time-frequency feature of the radio frequency signal; calculating the amplitude and phase of the radio frequency signal received by each satellite terminal user point in the full time domain, and arranging and mapping the amplitudes and phases according to the serial numbers of the satellite terminal user points to obtain a contour-like three-dimensional stereogram to represent the amplitude and phase feature of the radio frequency signal; calculating the cross-entropy values of the frequency domain feature, the time-frequency feature and the amplitude and phase feature respectively, and generating a radio frequency semantic graph based on the feature with the maximum cross-entropy value; acquiring a pre-trained low-orbit satellite radio frequency semantic graph recognition and classification model, wherein the low-orbit satellite radio frequency semantic graph recognition and classification model comprises an input layer, a convolution layer, a pooling layer, a full connection layer and an activation function connected in sequence, inputting the radio frequency semantic graph into the pre-trained low-orbit satellite radio frequency semantic graph recognition and classification model to obtain an output modulation method recognition result; wherein generating the feature radio frequency semantic graph comprises mapping the element values in the sub-image and the total gain of the element value change trend into the chroma and brightness of the semantic graph elements according to a set proportion to obtain a frequency domain feature radio frequency semantic graph; calculating the average received power and the received signal-to-noise ratio of the radio frequency signal according to priori, comparing the average received power and the received signal-to-noise ratio with a set value to determine whether the radio frequency signal is error-coded, and constructing a time-frequency feature radio frequency semantic graph; filling the element values in the contour-like three-dimensional stereogram into the semantic graph according to the edge-center principle from the edge to the center to obtain an amplitude and phase feature radio frequency semantic graph; the edge-center principle refers to that the element values gradually decrease from the edge to the center in the amplitude and phase feature radio frequency semantic graph, and the color of the element values gradually deepens from the edge to the center in the amplitude and phase feature radio frequency semantic graph.

[0006] In some embodiments, the sharpening and hiding operation is performed after statistical analysis of the elements in each two-dimensional visualized picture, including,

[0007] calculating the statistical variance of the element values of all two-dimensional visualized pictures of each satellite terminal user point in the same time domain and frequency domain, removing the element values with a deviation greater than a statistical variance threshold, and then calculating the statistical variance of the remaining element values until the statistical variance of the remaining element values is less than the statistical variance threshold, and calculating the statistical average of the remaining element values, and filling the statistical average to the position of the removed element values, if the number of times of calculating the statistical variance of the element values exceeds a set number of times and the statistical variance of the remaining element values is still greater than the statistical variance threshold, then setting all element values in the two-dimensional visualized picture to zero.

[0008] In some embodiments, the two-dimensional visualized spectrum is sampled according to the positions of the elements in each sub-sequence to obtain a plurality of sub-images, and further comprising, performing a graph Fourier transform on each sub-sequence corresponding to a sub-image to obtain a transformed sequence of each sub-sequence,

[0009] The calculation formula of the transformed sequence is:

[0010] ;

[0011] ;

[0012] wherein, represents the transformed sequence, represents the total amount of sampled element values, represents the sub-sequence, represents the frequency, represents the number of sub-sequences, and u represents the element value in the sub-sequence.

[0013] In some embodiments, the Laplacian operator of each sampled element value on each sub-image is calculated and combined to obtain a Laplacian matrix of all sampled element values, including, the calculation formula of the Laplacian operator of each sampled element is:

[0014] ;

[0015] ;

[0016] wherein, represents the Laplacian operator, and m represents the sequence number of the sampled element value, represents the edge weight, represents any sampled element value on the sub-image, This represents the eight sampled element values ​​adjacent to any sampled element value on the subimage, where i represents positive and negative integers from -4 to 4.

[0017] In some embodiments, the three-dimensional spectral array is cut along the direction of the absolute value of the element numerical gradient from the full time domain and the full frequency domain. The influence of multiple neighboring element values ​​on any element value on the cut surface is calculated and noise interpolation is performed. The noise values ​​and each element value are rendered into a heatmap according to their numerical magnitude to characterize the time-frequency characteristics of the radio frequency signal, including...

[0018] At a fixed frequency, the position of the largest element value on each two-dimensional visualization image and the eight element values ​​adjacent to the largest element value are found. The gradient between the largest element value and the eight adjacent element values ​​is calculated, and the image is cut along the direction of the maximum gradient value to obtain the first cutting surface of the three-dimensional spectral array. At a fixed time, the position of the largest element value on each two-dimensional visualization image and the eight element values ​​adjacent to the largest element value are found. The gradient between the largest element value and the eight adjacent element values ​​is calculated, and the image is cut along the direction of the maximum gradient value to obtain the second cutting surface of the three-dimensional spectral array.

[0019] In some embodiments, the three-dimensional spectral array is cut along the direction of the absolute value of the element numerical gradient from the full time domain and the full frequency domain; the influence of multiple neighboring element values ​​on any element value on the cut surface is calculated and noise interpolation is performed; and the noise values ​​and each element value are rendered into a heatmap according to their numerical magnitude to characterize the time-frequency characteristics of the radio frequency signal. The method also includes...

[0020] The gradient vector and distance vector between any element value on the first cutting surface and the next eight adjacent element values ​​are calculated respectively. The gradient vector is multiplied by the corresponding distance vector and accumulated to obtain the influence value of the eight adjacent element values ​​on the element value. The influence value is interpolated with noise value using a nonlinear interpolation function to make the surfaces of the first cutting surface and the next cutting surface smoother. The noise value and each element value on the three-dimensional spectral array are rendered into a heat map according to their numerical values ​​to represent the time-frequency characteristics.

[0021] In some embodiments, the amplitude and phase of the radio frequency signal received by each satellite terminal user point in the full time domain are calculated, and a contour-like three-dimensional map is obtained by arranging and mapping the data according to the serial number of each satellite terminal user point to characterize the amplitude and phase features of the radio frequency signal, including...

[0022] The magnitude of the real and imaginary parts of the radio frequency signal received by each satellite terminal user point in the full time domain and the arctangent trigonometric function of the real and imaginary parts are calculated to obtain the signal amplitude and phase of the radio frequency signal of each satellite terminal user point. The signal amplitude and amplitude of each satellite terminal user point are plotted as a contour map with contour lines representing the phase and fill color representing the amplitude. The contour maps of each satellite terminal user point are sorted and mapped in sequence according to the serial number of each satellite terminal user point to obtain a three-dimensional contour map of the radio frequency signal.

[0023] In some embodiments, the average received power and received signal-to-dryness ratio of the radio frequency signal are calculated a priori, and the average received power and the received signal-to-dryness ratio are compared with a set value to determine whether the radio frequency signal has bit errors, and a time-frequency characteristic radio frequency semantic map is constructed, including,

[0024] The average received power and received signal-to-dryness ratio of the radio frequency signal are calculated based on the broadband channel model, interference fluctuations, and random noise priors. The received signal-to-dryness ratio is compared with a set value to determine whether each frequency point of the communication bandwidth is occupied by interference. If it is occupied by interference, the signal at that frequency point is a bit error signal; if it is not occupied by interference, the signal at that frequency point is a bit error-free signal. Different colors are assigned to bit error signals and bit error-free signals, and the time-frequency feature radio frequency semantic map is constructed.

[0025] On the other hand, the present invention also provides an electronic device including a processor and a memory, wherein the memory stores computer instructions, and the processor is used to execute the computer instructions stored in the memory, wherein when the computer instructions are executed by the processor, the device implements the steps of the above method.

[0026] On the other hand, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0027] The beneficial effects of the present invention are at least as follows:

[0028] The low-Earth orbit (LEO) satellite signal modulation recognition method and apparatus based on radio frequency (RF) learning of this invention calculates element values ​​from the RF signal and constructs a three-dimensional spectral array. Frequency domain features, time-frequency features, and amplitude-phase features are extracted from the 3D spectral array. The signal features for generating an RF semantic map are determined by calculating the cross-entropy values ​​of each feature. A frequency domain feature RF semantic map is obtained based on the frequency domain features, a time-frequency feature semantic map based on the time-frequency features, and an amplitude-phase feature RF semantic map based on the amplitude-phase features. A LEO satellite RF semantic map recognition and classification model is constructed. The feature semantic map corresponding to the maximum cross-entropy value is input into the LEO satellite RF semantic map recognition and classification model for classification and recognition, outputting the modulation recognition result. This transforms the processing of RF signals into image processing, improving the accuracy of signal modulation method recognition in LEO satellite communication scenarios.

[0029] Furthermore, a decision scheme for sharpening and concealing element values ​​under simultaneous frequency domain conditions was designed to conceal invalid element values ​​due to interference and sharpen valid element values, thereby eliminating the influence of channel noise interference and improving the accuracy of signal modulation recognition.

[0030] Furthermore, different extraction methods were used to extract different dominant features of the radio frequency signal in the three-dimensional spectral array, which improved the accuracy of signal feature extraction.

[0031] Furthermore, different methods are used to construct radio frequency semantic maps for different explicit features, so that the actual physical meaning represented by different explicit features is different. This makes it easier for the low-orbit satellite radio frequency semantic map identification and classification model to identify the modulation method of the signal in the radio frequency semantic map, thereby improving the identification accuracy.

[0032] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the description, or may be learned by practice of the invention. The objects and other advantages of the invention can be realized and obtained by means of the structures specifically pointed out in the description and drawings.

[0033] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0034] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, are not intended to limit the scope of the invention. In the drawings:

[0035] Figure 1 This is a flowchart of a low-orbit satellite signal modulation identification method based on radio frequency learning according to an embodiment of the present invention.

[0036] Figure 2 This is an embodiment of the low-orbit satellite signal modulation identification method based on radio frequency learning according to one of the present invention. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0038] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0039] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0040] It should also be noted that, unless otherwise specified, the term "connection" in this article can refer not only to a direct connection, but also to an indirect connection involving an intermediary.

[0041] In the following description, embodiments of the invention will be illustrated with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts, or the same or similar steps.

[0042] The ultimate goal of sixth-generation mobile communication networks is to achieve the Internet of Everything, a fully connected world integrating terrestrial mobile wireless communication and satellite communication, achieving seamless global coverage by integrating satellite communication into 6G mobile communication. In this application scenario, low-Earth orbit (LEO) satellite communication will enable broadband internet access to support high-speed, ultra-low-latency service demands. Given the unique characteristics of LEO satellite communication compared to terrestrial mobile communication, its performance heavily depends on channel characteristics. Accurate knowledge of the space-time-frequency non-stationary channel response and Doppler frequency offset, along with targeted channel equalization and frequency synchronization, will significantly improve the adaptability of LEO satellite internet to channel degradation. By performing feature learning on the RF received signal and based on the learning results and the acquired space-time-frequency non-stationary channel response, channel estimation equalization and frequency offset compensation operations can be pre-deployed for the LEO satellite internet to ensure its communication performance requirements. Therefore, feature learning of the RF received signal is essential to ensuring that the communication network performance meets the demands of efficient services. The RF received signal in the LEO satellite communication scenario will be severely affected by the characteristics of space-time-frequency non-stationary channels. The received radio frequency (RF) signal is submerged in interference noise due to power loss, resulting in an excessively low received signal level and thus subtle overall RF spectrum features, which is detrimental to feature learning. The relative motion between the LEO satellite transmitter and the receiving mobile terminal causes phase and frequency shifts in the received RF signal, leading to irregular jitter of various spectrogram elements in the RF spectrum, which is also unfavorable for signal feature extraction. Due to the fast time-varying characteristics of LEO satellite channels, the received channel characteristics of the RF signal will vary significantly at different times and for different durations. Different channel characteristics will also have different effects on the received RF signal. For example, a harsh electromagnetic environment will cause sudden changes in signal intermodulation, and the relative motion between the transmitter and receiver will cause signal frequency shifts.

[0043] Existing modulation identification technologies fall into two main categories: feature extraction-based modulation identification methods and statistical likelihood-based modulation identification methods. The former requires attention to specific extracted features, such as higher-order cumulant features, cyclic statistics features, wavelet features, and signal spectrum correlation features. Only highly targeted and high-quality features can achieve the required recognition accuracy. The latter relies on complete prior information about the channel state and is sensitive to the effectiveness of model parameter estimation. It requires a large amount of computational prior data to support algorithm performance, thus exhibiting poor robustness to channels with non-ideal characteristics such as frequency shift, phase shift, and time-varying fading. Both types of modulation identification technologies are based on signal-level data features, and their recognition accuracy and time overhead are highly dependent on the feature type and algorithm computational power. Since signal sensing and detection is essentially blind detection of RF received signals, signal-level feature analysis of unprocessed RF received data is required. With the expansion of communication bandwidth and the application of frequency multicolor multiplexing in low-Earth orbit satellite communication scenarios, the difficulty of signal-level analysis and processing of RF received signals increases sharply with the complexity of RF received data. Therefore, the applicability of signal-level analysis and processing methods to high-bandwidth, high-dynamic low-Earth orbit satellite communication scenarios is greatly reduced. Therefore, applying deep learning to modulation recognition is a new research direction in the field of signal sensing and detection. Most research institutions' radio frequency (RF) learning for image-level analysis and processing is generally based on single-indicator images with limited information, such as spectrograms, constellation diagrams, and signal envelope diagrams. If the number of modulation types to be identified remains constant, the smaller the amount of spectrogram information, the greater the difficulty in developing subsequent deep learning algorithms. It is evident that existing deep learning modulation recognition methods have limited spectrogram information and cannot accurately identify signal modulation methods in high-bandwidth, high-dynamic low-Earth orbit (LEO) satellite communication scenarios. Therefore, this invention provides a method and apparatus for LEO satellite signal modulation recognition based on RF learning to address the inapplicability of existing signal modulation recognition methods to RF signal modulation method identification in high-bandwidth, high-dynamic LEO satellite communication scenarios.

[0044] One aspect of the present invention provides a method for low-Earth orbit satellite signal modulation identification based on radio frequency learning, such as... Figure 2 As shown, the method includes steps S101 to S104:

[0045] S101: The radio frequency signal is a modulated radio wave with a certain transmission frequency. The radio frequency signals of multiple satellite terminal user points are acquired. The discrete Fourier transform of the radio frequency signal of each satellite terminal user point is performed to obtain the spectrum of each satellite terminal user point at each time point. The spectrum of each satellite terminal user at each time point is evolved along time and stacked into a planar spectrum. The grayscale value of each element in the planar spectrum is processed to obtain a two-dimensional visualization image. After statistical analysis of the element values ​​in each two-dimensional visualization image, a sharpening and hiding operation is performed. The images are then stacked into a three-dimensional spectral array according to the sequence number of each satellite terminal user point. S102: Project the 3D spectral array onto the 2D plane containing the time and frequency domain to obtain a 2D visualized spectrum map superimposed from each satellite terminal user point. Sample the element values ​​in the 2D visualized spectrum map and divide the sampled element values ​​into multiple subsequences. Sample the 2D visualized spectrum map according to the position of the elements in each subsequence to obtain multiple sub-images. Calculate the Laplace operator for each sampled element value in each sub-image and combine them to obtain the Laplace matrix of all sampled element values. The Laplace operator in the Laplace matrix is ​​used to characterize the frequency domain characteristics of the radio frequency signal. Perform a full-time and full-frequency analysis of the 3D spectral array along the elements... The absolute value of the prime gradient is used for cutting. The influence of multiple neighboring element values ​​on any element value on the cut surface is calculated and noise interpolation is performed. The noise value and each element value are rendered into a heatmap according to their numerical magnitude to characterize the time-frequency characteristics of the radio frequency signal. The amplitude and phase of the radio frequency signal received by each satellite terminal user point in the full time domain are calculated and mapped according to the serial number of each satellite terminal user point to obtain a contour-like three-dimensional map to characterize the amplitude and phase characteristics of the radio frequency signal. The cross-entropy values ​​of the frequency domain characteristics, time-frequency characteristics, and amplitude and phase characteristics are calculated respectively, and the feature with the largest cross-entropy value is used to generate the radio frequency semantic map. S103: Obtain the pre-trained low-Earth orbit satellite radio frequency semantic map recognition and classification model. The low-Earth orbit satellite radio frequency semantic map recognition and classification model includes an input layer, convolutional layer, pooling layer, fully connected layer and activation function connected in sequence. The radio frequency semantic map is input into the pre-trained low-Earth orbit satellite radio frequency semantic map recognition and classification model to obtain the output modulation method recognition result. S104: The generation of the feature radio frequency semantic map includes mapping the element values ​​and the total gain of the element value change trend in the sub-image to the chromaticity and brightness of the semantic map elements according to a set ratio to obtain the frequency domain feature radio frequency semantic map; calculating the average received power and received signal-to-dryness ratio of the radio frequency signal according to prior calculation, and comparing the average received power and received signal-to-dryness ratio with the set value to determine whether the radio frequency signal has a bit error, and constructing the time-frequency feature radio frequency semantic map; filling the element values ​​in the contour-like three-dimensional solid map from the edge to the center in the semantic map according to the edge-center principle to obtain the amplitude-phase feature radio frequency semantic map; the edge-center principle means that the element values ​​gradually decrease from the edge to the center in the amplitude-phase feature radio frequency semantic map, and the color of the element values ​​gradually becomes darker from the edge to the center in the amplitude-phase feature radio frequency semantic map.

[0046] In step S101, each satellite terminal user receives a radio frequency (RF) signal for a set duration. Discrete Fourier transforms (DFTs) are performed on the RF signal at each time point. The result of a single DFT can be considered as a "line" representing the spectrum, i.e., the element value, at that time point. The set duration contains multiple time points. The spectra of multiple consecutive time points are stacked and combined according to their temporal evolution to form a "plane," i.e., a planar spectrum. Each planar spectrum is used to represent the time-frequency domain information characteristics of the RF signal under the corresponding satellite terminal user. Image grayscale extraction can enhance two-dimensional visualization images and simultaneously reduce their dimensionality, thus reducing computational load. The satellite terminal user point refers to the geographical location of the ground user terminal.

[0047] In some embodiments, after performing statistical analysis on the elements in each two-dimensional visualization image, a sharpening and concealment operation is performed, including...

[0048] The statistical variance of element values ​​in all two-dimensional visualization images of each satellite terminal user point in the same time and frequency domains is calculated. Element values ​​with deviations greater than a statistical variance threshold are removed, and the statistical variance of the remaining element values ​​is calculated until it is less than the threshold. The statistical average of the remaining element values ​​is then calculated and filled into the positions of the removed element values, achieving element value sharpening. If the number of times the statistical variance of element values ​​is calculated exceeds a set number and the statistical variance of the remaining element values ​​is still greater than the statistical variance threshold, all element values ​​in the two-dimensional visualization image are set to zero, and the corresponding colors are hidden. Element values ​​invalidated by noise interference in the two-dimensional visualization image are hidden, while valid element values ​​are sharpened, thereby eliminating channel noise interference. The element values ​​in two-dimensional visualization images of different satellite terminal user points at the same time and frequency are sequentially sharpened and hidden, and image grayscale extraction is performed to obtain a three-dimensional spectrogram.

[0049] In other embodiments, the statistical variance is calculated for element values ​​in a two-dimensional visualization image as follows:

[0050] ;

[0051] ;

[0052] in, Indicates statistical variance. Indicates the number of satellite terminal user points. This represents all element values ​​at a fixed frequency and time, where i represents time and j represents frequency. It represents the average value of all element values ​​on a plane perpendicular to the two-dimensional visualization image under fixed frequency and fixed time conditions, where k represents the satellite terminal user point.

[0053] Remove deviation Greater than the current statistical variance element value Then continue to calculate the statistical variance of the remaining element values. Remove deviation degree Greater than the current statistical variance element value Repeat the above operation until the variance of the remaining element values ​​is calculated. Less than the preset statistical variance threshold This indicates that the remaining element values ​​can accurately represent the element values ​​under this time-frequency domain condition. The statistical average of the remaining element values ​​is calculated using the following formula:

[0054] ;

[0055] ;

[0056] The statistical average is used as the element value under this time-frequency domain condition, and this set of element values ​​is considered valid. The statistical average is then used to fill in the positions of the removed element values. The above describes the element-value sharpening operation; if we calculate the number of statistical variances... More than the preset limit number of times There is still a degree of deviation. Greater than the preset statistical variance threshold If the data is invalid, it indicates that the original data used to calculate the value of this group of elements was interfered with to varying degrees. Therefore, the value of this group of elements is considered invalid, and zeros are filled in the position of the value of this group of elements and the color is hidden.

[0057] In some embodiments, image grayscale extraction is performed on the element values ​​in the planar atlas to obtain a two-dimensional visualization image, including the calculation formula for image grayscale extraction:

[0058] ;

[0059] ;

[0060] ;

[0061] In this representation, RGB represents three channels: red, green, and blue, i represents time, and j represents frequency. Indicates the start time. Indicates the end time. Indicates the starting frequency. This indicates the ending frequency. Image grayscale extraction can convert the magnitude of element values ​​into grayscale images of varying shades.

[0062] In step S102, the frequency domain features of the radio frequency signal are extracted by combining graph fast Fourier transform with graph Laplace matrix. That is, the three-dimensional spectral array is projected onto the two-dimensional plane containing the time and frequency domain to obtain a two-dimensional visualized spectrum of each satellite terminal user point superimposed. The two-dimensional visualized spectrum is sampled from dark to light according to the gray color. The three-dimensional spectral array is projected onto the two-dimensional plane. The element values ​​on the three-dimensional spectral array will also be superimposed. That is, the obtained sampled element values ​​may be obtained by superimposing multiple element values. Therefore, it is necessary to use graph fast Fourier transform to calculate the element values ​​on the sub-image for subsequent calculation of the Laplace operator of the sampled element values ​​and characterization of the frequency domain features of the radio frequency signal. A three-dimensional thermal texture mapping method is applied to extract the time-frequency characteristics of the received radio frequency signal. The three-dimensional spectral array is cut along the direction of maximum element value gradient from both the full time and full frequency domains, resulting in two types of cut surfaces. One type involves finding the position of the maximum element value in each two-dimensional visualization image at a fixed frequency, along with the eight adjacent element values, calculating the gradient between the maximum and eight adjacent element values, and cutting along the direction of the maximum gradient value. This yields the first cut surface of the three-dimensional spectral array. The other type involves finding the position of the maximum element value in each two-dimensional visualization image at a fixed time, along with the eight adjacent element values, calculating the gradient between the maximum and eight adjacent element values, and cutting along the direction of the maximum gradient value. This yields the second cut surface of the three-dimensional spectral array. Both the first and second cut surfaces are uneven, composed of multiple discrete element values, with the height of the cut surface representing the magnitude of the element values. Since different radio frequency signal modulation methods have different characteristics, such as the significant frequency domain characteristics of FSK signals, the significant time-frequency characteristics of QAM signals, and the significant amplitude and phase characteristics of APSK signals, after obtaining the characteristics of the radio frequency signals, the target decision function value is used to determine which feature to generate the radio frequency semantic map.

[0063] In some embodiments, sampling the two-dimensional visualized spectrogram according to the position of elements in each sub-sequence to obtain multiple sub-images, further includes performing graph fast Fourier transform on the sub-sequences corresponding to each sub-image to obtain the transform sequence of each sub-sequence.

[0064] The formula for calculating the transformed sequence is:

[0065] ;

[0066] ;

[0067] in, Represents the transformation sequence. This represents the total number of sampled element values. Represents a subsequence. Indicates frequency, The number of subsequences is represented by , and u represents the element value in the subsequence. Performing a graphical Fast Fourier Transform on the subsequences corresponding to a subimage can characterize the frequency domain features of the subimage, reduce noise, identify hidden features of the signal, and obtain the element values ​​in the subimage.

[0068] In some embodiments, the Laplacian operator for each sampled element value in each sub-image is calculated and combined to obtain the Laplacian matrix of all sampled element values, including the following formula for calculating the Laplacian operator for each sampled element value:

[0069] ;

[0070] ;

[0071] in, This represents the Laplace operator, where m represents the ordinal number of the sampled element value. Represents edge weight, Represents any sampled element value on the sub-image. Let represent the eight sampled element values ​​adjacent to any given sampled element value in the subimage, where i represents a positive or negative integer from -4 to 4. Each sampled element value in the subimage is adjacent to eight other sampled element values, and the edge weight is the square of the difference between any two adjacent sampled element values. The Laplacian operator for all sampled element values ​​is calculated to obtain the Laplacian matrix. The Laplacian operator in the Laplacian matrix intuitively represents the total gain obtained after a small perturbation of a given element value and the change trend of the element value in the frequency domain.

[0072] In other embodiments, the number of sample element values ​​is an integer power of 2. To meet the requirements of the Fast Fourier Transform, the extracted sampled element values ​​are also divided into integer powers of 2. If there are n subsequences, then each subsequence can be represented as:

[0073]

[0074]

[0075]

[0076]

[0077] in, arrive Represents a subsequence. arrive This represents the value of the sampled element; one subsequence corresponds to one subimage.

[0078] In some embodiments, the three-dimensional spectral array is cut along the direction of the absolute value of the element numerical gradient from the full time domain and the full frequency domain; the influence of multiple neighboring element values ​​on any element value on the cut surface is calculated and noise interpolation is performed; the noise value and each element value are rendered into a heatmap according to their numerical magnitude to characterize the time-frequency characteristics of the radio frequency signal; the method also includes...

[0079] The gradient vector and distance vector between any element value and its eight neighboring element values ​​are calculated on the first and second cutting surfaces, respectively. The gradient vector is multiplied by the corresponding distance vector and summed to obtain the influence value of the eight neighboring element values ​​on that element value. A nonlinear interpolation function is used to interpolate the influence value to reduce noise, making the first and second cutting surfaces smoother. The noise value and each element value on the 3D spectral array are rendered into a heatmap according to their numerical magnitude to represent time-frequency characteristics. For each element value on the first and second cutting surfaces, which is adjacent to eight element values, the gradient vector and distance vector of the eight element values ​​are calculated, and their dot product is calculated. The sign of the dot product indicates the sign of the influence value of the eight element values ​​on that element value. A Berlin noise function is used to interpolate the influence value to make the first and second cutting surfaces smoother.

[0080] In some embodiments, the amplitude and phase of the radio frequency signal received by each satellite terminal user point in the full time domain are calculated, and a contour-like three-dimensional map is obtained by arranging and mapping the data according to the serial number of each satellite terminal user point to characterize the amplitude and phase features of the radio frequency signal, including...

[0081] The magnitude of the imaginary and real parts of the radio frequency (RF) signal received by each satellite terminal user point in the full time domain, along with the arctangent trigonometric function of the imaginary and real parts, are calculated to obtain the signal amplitude and phase of the RF signal at each satellite terminal user point. The signal amplitude and phase of each satellite terminal user point are then plotted as a contour map, with contour lines representing phase and fill colors representing amplitude. These contour maps are then sorted and mapped sequentially according to the satellite terminal user point's index to obtain a contour-like 3D model of the RF signal. In this contour-like 3D model, higher elevations represent larger amplitudes, the number of elevations represents the number of amplitude magnitudes of the received RF signal, and the number of discrete closed contour lines represents the number of phase magnitudes of the received RF signal.

[0082] In some embodiments, the cross-entropy values ​​of frequency domain features, time-frequency features, and amplitude-phase features are calculated respectively, and the feature corresponding to the maximum cross-entropy value is used to generate a radio frequency semantic map, including,

[0083] The formula for calculating cross-entropy is:

[0084] ;

[0085] in, This represents the cross-entropy value, and N represents the number of experiments performed. This indicates the selection weight for each radio frequency semantic map type. This represents the value of each spectral element in the radio frequency spectrum. This represents the probability of selecting the radio frequency semantic map. The objective decision function value is calculated using cross-entropy, and the feature with the largest cross-entropy value is used to generate the radio frequency semantic map.

[0086] In step S103, in the pre-trained low-Earth orbit satellite radio frequency semantic map recognition and classification model, the input layer receives various generated radio frequency semantic maps; the convolutional layer is designed to include the actual physical meaning represented by the spectrogram elements in various radio frequency semantic maps; the pooling layer is designed with a max pooling scheme to solve the problems of repetition and partial invalidity of the physical meaning represented by the spectrogram elements after convolution calculation and to achieve effective substitution; the fully connected layer is designed to reorganize all processed semantic map elements into a one-dimensional data sequence to meet the requirement that graph convolutional neural networks can only recognize and classify one-dimensional data sequences; the activation function is designed as follows: A function is used to ensure that the backpropagation gradient is never zero during the iterative optimization of the classification weight parameter matrix. The weight matrix is ​​continuously optimized iteratively until the loss function meets the requirements. Finally, the parameters and structure of the pre-set low-Earth orbit (LEO) satellite radio frequency semantic map recognition and classification model are determined. The radio frequency semantic map is input into the pre-trained LEO satellite radio frequency semantic map recognition and classification model, and the output is the probability of obtaining a certain modulation method of the radio frequency signal. The method corresponding to the highest probability is the modulation method of the received radio frequency signal.

[0087] In some embodiments, when training the low-Earth orbit (LEO) satellite radio frequency (RF) semantic map recognition and classification model, a large number of frequency domain feature RF semantic maps, amplitude and phase feature RF semantic maps, and time-frequency feature RF semantic maps are used to train the LEO satellite RF semantic map recognition and classification model, outputting the modulation method of the RF signal. After training to obtain a high-performance satellite RF semantic map recognition and classification model, the frequency domain feature RF semantic maps, amplitude and phase feature RF semantic maps, and time-frequency feature RF semantic maps are used as labels for the LEO satellite RF semantic map recognition and classification model to identify the modulation method of the signal to which the subsequently input RF semantic map belongs.

[0088] In step S104, the type of the generated radio frequency semantic map is determined based on the cross-entropy value. If the feature with the largest cross-entropy value is a frequency domain feature, then according to... (Simplified Information Aggregation) Based on the stability property of the rate of change of the continuous dynamic change law and the principle of consistency of change direction, the element values ​​and the total gain of the element value change trend in the sub-image are mapped to the chromaticity and brightness of the semantic map elements according to a set ratio relative to the overall data set, thus obtaining the frequency domain feature radio frequency semantic map. If the time-frequency feature has the largest cross-entropy value, the average received power and received signal-to-dryness ratio of the radio frequency signal are calculated according to the broadband channel model, interference fluctuations, and random noise priors. The received signal-to-dryness ratio is compared with the set value to determine whether each frequency point of the communication bandwidth is occupied by interference. If it is occupied by interference, the signal at that frequency point is a bit error signal; if it is not occupied by interference, the signal at that frequency point is a bit error-free signal. Different colors are assigned to bit error signals and bit error-free signals, and a time-frequency feature radio frequency semantic map is constructed. If the amplitude-phase feature has the largest cross-entropy value, the corresponding amplitude-phase values ​​are filled into the corresponding positions in the semantic map according to the edge-center expansion principle based on the mapping relationship between the number of "altitude" and the number of closed "altitude horizontal lines" in the three-dimensional contour model, thus obtaining the amplitude-phase feature radio frequency semantic map. A contour-based 3D stereogram is a mountain-like structure with concentric rings rising upwards, from larger bottom rings to smaller top rings. Rings with smaller height but larger radii represent smaller phases and larger amplitudes. The phase value within a ring is fixed. Height represents the phase value, and radius represents the amplitude value. For generating amplitude-phase feature radio frequency semantic maps, rings with larger radii are on the outer edge, representing larger amplitudes but smaller phases; rings with smaller radii are closer to the interior, representing smaller amplitudes but larger phases. The amplitude-phase feature radio frequency semantic map is two-dimensional, but each location has two variables: the magnitude of the element value and the intensity of its color. The element values ​​in the amplitude-phase feature radio frequency semantic map gradually decrease from the edge to the center, representing a decrease in amplitude; the color intensity of the map elements gradually increases from the edge to the center, representing an increase in phase. In the frequency domain characteristic RF semantic map, each element represents the chromaticity and brightness of a pixel. In the time-frequency characteristic RF semantic map, each element represents whether interference occupies 1 bit of channel information, i.e., whether the RF signal at each frequency point has bit errors. The position and color of each element in the amplitude-phase characteristic RF semantic map represent the phase and amplitude of the RF signal, respectively. Among them, the 1 bit of information represents the information that should be transmitted normally at each frequency point.

[0089] Furthermore, the formula for calculating the average received power is:

[0090]

[0091] The method for calculating the signal-to-interference-plus-noise ratio (SINR) at the receiver is as follows:

[0092]

[0093]

[0094]

[0095]

[0096] in, Indicates the transmission power. Indicates the satellite transmitting antenna gain. Indicates path loss. This indicates the weakening of clouds and rain. Indicates atmospheric decay, Indicates the flickering effect. Indicates the fading of the shadow. This represents power loss (cable loss, device loss, antenna loss, etc.) that is independent of the fading type in the propagation path. Indicates the antenna gain of the receiving terminal. Indicates the gain of the airborne antenna. Indicates signal wavelength , Indicates the distance between the satellite transmitter and the ground receiver. Power, gain, and loss are all in units of 1. . Indicates the received signal dryness ratio (SINR). Indicates the received signal power. This represents the length of the two sides of the time-frequency feature radio frequency semantic map.

[0097] The following description uses a specific embodiment as an example.

[0098] This invention provides a method for low-Earth orbit satellite signal modulation identification based on radio frequency learning, such as... Figure 1 As shown, including,

[0099] Step 1: Generation of Spatiotemporal 3D Spectral Array

[0100] The main approach is as follows: a three-dimensional model is constructed based on the constraints of the coordinate axis variables of the three-dimensional spectral array; a decision scheme for sharpening and concealing spectrogram elements of multiple users under the simultaneous frequency domain conditions in the low-orbit satellite communication scenario is designed, and the decision value of the spectrogram element is adaptively adjusted according to the calculation results each time, and the decision results are superimposed sequentially; finally, a spatiotemporal three-dimensional spectral array diagram is generated based on the three-dimensional model and the decision criteria of the spectrogram element.

[0101] (1) Construct a three-dimensional solid model based on the constraints of each coordinate variable of the three-dimensional spectral array: the three coordinate axis variables are of the time type. ,frequency Satellite terminal user points First, modeling is performed at the same satellite terminal user locations. The time-frequency domain model at that location receives data for a fixed period of time. To process the received data at a specific point in time. ,once The result of the calculation can be visualized as a "line" representing the spectrum at this point in time, and multiple consecutive time points... The evolution of received data DFT results over time Stacked and combined to form "a plane," simultaneously expressing time-frequency domain information features. In this two-dimensional plane, the horizontal axis represents the frequency domain, the vertical axis represents the time domain, and the element values ​​are... Calculation results. Convert the values ​​of each element. By image A two-dimensional visualization image is drawn using color conversion methods. The color conversion method is as follows:

[0102]

[0103] ;

[0104] Then, plot the different satellite terminal user points under the same time-frequency domain conditions. The two-dimensional visualization image shows the user locations arranged by number in ascending order. By overlaying numerous two-dimensional visualizations, the two-dimensional plane is extended into a three-dimensional space, where each numerical element is... Finally, each DFT data point in each three-dimensional space is represented using different gray levels based on its size.

[0105] (2) Design of sharpening and concealment of three-dimensional spectrogram elements for multiple users Judgment Scheme: In a two-dimensional visualized image, the depth of the spectral color at each element's location represents different... The calculation results will be used to visualize the various two-dimensional images of different satellite terminal user locations under the same time-frequency domain conditions. Statistical analysis of the calculation results is beneficial for this group. The data is then sharpened and hidden. First, a set of user serial numbers is obtained. The statistical variance of the DFT data set is calculated using the following method:

[0106] ;

[0107] Remove deviation Greater than the current statistical variance DFT values Then continue to calculate the statistical variance of the remaining DFT data. Remove deviation degree Greater than the current statistical variance DFT values Repeat the above steps until the variance of the remaining data is calculated. Less than the preset statistical variance threshold This indicates that the remaining data can accurately characterize the DFT operation results under this time-frequency domain condition, and the remaining data can be calculated. The statistical average of the data is calculated as follows:

[0108]

[0109] ;

[0110] If the result is used as the DFT calculation result under this time-frequency domain condition, then the set of DFT data is considered valid, and the statistical average is used to fill in the positions of the removed data. The above describes the numerical sharpening operation; if we calculate the number of statistical variances... More than the preset limit number of times There is still a degree of deviation. Greater than the preset statistical variance threshold The data indicates that the original data used to calculate this set of DFT data was subject to varying degrees of interference, suggesting that this set of data... Invalid data in this group Data locations are filled with zeros and then hidden by color to reduce interference.

[0111] Step 2: Extract semantic information from explicit random feature maps

[0112] The main approach is as follows: design an explicit random feature extraction scheme, apply a combination of graph fast Fourier transform and Laplace matrix to extract the frequency domain characteristics of the radio frequency signal of the low-Earth orbit satellite receiving terminal, apply a three-dimensional thermal texture mapping method to extract the time-frequency-bit characteristics of the radio frequency signal of the low-Earth orbit satellite receiving terminal, apply a three-dimensional contour mapping method to extract the amplitude and phase characteristics of the radio frequency signal of the low-Earth orbit satellite receiving terminal, and finally determine the semantic information used for the generation of radio frequency semantic map based on the target decision function.

[0113] (1) Extracting signal frequency domain features using a combination of graph fast Fourier transform and graph Laplace matrix: Project the spatiotemporal three-dimensional spectral array generated in step one onto the time-frequency plane to obtain a two-dimensional visualized spectrum of each user point superimposed. Sample the spectrum obtained by projection according to the color of the spectral elements from dark to light, and the number of samples is... an integer power of 2 To meet the requirements of the Fast Fourier Transform, the extracted sample data sequence will also be used. Divide into integer powers of 2 Subsequences:

[0114]

[0115]

[0116]

[0117]

[0118] Each subsequence Mapped into a two-dimensional visualized spectrogram, it becomes a part of the overall image. Then, the original image needs to be segmented according to the image parts corresponding to the subsequence to obtain sub-images, and the subsequences corresponding to the sub-images need to be subjected to discrete Fourier transform.

[0119] The calculation formula is:

[0120]

[0121] Therefore, we obtain Subsequences Transform the sequence, and then... indivual Transformed sequences combined into sample data Transform the sequence. Simultaneously, calculate the Laplacian operator at each sampled data point in the sub-image to obtain the graphical Laplacian matrix. In the two-dimensional visualization plane obtained by projection and sampling, each sampled data point is adjacent to 8 other sampled data points. The difference between two adjacent sampled data points is... The square of the edge weight is defined as the edge weight. The weighted degree of this sample data is the sum of the edge weights of its eight adjacent sample data. Because there are two-dimensional visualization planes... If there are 1 sample data, then define the adjacency matrix. for The matrix, matrix elements The edge weights are the edge weights of two adjacent sample data points. Weighted degree matrix. If the main diagonal elements are the weighted values ​​of each sample data point, and the other elements are 0, then the Laplace operator for this sample data is:

[0122] ;

[0123] ;

[0124]

[0125] in The adjacency matrix is ​​the first Okay, from the above reasoning, for all sampled data, the Laplace matrix is:

[0126]

[0127]

[0128] The Laplace operator in the Laplace matrix intuitively represents the total gain obtained after a small perturbation of a certain spectral element and the change trend of the frequency domain DFT value.

[0129] (2) Extracting the time-frequency characteristics of the image signal using the three-dimensional thermal texture mapping method: The spatiotemporal three-dimensional spectral array generated in step one is processed from both the full time domain and the full frequency domain according to the direction of the absolute value of the numerical gradient of the three-dimensional spectral array elements. Cut it.

[0130] ;

[0131] ;

[0132]

[0133]

[0134] ;

[0135] in, This is the time-frequency diagram corresponding to the k-th user terminal. With a fixed frequency, find the maximum DFT value at all times within that frequency. This is a numerical search in the time domain. The above operation is repeated for the time-frequency plots at all users, which is represented as an uneven cut in the 3D plot. This is the time-frequency graph for the k-th user terminal. With a fixed time frame, find the maximum DFT value at all frequencies at that given moment. This search is performed in the frequency domain. This operation is performed on the time-frequency plots of all users, resulting in a textured, uneven surface in the 3D plot. Based on the first type of cutting surface mentioned above Any DFT value in the cut plane is adjacent to eight DFT data points. Calculate the maximum gradient between the DFT value and the eight adjacent data points. This corresponds to the second type of cutting diagram.

[0136] This results in two types of uneven cut surfaces. The elevation of each point on the cut surface represents the magnitude of the DFT value. To significantly characterize the difference in elevation of the cut surface, a three-dimensional thermal texture mapping is implemented. A Berlin noise function is defined in the three-dimensional spectral array space. The cut surface is composed of several discrete DFT data points with unequal elevations. Any one of these points... Each data point is adjacent to eight other data points. Calculate the gradient vector of this DFT data point with respect to its eight adjacent data points. With distance vector Take the dot product of the corresponding gradient vector and the distance vector. Then we get the data from the eight adjacent data points. The influence value of the data is represented by the sign of the dot product, and then the influence values ​​of 8 adjacent data points are interpolated using nonlinear interpolation. function Noise interpolation To make the cut surface smoother, and finally compare the noise value with the original The numerical values ​​are rendered into a heatmap according to their magnitude, and the color distribution in the heatmap represents the time-frequency characteristics of the signal.

[0137] (3) Extracting the amplitude and phase features of the radio frequency received signal using the three-dimensional contour mapping method: The low-orbit satellite receiving terminal user point The received radio frequency signal data in the entire time domain is moduloed by the virtual and real parts. Real and imaginary parts of the arctangent trigonometric function Calculate the amplitude and phase of the data set, then draw a contour-like plot using contour lines to represent the phase and fill colors to represent the amplitude. Similarly, calculate the amplitude for all user points. Full-time domain radio frequency received signal amplitude and phase Feature-type contour maps, these feature-type contour maps are sorted by user number. The sorting mapping yields a 3D contour map, showing the elevation. Higher represents amplitude The larger the number of discrete "altitudes", the higher the number of altitudes. Represents the number of radio frequency received signal amplitudes Number of discrete closed contour lines Represents the number of phases of the radio frequency received signal Thus, the amplitude and phase characteristics of the radio frequency received signal are obtained by mapping from the contour-equivalent three-dimensional solid image.

[0138] (4) Determining the semantic information used for generating the radio frequency semantic map using the objective decision function: Different modulation signals have different dominant random features. Therefore, the most significant and suitable feature among the three features mentioned above needs to be selected as the semantic information for generating the radio frequency semantic map through the objective decision function. The objective decision function value is calculated using cross-entropy, and the selection weights of various dominant features for different modulation methods are determined. The difference lies in the selection probability of the corresponding dominant feature for a certain modulation scheme of radio frequency received signal data. They are also different. The following is the cross-entropy. Calculation formula:

[0139] ;

[0140] in, This represents the cross-entropy value, and N represents the number of experiments performed. This indicates the selection weight for each radio frequency semantic map type. This represents the value of each spectral element in the radio frequency spectrum. This indicates the probability of selecting a radio frequency semantic map. When the cross-entropy value corresponding to a certain dominant feature is greater than the cross-entropy values ​​of other dominant features, that dominant feature is selected to generate the radio frequency semantic map.

[0141] Step 3: Design a radio frequency semantic information graph construction scheme

[0142] The main approach is as follows: determine the actual physical meaning represented by each element in the low-Earth orbit satellite internet radio frequency semantic information map; each element in the semantic map matrix extracted from frequency domain characteristics represents the chromaticity and brightness of a pixel; each element in the semantic map matrix extracted from time-frequency features represents whether interference occupies 1 bit of channel information; and the position and color of each element in the semantic map matrix extracted from amplitude and phase features respectively represent the phase and amplitude of the image signal; based on the actual meaning represented by the semantic map elements and prior model knowledge (periodicity, RSS dynamic change law, consistency of change direction, etc.), apply the map construction method to construct the low-Earth orbit satellite internet radio frequency semantic information map.

[0143] (1) Frequency domain feature radio frequency semantic map construction scheme: In step two, the method of combining graph FFT and graph Laplacian matrix is ​​used to intuitively and accurately reflect the spectral change trend of each part of the entire image, based on (Simplified Information Aggregation) Continuous "push" of related "preference" information based on the continuous dynamic change pattern.

[0144]

[0145] and the stability property of the rate of change based on the principle of consistency of direction of change ( (Critical value of rate of change):

[0146]

[0147] sequentially on the sub-images Numerical and Total gain of numerical change situation The chromaticity and luminance of the semantic map elements are mapped to their respective proportions relative to the overall dataset, thus obtaining the frequency domain characteristic radio frequency semantic map.

[0148] (2) Time-Frequency Feature RF Semantic Map Construction Scheme: The three-dimensional heatmap obtained by applying the three-dimensional heatmap mapping method in step two accurately characterizes the time-frequency characteristics of the RF received signal based on the prior broadband channel model. Interference fluctuations such as adjacent channel interference, intermodulation interference, and spurious interference; random noise such as electromagnetic noise, impulse noise, and fluctuation noise; and prior knowledge such as two-dimensional texture mapping are used to calculate the average received power. With the received signal-to-interference-plus-noise ratio Knowledge of inference algorithms, and compare the knowledge of inference algorithms with preset values ​​respectively. A comparison is made, and the semantic graph elements extracted from time-frequency features are judged based on the comparison results. Whether or not the 1-bit information representing interference occupying the channel is used to obtain the time-frequency characteristic radio frequency semantic map.

[0149] The average received power is calculated as follows:

[0150]

[0151] The method for calculating the signal-to-interference-plus-noise ratio (SINR) at the receiver is as follows:

[0152]

[0153]

[0154]

[0155]

[0156] Among them, among them, Indicates the transmission power. Indicates the satellite transmitting antenna gain. Indicates path loss. This indicates the weakening of clouds and rain. Indicates atmospheric decay, Indicates the flickering effect. Indicates the fading of the shadow. This represents power loss (cable loss, device loss, antenna loss, etc.) that is independent of the fading type in the propagation path. Indicates the antenna gain of the receiving terminal. Indicates the gain of the airborne antenna. Indicates signal wavelength , Indicates the distance between the satellite transmitter and the ground receiver. Power, gain, and loss are all in units of 1. . Indicates the received signal dryness ratio (SINR). Indicates the received signal power. This represents the length of the two sides of the time-frequency feature radio frequency semantic map.

[0157] (3) Amplitude and Phase Feature RF Semantic Graph Construction Scheme: The three-dimensional contour mapping method used in step two to obtain a contour-like three-dimensional solid map accurately represents the amplitude and phase features of the RF received signal. This is based on the number of "altitude" and the number of closed "horizontal lines" in the three-dimensional contour model. The mapping relationship will correspond to amplitude and phase. Numerical values ​​are filled into the corresponding positions in the semantic graph according to the edge-center expansion principle. This yields the amplitude and phase characteristic radio frequency semantic map.

[0158] Step 4: Design of Modulation Recognition Algorithm Based on Graph Convolutional Neural Network

[0159] The main approach is as follows: First, a set of radio frequency semantic information maps generated by various explicit semantic features is formed. Then, a modulation recognition model based on graph convolutional neural networks is constructed, and input layers, convolutional layers, pooling layers, fully connected layers and activation functions are designed. The model is trained and the weight parameter matrix is ​​iteratively optimized. Finally, the parameters and structure of the neural network model are determined. Finally, the modulation recognition result is output by inputting the low-orbit satellite internet radio frequency semantic information map into the input layer.

[0160] (1) Construct a set of explicit feature radio frequency semantic maps Based on the three types of low-Earth orbit satellite internet radio frequency semantic maps generated in step three, a standard image set is constructed. This image set is then used as the input layer of a graph convolutional neural network to generate the standard image map labels built into the graph convolutional neural network. .

[0161] (2) Constructing a low-orbit satellite internet radio frequency semantic graph recognition and classification model based on graph convolutional neural networks: The graph convolutional neural network input layer is designed to receive various radio frequency semantic graphs generated in step three; the convolutional layer is designed to include the actual physical meaning represented by the spectrogram elements in various radio frequency semantic graphs; the pooling layer is designed as a max pooling scheme to solve the problems of repetition of the physical meaning represented by the spectrogram elements after convolution calculation and the invalidity of some data, and to achieve effective substitution; the fully connected layer is designed to reorganize all processed semantic graph elements into a one-dimensional data sequence to meet the requirement that the graph convolutional neural network can only recognize and classify one-dimensional data sequences; the activation function is designed as... A function is used to ensure that the backpropagation gradient is never zero when iterating the classification weight parameter matrix. The weight matrix is ​​continuously optimized iteratively until the loss function meets the requirements. Finally, the parameters and structure of the low-orbit satellite internet radio frequency semantic map recognition and classification model based on graph convolutional neural network are determined.

[0162] (3) Modulation mode identification strategy based on the radio frequency semantic graph identification classification model of low-orbit satellite Internet: input the radio frequency semantic graph into the constructed modulation recognition graph convolutional neural network model. The output of the graph convolutional neural network will be the probability value of a certain modulation mode. The modulation mode corresponding to the maximum probability value is the modulation mode of the received radio frequency signal.

[0163] In summary, the low-Earth orbit (LEO) satellite signal modulation recognition method and apparatus based on radio frequency (RF) learning described in this invention calculates element values ​​from the RF signal and constructs a three-dimensional spectral array. It then extracts frequency domain features, time-frequency features, and amplitude-phase features from the 3D spectral array. The signal features for generating the RF semantic map are determined by calculating the cross-entropy values ​​of each feature. A frequency domain feature RF semantic map is obtained based on the frequency domain features, a time-frequency feature semantic map based on the time-frequency features, and an amplitude-phase feature RF semantic map based on the amplitude-phase features. A LEO satellite RF semantic map recognition and classification model is constructed. The feature semantic map corresponding to the maximum cross-entropy value is input into the LEO satellite RF semantic map recognition and classification model for classification and recognition, outputting the modulation recognition result. This transforms the processing of RF signals into image processing, improving the accuracy of signal modulation method recognition in LEO satellite communication scenarios.

[0164] Furthermore, a decision scheme for sharpening and concealing element values ​​under simultaneous frequency domain conditions was designed to conceal invalid element values ​​due to interference and sharpen valid element values, thereby eliminating the influence of channel noise interference and improving the accuracy of signal modulation recognition.

[0165] Furthermore, different extraction methods were used to extract different dominant features of the radio frequency signal in the three-dimensional spectral array, which improved the accuracy of signal feature extraction.

[0166] Furthermore, different methods are used to construct radio frequency semantic maps for different explicit features, so that the actual physical meaning represented by different explicit features is different. This makes it easier for the low-orbit satellite radio frequency semantic map identification and classification model to identify the modulation method of the signal in the radio frequency semantic map, thereby improving the identification accuracy.

[0167] Corresponding to the above method, the present invention also provides an apparatus comprising a computer device including a processor and a memory, the memory storing computer instructions, the processor executing the computer instructions stored in the memory, and the apparatus performing the steps of the method as described above when the computer instructions are executed by the processor.

[0168] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned edge computing server deployment method. The computer-readable storage medium can be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, floppy disks, hard disks, removable storage disks, CD-ROMs, or any other form of storage medium known in the art.

[0169] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0170] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0171] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0172] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations can be made to the embodiments of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for low-Earth orbit satellite signal modulation identification based on radio frequency learning, characterized in that, The method includes the following steps: Radio frequency (RF) signals from multiple satellite terminal user points are acquired. A Discrete Fourier Transform (DFT) is performed on the RF signal of each satellite terminal user point to obtain its spectrum at each time point. The spectra of each satellite terminal user point at each time point are then evolved over time and stacked into a planar spectrum. Image grayscale extraction processing is performed on the element values ​​of each element in the planar spectrum to obtain a two-dimensional visualization image. Statistical analysis is performed on the element values ​​in each two-dimensional visualization image, followed by sharpening and masking operations. The images are then stacked into a three-dimensional spectral array according to the sequence number of each satellite terminal user point. The three-dimensional spectral array is projected onto the two-dimensional plane containing the time-frequency domain to obtain a two-dimensional visualization spectrum of the superimposed satellite terminal user points. The element values ​​in the two-dimensional visualization spectrum are sampled, and the sampled elements are... The values ​​are divided into multiple subsequences; the two-dimensional visualized spectrum is sampled according to the position of the elements in each subsequence to obtain multiple sub-images; the Laplacian operator of each sampled element value in each sub-image is calculated and combined to obtain the Laplacian matrix of all sampled element values, the Laplacian operator in the Laplacian matrix is ​​used to characterize the frequency domain characteristics of the radio frequency signal; the three-dimensional spectrum is cut along the direction of the absolute value of the element numerical gradient from the full time domain and the full frequency domain, the influence value of multiple neighboring element values ​​of any element value on the cut surface is calculated and noise interpolation is performed, and the noise value and each element value are rendered into a heat map according to the numerical magnitude to characterize the time-frequency characteristics of the radio frequency signal; the radio frequency signal received by each satellite terminal user point in the full time domain is calculated. The amplitude and phase of the signal are determined and mapped according to the serial number of each satellite terminal user point to obtain a contour-like three-dimensional image to characterize the amplitude and phase features of the radio frequency signal; the cross-entropy values ​​of the frequency domain features, the time-frequency features, and the amplitude and phase features are calculated respectively, and the feature with the largest cross-entropy value is used to generate a radio frequency semantic map; a pre-trained low-Earth orbit satellite radio frequency semantic map recognition and classification model is obtained, which includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an activation function connected in sequence; the radio frequency semantic map is input into the pre-trained low-Earth orbit satellite radio frequency semantic map recognition and classification model to obtain the output modulation method recognition result; wherein, generating the radio frequency semantic map includes converting the element values ​​in the sub-image into... The total gain of the element value change trend is mapped to the chromaticity and brightness of the semantic map elements according to a set ratio to obtain the frequency domain characteristic radio frequency semantic map; the average received power and received signal-to-dryness ratio of the radio frequency signal are calculated based on priors, and the average received power and received signal-to-dryness ratio are compared with the set values ​​to determine whether the radio frequency signal has bit errors, and a time-frequency characteristic radio frequency semantic map is constructed; the element values ​​in the contour-like three-dimensional solid map are filled into the semantic map from the edge to the center according to the edge-center principle to obtain the amplitude-phase characteristic radio frequency semantic map; the edge-center principle means that the element values ​​in the amplitude-phase characteristic radio frequency semantic map gradually decrease from the edge to the center, and the color of the element values ​​in the amplitude-phase characteristic radio frequency semantic map gradually becomes darker from the edge to the center.

2. The method for low-Earth orbit satellite signal modulation identification based on radio frequency learning according to claim 1, characterized in that, After performing statistical analysis on the elements in each 2D visualization image, a sharpening and masking operation is performed, including... Calculate the statistical variance of element values ​​in all two-dimensional visualization images of each satellite terminal user point in the same time and frequency domain. After removing element values ​​whose deviation is greater than the statistical variance threshold, continue to calculate the statistical variance of the remaining element values ​​until the statistical variance of the remaining element values ​​is less than the statistical variance threshold. Calculate the statistical average of the remaining element values ​​and fill the positions of the removed element values ​​with the statistical average. If the number of times the statistical variance of the element values ​​is calculated exceeds a set number and the statistical variance of the remaining element values ​​is still greater than the statistical variance threshold, then set all element values ​​in the two-dimensional visualization image to zero.

3. The method for low-Earth orbit satellite signal modulation identification based on radio frequency learning according to claim 1, characterized in that, The method further includes sampling the two-dimensional visualized spectrogram according to the position of elements in each sub-sequence to obtain multiple sub-images, and performing graph fast Fourier transform on the sub-sequences corresponding to each sub-image to obtain the transform sequence of each sub-sequence. The formula for calculating the transformation sequence is: ； ； in, Represents the transformation sequence, This represents the total number of sampled element values. Represents the subsequence, Indicates frequency, The number of subsequences is represented by , and u represents the element value in the subsequence.

4. The method for low-Earth orbit satellite signal modulation identification based on radio frequency learning according to claim 3, characterized in that, Calculate the Laplacian operator for each sampled element value in each sub-image and combine them to obtain the Laplacian matrix of all sampled element values. The formula for calculating the Laplacian operator for each sampled element is as follows: ； ； in, This refers to the Laplace operator, where m represents the ordinal number of the sampled element value. Represents edge weight, Represents the value of any sample element on the sub-image. This represents the eight sampled element values ​​adjacent to any sampled element value on the subimage, where i represents positive and negative integers from -4 to 4.

5. The method for low-Earth orbit satellite signal modulation identification based on radio frequency learning according to claim 1, characterized in that, The three-dimensional spectral array is segmented along the direction of the absolute value of the element numerical gradient in both the full time and full frequency domains. The influence of multiple neighboring element values ​​on any element value at the segmented surface is calculated and noise interpolation is performed. The noise values ​​and each element value are then rendered into a heatmap according to their numerical magnitude to characterize the time-frequency characteristics of the radio frequency signal. At a fixed frequency, the position of the largest element value on each two-dimensional visualization image and the eight element values ​​adjacent to the largest element value are found. The gradient between the largest element value and the eight adjacent element values ​​is calculated, and the image is cut along the direction of the maximum gradient value to obtain the first cutting surface of the three-dimensional spectral array. At a fixed time, the position of the largest element value on each two-dimensional visualization image and the eight element values ​​adjacent to the largest element value are found. The gradient between the largest element value and the eight adjacent element values ​​is calculated, and the image is cut along the direction of the maximum gradient value to obtain the second cutting surface of the three-dimensional spectral array.

6. The method for low-Earth orbit satellite signal modulation identification based on radio frequency learning according to claim 5, characterized in that, The three-dimensional spectral array is cut along the direction of the absolute value of the element numerical gradient in both the full time and full frequency domains. The influence of multiple neighboring element values ​​on any element value at the cut surface is calculated and noise interpolation is performed. The noise values ​​and each element value are then rendered into a heatmap according to their numerical magnitude to characterize the time-frequency characteristics of the radio frequency signal. This also includes... The gradient vector and distance vector between any element value on the first cutting surface and the next eight adjacent element values ​​are calculated respectively. The gradient vector is multiplied by the corresponding distance vector and accumulated to obtain the influence value of the eight adjacent element values ​​on the element value. The influence value is interpolated with noise value using a nonlinear interpolation function to make the surfaces of the first cutting surface and the next cutting surface smoother. The noise value and each element value on the three-dimensional spectral array are rendered into a heat map according to their numerical values ​​to represent the time-frequency characteristics.

7. The method for low-Earth orbit satellite signal modulation identification based on radio frequency learning according to claim 1, characterized in that, The amplitude and phase of the radio frequency signal received by each satellite terminal user point in the full time domain are calculated, and a contour-like three-dimensional map is obtained by arranging and mapping the data according to the serial number of each satellite terminal user point to characterize the amplitude and phase features of the radio frequency signal, including... The magnitude of the real and imaginary parts of the radio frequency signal received by each satellite terminal user point in the full time domain and the arctangent trigonometric function of the real and imaginary parts are calculated to obtain the signal amplitude and phase of the radio frequency signal of each satellite terminal user point. The signal amplitude and amplitude of each satellite terminal user point are plotted as a contour map with contour lines representing the phase and fill color representing the amplitude. The contour maps of each satellite terminal user point are sorted and mapped in sequence according to the serial number of each satellite terminal user point to obtain a three-dimensional contour map of the radio frequency signal.

8. The method for low-Earth orbit satellite signal modulation identification based on radio frequency learning according to claim 1, characterized in that, The average received power and received signal-to-dryness ratio of the radio frequency signal are calculated based on prior a priori calculations. These are then compared with set values ​​to determine if the radio frequency signal has bit errors, and a time-frequency characteristic radio frequency semantic map is constructed, including... The average received power and received signal-to-dryness ratio of the radio frequency signal are calculated based on the broadband channel model, interference fluctuations, and random noise priors. The received signal-to-dryness ratio is compared with a set value to determine whether each frequency point of the communication bandwidth is occupied by interference. If it is occupied by interference, the signal at that frequency point is a bit error signal; if it is not occupied by interference, the signal at that frequency point is a bit error-free signal. Different colors are assigned to bit error signals and bit error-free signals, and the time-frequency feature radio frequency semantic map is constructed.

9. A method and apparatus for low-Earth orbit satellite signal modulation identification based on radio frequency learning, comprising a processor and a memory, characterized in that, The memory stores computer instructions, and the processor executes the computer instructions stored in the memory. When the computer instructions are executed by the processor, the device implements the steps of the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1 to 8.

Images