Satcom antenna tracking method and system fusing monopulse phase comparison and inertial navigation information

Abstract

The application relates to the technical field of satellite communication antenna tracking, and discloses a moving satellite communication antenna tracking method and system fusing single-pulse phase comparison and inertial navigation information. The method comprises the following steps: processing satellite signals and inertial navigation information to generate an initial observation sequence, obtaining a space state vector and a time correlation matrix through double-channel processing, and forming a state fusion mapping through cross projection. Multi-scale feature extraction is performed on the state fusion mapping to obtain steady-state and dynamic feature components. The steady-state features are used to construct an error evolution manifold, and the error evolution manifold and the dynamic features are input into a compensation decision network. The network outputs a multi-step prediction control instruction sequence based on the interaction between the error evolution manifold and the dynamic features. The sequence is dynamically compensated and optimized in combination with a real-time attitude change rate to generate an anti-disturbance control instruction and is executed, and finally, the state fusion mapping is corrected online through closed-loop feedback data. The method improves the tracking precision and stability of the antenna under complex motion through accurate modeling of the dynamic error and implementation of intelligent control.

Description

Technical Field

[0001] This invention relates to the field of satellite communication antenna tracking technology, specifically to a method and system for tracking mobile communication antennas that integrates monopulse phase ratio and inertial navigation information. Background Technology

[0002] In mobile satellite communication systems, maintaining the antenna's pointing precisely at the satellite is a core technological challenge. Existing methods generally integrate monopulse phase comparison and inertial navigation information. The monopulse system provides real-time angular deviation measurement, while the inertial navigation system outputs the angular velocity and angular acceleration information of the carrier. This information, after filtering, is typically used as input to classical control algorithms to drive the antenna servo mechanism, thereby counteracting the effects of carrier motion.

[0003] These conventional technical solutions have inherent limitations. Their core control models mostly rely on linear or quasi-linear state-space equations, making it difficult to accurately describe the dynamic evolution of errors under complex vehicle maneuvers, external wind disturbances, and the platform's own nonlinear coupling. This inadequate model descriptive capability directly limits the theoretical upper limit of control accuracy. Furthermore, mainstream control strategies are essentially based on passive feedback compensation of current or recent errors, lacking the ability to proactively predict and plan for the future development trend of the system state, leading to tracking lag and overshoot problems when dealing with rapid and abrupt movements. Summary of the Invention

[0004] The purpose of this invention is to provide a tracking method and system for a moving antenna that integrates single-pulse phase ratio and inertial navigation information, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, this invention provides a tracking method for a mobile antenna that integrates single-pulse phase comparison and inertial navigation information. The method includes: processing satellite signals received by the mobile antenna and navigation information output by an inertial navigation device to generate an initial observation sequence; performing dual-channel parallel processing on the initial observation sequence, generating a spatial state vector characterizing the antenna pointing state and a time correlation matrix characterizing the dynamic change trend through a spatial spectrum decomposition channel and a time correlation matrix, respectively; at the state fusion layer, cross-projecting the spatial state vector and the time correlation matrix to generate a state fusion mapping containing multi-dimensional state features; and performing multi-scale feature extraction on the state fusion mapping, extracting steady-state feature components corresponding to the macroscopic motion trend and features related to the dynamic change trend. The system analyzes the dynamic characteristic components of high-frequency disturbances; based on the extracted steady-state characteristic components, it constructs an error evolution manifold to describe the changes in antenna alignment error; it inputs the error evolution manifold and dynamic characteristic components into a preset compensation decision network, which outputs a control command sequence with multi-step future prediction characteristics based on their interaction; it dynamically compensates the control command sequence based on the control command sequence and the attitude change rate in the current navigation information to generate real-time control commands optimized for disturbance resistance; it executes the real-time control commands and collects the actual antenna pointing data after the execution of the real-time control commands to form a closed-loop feedback flow, which is injected into the state fusion layer to correct the state fusion mapping online.

[0006] Preferably, the step of processing the satellite signals received by the on-the-move antenna and the navigation information output by the inertial navigation device to generate an initial observation sequence includes: performing single-pulse phase comparison processing on the received satellite signals, calculating the phase difference between signals from different feed sources, and obtaining a pointing error phase difference sequence; synchronously acquiring the carrier attitude angle and angular velocity information output by the inertial navigation device to obtain a carrier motion sequence; aligning the pointing error phase difference sequence and the carrier motion sequence with timestamps and unifying the data format, combining them into a time-synchronized, format-standardized multidimensional data sequence, and using the multidimensional data sequence as the initial observation sequence.

[0007] Preferably, the dual-channel parallel processing of the initial observation sequence includes: in the spatial spectrum decomposition channel, performing spatial spectrum estimation on the phase difference data of different feed sources, separating the principal components of the spatial spectrum related to the antenna main beam pointing and the sidelobe components reflecting environmental interference, and constructing the spatial state vector from the principal components of the spatial spectrum; in the time correlation channel, performing autocorrelation and cross-correlation analysis on the continuous carrier motion sequence, identifying the similarity and difference of the carrier motion mode in different time segments, and forming the time correlation matrix.

[0008] Preferably, in the state fusion layer, the cross-projection of the spatial state vector and the temporal correlation matrix includes: projecting the spatial state vector onto a subspace spanned by the eigenvectors of the temporal correlation matrix and calculating projection coefficients; projecting the temporal correlation matrix onto a space spanned by the basis vectors of the spatial state vector and calculating the projection energy distribution; and constructing a tensor structure that integrates the spatiotemporal characteristics of both as the state fusion mapping based on the projection coefficients and the projection energy distribution.

[0009] Preferably, the multi-scale feature extraction of the state fusion mapping includes: applying a low-pass filter bank to perform downsampling smoothing on the state fusion mapping to extract low-frequency components reflecting the main trend of motion as the steady-state feature components; and applying a high-pass filter bank to perform differential processing on the state fusion mapping to extract high-frequency components reflecting instantaneous fluctuations as the dynamic feature components.

[0010] Preferably, the step of constructing an error evolution manifold to describe the changes in antenna-to-satellite error based on the extracted steady-state feature components includes: arranging the steady-state feature components in time over a continuous observation period to form an error feature trajectory; using a manifold learning algorithm to perform dimensionality reduction and geometric structure modeling on the error feature trajectory to obtain a low-dimensional smooth surface, where each point on the low-dimensional smooth surface represents an error state, the tangent direction at that point represents the evolution direction of the error state at the corresponding time, and the slope of the tangent represents the evolution rate of the error state, thus using the low-dimensional smooth surface as the error evolution manifold.

[0011] Preferably, the step of inputting the error evolution manifold and dynamic feature components into a preset compensation decision network includes: constructing the compensation decision network, which comprises a long short-term memory network layer, an attention mechanism layer, and a fully connected policy layer; inputting the tangent vector sequence of the error evolution manifold at the most recent preset number of observation times into the long short-term memory network layer to learn the time-dependent pattern of error evolution; simultaneously, inputting the dynamic feature components at the current time into the attention mechanism layer to calculate the weights between them and the hidden states output by the long short-term memory network layer, so as to focus on the historical error evolution information most relevant to the current disturbance; inputting the weighted and fused feature vector from the attention mechanism layer into the fully connected policy layer, which outputs a sequence of antenna angle adjustment amounts containing multiple future control cycles through a set of preset activation function mappings, and using the sequence of antenna angle adjustment amounts as the control command sequence.

[0012] Preferably, the step of dynamically compensating the control command sequence based on the control command sequence and the attitude change rate in the current navigation information to generate real-time control commands optimized for disturbance resistance includes: acquiring the real-time roll rate, pitch rate, and azimuth rate of the vehicle in the current navigation information; inputting the attitude change rate into a pre-trained disturbance prediction model, which predicts the additional pointing deviation caused by vehicle maneuvering within a future time window based on the time series of the change rate; vector-superimposing the predicted additional pointing deviation with the command value at the corresponding moment in the control command sequence to obtain a superimposed command sequence; and performing amplitude limiting and filtering smoothing processing on the superimposed command sequence to generate real-time control commands optimized for disturbance resistance.

[0013] Preferably, the step of executing the real-time control command and collecting the actual antenna pointing data after the execution of the real-time control command to form a closed-loop feedback flow, and injecting the closed-loop feedback flow into the state fusion layer to perform online correction of the state fusion mapping, includes: executing the disturbance-resistant optimized real-time control command to drive the servo mechanism of the mobile antenna to make adjustments; after the real-time control command is executed, collecting the actual pointing angle data of the antenna on the azimuth and elevation axes through the antenna's angle sensor to obtain the actual pointing sequence; and calculating the antenna pointing within the current control cycle based on the actual pointing sequence and the expected theoretical satellite pointing value. The antenna pointing residual sequence is associated and aligned with the carrier motion sequence corresponding to the generation of the real-time control command to construct a feedback data pair. The feedback data pair is processed by a feedback injection network, which maps the feedback data pair to a feedback correction amount with the same tensor structure as the state fusion mapping. In the state fusion layer, the feedback correction amount is weighted and fused with the state fusion mapping at the current time, wherein the weight of the feedback correction amount is calculated and determined in real time by an adaptive algorithm based on the confidence of the feedback data pair, thereby completing the online correction of the state fusion mapping.

[0014] Preferably, the present invention also includes a mobile antenna tracking system that integrates monopulse phase comparison and inertial navigation information. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the mobile antenna tracking method that integrates monopulse phase comparison and inertial navigation information as described above.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows.

[0016] An error evolution manifold is constructed to describe the variation of antenna alignment error. A high-dimensional differential manifold structure is mathematically defined using macroscopic steady-state characteristics extracted from the system state. This structure characterizes the dynamic change of the error as an evolutionary process along a specific path on the manifold surface. This approach breaks through the traditional linear state equation framework, enabling a more accurate depiction of the error's evolution trajectory under the combined influence of nonlinear system dynamics and complex external disturbances. An internal reference model with stronger mathematical representation capabilities is established for the control system, allowing state evaluation based on the intrinsic evolution of the error, rather than solely relying on external instantaneous observations, thereby improving the accuracy and consistency of the state description.

[0017] The error evolution manifold and dynamic feature components are input into a pre-defined compensation decision network. Based on their interaction, this network outputs a sequence of control commands with multi-step future prediction capabilities. Through models such as neural networks, a deep fusion and joint analysis of steady-state evolution trends and transient disturbance characteristics is achieved. Its decision mechanism can deduce the future state of the system based on the current state's position on the manifold and the direction and intensity of dynamic disturbances, generating corresponding forward-looking control sequences. This transforms the control logic from compensating for past errors to actively shaping the future state. The control commands can preemptively offset expected deviations caused by both internal dynamics and external disturbances, reducing system response lag, enhancing the ability to suppress sudden movements and persistent disturbances, and improving the smoothness of the tracking process and the steady-state accuracy of target pointing. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the working principle of the mobile antenna tracking method that integrates single-pulse phase comparison and inertial navigation information as described in this invention.

[0019] Figure 2 A flowchart for generating the initial observation sequence;

[0020] Figure 3 A flowchart for generating state fusion mappings for cross projection;

[0021] Figure 4 A line graph showing the dynamic changes of the three-axis angular velocity of the on-the-go antenna carrier;

[0022] Figure 5 A periodic comparison histogram of the pointing residual of a moving-at-the-moment antenna. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Please see Figure 1 This invention provides a tracking method for a mobile antenna that integrates single-pulse phase comparison and inertial navigation information. The method includes: processing satellite signals received by the mobile antenna and navigation information output by an inertial navigation device to generate an initial observation sequence; performing dual-channel parallel processing on the initial observation sequence, generating a spatial state vector characterizing the antenna pointing state and a temporal correlation matrix characterizing the dynamic change trend through a spatial spectrum decomposition channel and a temporal correlation matrix, respectively; cross-projecting the spatial state vector and the temporal correlation matrix in a state fusion layer to generate a state fusion mapping containing multi-dimensional state features; and performing multi-scale feature extraction on the state fusion mapping to extract steady-state features corresponding to the macroscopic motion trend. The dynamic feature components corresponding to the quantity and high-frequency disturbance are extracted. Based on the extracted steady-state feature components, an error evolution manifold is constructed to describe the change of antenna alignment error. The error evolution manifold and the dynamic feature components are input into a preset compensation decision network. Based on the interaction between the two, the compensation decision network outputs a control command sequence with future multi-step prediction characteristics. The control command sequence is dynamically compensated according to the control command sequence and the attitude change rate in the current navigation information to generate real-time control commands optimized for disturbance resistance. The real-time control commands are executed and the actual antenna pointing data after the command execution is collected to form a closed-loop feedback flow. The closed-loop feedback flow is injected into the state fusion layer to correct the state fusion mapping online.

[0025] In one embodiment of the present invention, see [reference] Figure 2 Generating the initial observation sequence involves performing single-pulse phase comparison processing on the received satellite signals, calculating the phase difference between signals from different feed sources to obtain the pointing error phase difference sequence. The specific formula is as follows: .in, express Time of the first The first feed and the first The phase difference of the signals received by each feed source for Time of the first The phase value of the signal received by each feed source for Time of the first The phase values ​​of the signals received by each feed source are obtained. Simultaneously, the carrier attitude angle and angular velocity information output by the inertial navigation device are acquired to obtain the carrier motion sequence. The pointing error phase difference sequence and the carrier motion sequence are time-stamped and formatted to form a time-synchronized, formatted, multi-dimensional data sequence as the initial observation sequence. During dual-channel parallel processing of the initial observation sequence, the phase difference data from different feed sources are spatially estimated in the spatial spectrum decomposition channel, separating the principal components of the spatial spectrum related to the antenna main beam pointing and the sidelobe components reflecting environmental interference. A spatial state vector is constructed from the principal components of the spatial spectrum. In the time correlation channel, autocorrelation and cross-correlation analysis are performed on the continuous carrier motion sequence to identify the similarities and differences in carrier motion patterns at different time segments, forming a time correlation matrix.

[0026] In specific implementation, the method involves processing satellite signals received by the on-the-go communication antenna and navigation information output by the inertial navigation device to generate an initial observation sequence, and performing dual-channel parallel processing on the initial observation sequence. A specific implementation example is an on-the-go communication antenna system installed on an off-road vehicle. The on-the-go communication antenna operates in the Ku-band and continuously receives beacon signals from geostationary orbit communication satellites. The inertial navigation device outputs the three-axis attitude angles and three-axis angular velocity information of the off-road vehicle in real time. In specific implementation, the step of generating the initial observation sequence is to perform single-pulse phase comparison processing on the satellite signals received by the on-the-go communication antenna. The single-pulse phase comparison processing calculates the phase between satellite signals from four independent feed sources. The phase difference is measured by outputting instantaneous phase difference values ​​in the azimuth and pitch directions through the phase detection circuit. The phase difference measurement values ​​are recorded at a sampling frequency of 100 Hz, forming a continuous pointing error phase difference sequence over time. Each data point in the pointing error phase difference sequence contains an azimuth error phase difference value and a pitch error phase difference value. In specific implementation, the carrier attitude angle and angular velocity information output by the inertial navigation device are acquired synchronously. The inertial navigation device outputs the roll angle, pitch angle, heading angle and corresponding angular velocity of the off-road vehicle at a frequency of 100 Hz, forming a carrier motion sequence. Each data point in the carrier motion sequence is a six-dimensional vector containing three attitude angles and three angular velocity values.

[0027] In some embodiments, the pointing error phase difference sequence and the carrier motion sequence are time-stamp aligned and data format unified. The time-stamp alignment operation involves marking each sampling point of the pointing error phase difference sequence with a precise time stamp from the antenna signal processing unit, and marking each data point of the carrier motion sequence with a precise time stamp from the inertial navigation device clock. The data processing unit uses a common reference clock as a reference to interpolate and align the time stamps of the two sequences, so that data points with the same index number in the two sequences correspond to the same physical time. During the data format unification operation, each data point of the pointing error phase difference sequence is converted from polar coordinates to rectangular coordinates, and the angular velocity unit of the carrier motion sequence is unified from degrees per second to radians per second. In a specific implementation, the combined data into a time-synchronized, format-standardized multidimensional data sequence is used as the initial observation sequence. In the combined initial observation sequence, the observation data at each moment is an eight-dimensional vector containing two phase differences, three attitude angles, and three angular velocities. The initial observation sequence is updated at a frequency of 100 Hz and sent to the dual-channel parallel processing module.

[0028] In some embodiments, the initial observation sequence is processed in parallel through two channels. In the spatial spectrum decomposition channel, the phase difference data of different feed sources are spatially estimated. Specifically, the phase difference data of multiple consecutive moments in the initial observation sequence are combined into a snapshot matrix. The covariance matrix of the snapshot matrix is ​​decomposed into eigenvalues. After decomposition, eigenvalues ​​and their corresponding eigenvectors are obtained in order of size. In a specific implementation, the principal component of the spatial spectrum related to the antenna main beam pointing and the sidelobe component reflecting environmental interference are separated. The principal component of the spatial spectrum corresponds to the signal subspace represented by the largest eigenvalue, and the sidelobe component corresponds to the noise subspace represented by the other smaller eigenvalues. A spatial state vector is constructed from the principal component of the spatial spectrum. The spatial state vector consists of the eigenvector corresponding to the largest eigenvalue and the largest eigenvalue itself. The spatial state vector represents the energy distribution center of the current antenna main beam pointing in the spatial spectrum.

[0029] In practical implementation, autocorrelation and cross-correlation analyses are performed on continuous carrier motion sequences in the time correlation channel. The carrier motion sequence includes three angular velocity components: roll, pitch, and yaw. The autocorrelation function of each angular velocity component is calculated at different time delays. Cross-correlation matrices are calculated for roll angular velocity and pitch angular velocity, roll angular velocity and yaw angular velocity, and pitch angular velocity and yaw angular velocity. In practical implementation, the similarity and difference of carrier motion patterns at different time segments are identified. By analyzing the peak value of the autocorrelation function at a specific time delay, the periodic pattern of carrier motion is identified. By analyzing the size of the off-diagonal elements of the cross-correlation matrix, the coupling strength between different axial motions is identified, forming a time correlation matrix. The time correlation matrix is ​​a comprehensive matrix. Its diagonal elements are composed of the integral values ​​of the autocorrelation coefficients of each angular velocity component within a preset time window, and its off-diagonal elements are composed of the integral values ​​of the cross-correlation coefficients between different angular velocity components within a preset time window.

[0030] In one embodiment of the present invention, see [reference] Figure 3 In the state fusion layer, the spatial state vector and the temporal correlation matrix are cross-projected. Specifically, the spatial state vector is projected onto the subspace spanned by the eigenvectors of the temporal correlation matrix to calculate the projection coefficients, and the temporal correlation matrix is ​​projected onto the space spanned by the basis vectors of the spatial state vector to calculate the projection energy distribution. Based on the projection coefficients and the projection energy distribution, a tensor structure that integrates the spatiotemporal characteristics of both is constructed as the state fusion mapping. When performing multi-scale feature extraction on the state fusion mapping, a low-pass filter bank is applied to downsample and smooth the state fusion mapping to extract low-frequency components reflecting the main trend of motion as steady-state feature components, and a high-pass filter bank is applied to differentially process the state fusion mapping to extract high-frequency components reflecting instantaneous fluctuations as dynamic feature components.

[0031] In specific implementation, the method involves cross-projecting the spatial state vector and the temporal correlation matrix in the state fusion layer, and performing multi-scale feature extraction on the state fusion mapping. Taking one working cycle as an example, the spatial state vector is a four-dimensional complex vector output from the spatial spectrum decomposition channel, containing energy center information representing the direction of the main beam. The temporal correlation matrix is ​​a three-row, three-column real symmetric matrix output from the temporal correlation channel, containing the temporal correlation characteristics of the carrier's three-axis angular velocity motion mode. In specific implementation, the spatial state vector is projected onto the subspace spanned by the eigenvectors of the temporal correlation matrix. First, the temporal correlation matrix is ​​eigenvalued to obtain three mutually orthogonal real eigenvectors. These eigenvectors together span a three-dimensional real subspace. Then, the four-dimensional complex spatial state vector is converted into a three-dimensional real vector through a mapping function. The projection length of this three-dimensional real vector in each eigenvector direction of the temporal correlation matrix is ​​calculated as the projection coefficient. The projection coefficient is a vector containing three real numbers.

[0032] In some embodiments, the temporal correlation matrix is ​​projected onto the space spanned by the basis vectors of the spatial state vector. The basis vectors of the spatial state vector are obtained by Gram-Schmidt orthogonalization of their real and imaginary vectors. These basis vectors span a four-dimensional complex space. The three-row, three-column temporal correlation matrix is ​​stacked column-wise into a nine-dimensional real vector. This nine-dimensional real vector is mapped onto the aforementioned four-dimensional complex space through a linear transformation matrix. The projection energy of the mapped vector on each basis vector of the spatial state vector is calculated. The projection energy distribution is a vector containing four non-negative real numbers. In a specific implementation, a tensor structure that integrates the spatiotemporal characteristics of both is constructed as a state fusion mapping based on the projection coefficients and the projection energy distribution. The construction method is to first perform an outer product operation on the projection coefficient vector and the projection energy distribution vector to obtain a three-row, four-column matrix. Then, this matrix is ​​subjected to a Kronecker product operation on the temporal correlation matrix to finally obtain a nine-row, twelve-column real matrix as the state fusion mapping. The state fusion mapping comprehensively encodes the coupling relationship between the spatial orientation state and the temporal motion mode.

[0033] It is understandable that when performing multi-scale feature extraction on the state fusion map, a low-pass filter bank is applied to downsample and smooth the state fusion map. The low-pass filter bank consists of two parallel elliptic low-pass digital filters. The cutoff frequency of the first elliptic low-pass digital filter is one-tenth of the system control bandwidth, and the cutoff frequency of the second elliptic low-pass digital filter is one-fifth of the system control bandwidth. Each column vector of the state fusion map is passed through these two elliptic low-pass digital filters, and the filtered output is sampled at intervals to achieve downsampling. The low-frequency components that reflect the main trend of motion are extracted as steady-state feature components. The steady-state feature components are composed of two downsampled matrices, corresponding to different smoothing degrees and time-domain resolutions.

[0034] It is understandable that a high-pass filter bank is applied to perform differential processing on the state fusion mapping. The high-pass filter bank consists of two parallel differential operators: the first is a first-order forward difference operator, and the second is a second-order central difference operator. Each row vector of the state fusion mapping is discretely convolved with these two differential operators to extract high-frequency components reflecting instantaneous fluctuations as dynamic feature components. The dynamic feature components consist of two differenced matrices, representing the first-order gradient and second-order curvature of the state change, respectively. In specific implementation, the multi-scale feature extraction process is mathematically expressed as follows: for the state fusion mapping tensor... The extracted steady-state feature components and dynamic feature components This is obtained through the following relationship: , Among them, symbols This represents a downsampling operation with a factor of 2, symbol... Representing the One low-pass filter core, The value can be 1 or 2, and the symbol is... Representing the A high-pass differential operator core, The value can be 1 or 2.

[0035] In one embodiment of the present invention, an error evolution manifold for describing the changes in antenna-to-satellite error is constructed based on the extracted steady-state feature components. Specifically, the steady-state feature components are arranged in time over a continuous observation period to form an error feature trajectory. A manifold learning algorithm is used to reduce the dimensionality and model the geometric structure of the error feature trajectory to obtain a low-dimensional smooth surface. Each point on the low-dimensional smooth surface and its tangent direction characterize the evolution direction and rate of a specific error state. This low-dimensional smooth surface is the error evolution manifold.

[0036] When the error evolution manifold and dynamic feature components are input into a pre-defined compensation decision network, the compensation decision network consists of a long short-term memory (LSM) network layer, an attention mechanism layer, and a fully connected policy layer. The tangent vector sequence of the error evolution manifold at the most recent predetermined number of observation times is input into the LSM network layer to learn the time-dependent pattern of error evolution. Simultaneously, the dynamic feature components at the current time step are input into the attention mechanism layer to calculate the weights between them and the hidden states output by the LSM network layer, thus focusing on the historical error evolution information most relevant to the current perturbation. The weighted and fused feature vector from the attention mechanism layer is input into the fully connected policy layer. The fully connected policy layer outputs a sequence containing antenna angle adjustments for multiple future control cycles as a control command sequence through a set of pre-defined activation functions. The weight calculation formula is as follows: .in, The weights between the hidden states output by the attention mechanism layer and the long short-term memory network layer are... The query vector is composed of dynamic feature components at the current moment. and These are all hidden state vectors output by the Long Short-Term Memory (LSTM) network layers. for Dimensions The function is used to normalize the weight distribution, thereby focusing on the historical error evolution information most relevant to the current perturbation, and inputting the weighted and fused feature vector from the attention mechanism layer into the fully connected policy layer.

[0037] In specific implementation, the method involves constructing an error evolution manifold to describe the changes in antenna alignment error based on the extracted steady-state feature components, and inputting the error evolution manifold and dynamic feature components into a preset compensation decision network. Taking an observation window containing fifty consecutive control cycles as an example, the steady-state feature components are two downsampled matrices output from the multi-scale feature extraction module. In specific implementation, the steady-state feature components are arranged in consecutive observation cycles to form an error feature trajectory. For each control cycle, the two steady-state feature matrices are expanded row by row and connected to form a one-dimensional feature vector. The one-dimensional feature vectors of the fifty consecutive cycles are arranged in chronological order to form a matrix with rows as time indexes and columns as feature dimensions. This matrix is ​​the error feature trajectory. Each row of the error feature trajectory represents the coordinate point of the antenna alignment error in the feature space at a specific moment.

[0038] In some embodiments, a manifold learning algorithm is used to reduce the dimensionality and model the geometric structure of the error feature trajectory. The manifold learning algorithm selected is the equidistant feature mapping algorithm. The equidistant feature mapping algorithm first calculates the Euclidean distance between all pairs of sample points in the error feature trajectory matrix to construct a distance matrix. Then, based on the distance matrix, it finds low-dimensional embedding coordinates that preserve the geodesic distance between sample points. In a specific implementation, a low-dimensional smooth surface is obtained. The low-dimensional smooth surface is achieved by mapping the high-dimensional error feature trajectory to a two-dimensional plane. On the mapped two-dimensional plane, fifty sample points are continuously distributed. The spline interpolation method is used to fit the surface to these discrete two-dimensional points, generating a smooth parametric surface passing through all points. Each point on the low-dimensional smooth surface and its tangent direction represent the evolution direction and rate of a specific error state. The low-dimensional smooth surface is the error evolution manifold. The coordinates of any point on the error evolution manifold correspond to the error state, the direction vector of the surface tangent at that point corresponds to the changing trend of the error state, and the magnitude of the tangent vector corresponds to the instantaneous rate of error evolution.

[0039] In its implementation, the compensation decision network consists of a long short-term memory (LSM) network layer, an attention mechanism layer, and a fully connected policy layer. The LSM network layer has 32 hidden units with an input dimension of two, corresponding to the two-dimensional tangent vector of the error evolution manifold. The attention mechanism layer is a network module based on dot product attention. The fully connected policy layer contains two linear transformation layers and one nonlinear activation function layer. The tangent vector sequence of the error evolution manifold at the most recent ten observation times is input into the LSM network layer to learn the time-dependent pattern of error evolution. The tangent vector sequence is obtained by calculating and normalizing the tangent directions at ten consecutive sample points on the error evolution manifold, forming an input sequence with ten time steps, each step being a two-dimensional vector. The LSM network layer processes this sequence sequentially and outputs the hidden state vector of the last time step.

[0040] It can be understood that the dynamic feature components at the current moment are input into the attention mechanism layer to calculate the weights between them and the hidden states output by the Long Short-Term Memory (LSTM) network layer. The dynamic feature components are two differenced matrices, which are expanded row-wise and concatenated into a one-dimensional vector as the query vector of the attention mechanism. The hidden state vector output by the LSTM network layer serves as the key vector and value vector. The attention mechanism layer obtains the attention weight distribution by calculating the dot product of the query vector and the key vector and applying the soft maximum function. In specific implementation, the focus is on the historical error evolution information most relevant to the current perturbation. The attention weights are used to perform a weighted summation of the hidden state sequence of the LSTM network layer over the past ten time steps to generate a context vector. This context vector integrates the part of the historical error evolution pattern most relevant to the current dynamic features.

[0041] It can be understood that the weighted and fused feature vector from the attention mechanism layer is input into the fully connected policy layer. The weighted and fused feature vector is formed by concatenating the context vector and the current dynamic feature component vector. The fully connected policy layer maps the input vector to 64 dimensions through a set of preset activation functions. The first linear transformation of the fully connected policy layer maps the input vector dimension to 64 dimensions and passes it through the hyperbolic tangent activation function. The second linear transformation maps the 64-dimensional vector to the output dimension, outputting a sequence of antenna angle adjustment amounts for the next five control cycles as the control command sequence. The control command sequence is a vector containing five elements, each of which is a two-dimensional real number pair, corresponding to the antenna azimuth and elevation axis angle adjustment command values ​​for the next five cycles, respectively. The calculation relationship of the forward propagation of the compensation decision network can be expressed as: Among them, symbols The symbol represents the output control command sequence vector. This represents the weight matrix of the first layer in the fully connected strategy layer, denoted by [symbol]. This represents the weight matrix of the second layer in the fully connected strategy layer, denoted by [symbol]. This represents the bias vector located after the first layer in the fully connected strategy layer, with the symbol... This represents the bias vector located after the second layer in the fully connected strategy layer, with the symbol... This represents the context vector output by the attention mechanism layer. With the current dynamic feature component vector The splicing operation.

[0042] In one embodiment of the present invention, the control command sequence is dynamically compensated based on the control command sequence and the attitude change rate in the current navigation information. Specifically, the real-time roll rate, pitch rate, and azimuth rate of the vehicle in the current navigation information are obtained. The attitude change rate is input into a pre-trained disturbance prediction model. The disturbance prediction model predicts the additional pointing deviation caused by the vehicle maneuver within a future time window based on the time series of the rate of change. The predicted additional pointing deviation is vector-superimposed with the command value at the corresponding moment in the control command sequence. The superimposed command sequence is subjected to amplitude limiting and filtering smoothing to generate real-time control commands optimized for disturbance resistance.

[0043] In specific implementation, the method involves dynamically compensating the control command sequence based on the control command sequence and the attitude change rate in the current navigation information. A specific implementation example is a scenario where a mobile antenna system installed on a high-speed vehicle is performing a slalom maneuver. The control command sequence is a sequence output from the compensation decision network containing antenna angle adjustments for the next five control cycles. The current navigation information comes from the vehicle's built-in inertial measurement unit. In this implementation, the real-time roll velocity, pitch velocity, and azimuth velocity of the vehicle are acquired from the current navigation information. These three angular velocity values ​​are directly read from the inertial measurement unit at a frequency of 100 times per second. These three angular velocity values ​​constitute a real-time three-dimensional angular velocity vector, representing the instantaneous rotational motion state of the vehicle at the current moment.

[0044] In some embodiments, the attitude change rate is input into a pre-trained disturbance prediction model, which is a fully connected neural network with two hidden layers. The disturbance prediction model is trained offline using a large amount of historical maneuver data before system deployment. The training data includes angular velocity time series under various typical maneuver modes and their corresponding antenna pointing deviations. The disturbance prediction model predicts the additional pointing deviation caused by the vehicle maneuver within a future time window based on the time series of the rate of change. The input rate of change time series is a 33-dimensional input vector composed of the past ten sampling periods and the current three-dimensional angular velocity vector. The predicted future time window covers the next five control periods, and the output is a 10-dimensional vector corresponding to the additional pointing deviation predicted in the azimuth and pitch dimensions within the next five periods.

[0045] In practice, the predicted additional pointing deviation is vector-superimposed with the corresponding command values ​​in the control command sequence. The additional pointing deviation vector and the control command sequence vector have the same dimensional structure. The superposition operation is element-wise vector addition. The superimposed command sequence is a new ten-dimensional vector containing the original adjustment command and the correction amount introduced to compensate for the predicted disturbance. The working process of the disturbance prediction model and the superposition relationship can be expressed by a functional relationship, denoted as... This is a sequence of attitude angular velocities from the past and present moments. If the original control instruction sequence is followed by a dynamically compensated instruction sequence, then... Obtained in the following way: Among them, symbols The nonlinear mapping function implemented by the pre-trained perturbation prediction model maps the angular velocity history sequence to an additional pointing bias in the prediction. The sign... This represents vector addition.

[0046] It is understandable that the superimposed command sequence undergoes amplitude limiting and filtering smoothing. Amplitude limiting involves setting upper and lower amplitude limits for each element of the superimposed command sequence. The amplitude limit for azimuth adjustment commands is ±0.5 degrees, and the amplitude limit for pitch adjustment commands is ±0.3 degrees. Any command value exceeding this range will be truncated to the boundary value. Filtering smoothing involves using a fifth-order Butterworth low-pass filter to perform time-domain filtering on the amplitude-limited command sequence. The cutoff frequency of the filter is set to twice the servo bandwidth of the control system to eliminate high-frequency fluctuations in the commands while retaining the main control trends, generating disturbance-resistant optimized real-time control commands. The disturbance-resistant optimized real-time control commands are ten-dimensional vectors that have undergone amplitude limiting and smoothing, and are directly sent to the antenna servo drive unit for execution.

[0047] It is understandable that the weight parameters relied upon by the pre-trained perturbation prediction model are determined based on the characteristics of a specific carrier platform. See Table 1, which shows a range of compensation weight coefficients learned on the example training dataset, corresponding to different angular velocity change modes.

[0048] Table 1. Range of Compensation Weight Coefficients for Hidden Layer Nodes in the Perturbation Prediction Model to Typical Angular Velocity Input

[0049]

[0050] See Figure 4 , Figure 4This is a line graph showing the dynamic changes of the three-axis angular velocities of a mobile satellite communication (SSM) antenna carrier, a specialized data chart in the field of attitude control for mobile satellite communication systems. A significant step jump in roll velocity occurs around 0.2 seconds, which, combined with fluctuations in pitch and azimuth angular velocities, conforms to the typical attitude change characteristics of a slalom maneuver. The fluctuations in the three-axis angular velocities show a certain correlation, reflecting the multi-dimensional coupling characteristics of the carrier's attitude during maneuvering. This is a key disturbance source that needs to be compensated for in SSM tracking. This chart is used to analyze the impact of carrier maneuvering on antenna pointing and serves as core data for training disturbance prediction models and verifying control command compensation algorithms in SSM systems.

[0051] In one embodiment of the present invention, real-time control commands are executed and the actual pointing data of the antenna after command execution is collected to form a closed-loop feedback flow. The closed-loop feedback flow is injected into the state fusion layer to perform online correction of the state fusion mapping. Specifically, real-time control commands optimized for disturbance resistance are executed to drive the servo mechanism of the on-the-move antenna for adjustment. After the real-time control commands are executed, the actual pointing angle data of the antenna on the azimuth and elevation axes are collected by the antenna's angle sensor to obtain the actual pointing sequence. Based on the actual pointing sequence and the expected theoretical value of satellite pointing, the actual pointing sequence is calculated using the formula... The antenna pointing residual sequence within the current control period is calculated. Wherein, for At time t, the antenna points towards the residual vector. for The actual pointing angle vector of the antenna at any given moment. for The theoretical value vector of the satellite's expected direction at any given time. Then, using the formula... Obtain feedback data pairs. Among them, for Real-time feedback data is correct. for The vector of the carrier motion state corresponding to the real-time control command generated at any time.

[0052] The antenna pointing residual sequence is correlated and aligned with the carrier motion sequence corresponding to the generation of real-time control commands to construct feedback data pairs. These feedback data pairs are then passed through a feedback injection network, which maps them to feedback correction values ​​with the same tensor structure as the state fusion map. In the state fusion layer, the feedback correction values ​​are weighted and fused with the current state fusion map. The weights of the feedback correction values ​​are calculated in real-time by an adaptive algorithm based on the confidence level of the feedback data pairs, following the formula: .in, As a weight for the feedback correction amount, The mean variance of the three-axis angular velocity components of the carrier. This represents the average of the absolute values ​​of the differences between adjacent points in the antenna pointing residual sequence. and The positive adjustment parameter is determined through experimental debugging. , This allows for online correction of the state fusion mapping.

[0053] In specific implementation, the method involves executing real-time control commands optimized for disturbance resistance and collecting actual antenna pointing data after command execution to form a closed-loop feedback flow. The closed-loop feedback flow is then injected into the state fusion layer to perform online correction of the state fusion mapping. A specific implementation example is a scenario where a mobile antenna carrier traverses a rough road. The optimized real-time control command is a sequence containing five control cycle adjustment values. In the implementation, the optimized real-time control command is executed to drive the servo mechanism of the mobile antenna to make adjustments. The servo mechanism includes motor drivers for the azimuth and pitch axes. The control system takes the command value of the first control cycle in the optimized real-time control command sequence, converts it into a pulse width modulation signal, and sends it to the corresponding motor driver to drive the motor to rotate, thereby adjusting the pointing of the antenna reflector. The duration of one control cycle is ten milliseconds.

[0054] In specific implementation, after the real-time control command optimized for anti-disturbance is executed, the actual pointing angle data of the antenna on the azimuth and elevation axes is collected by the antenna's angle sensor. The angle sensor is a high-precision photoelectric encoder mounted on two rotating shafts. At the end of the control cycle, the photoelectric encoder reads and outputs the absolute angle position of the antenna to obtain the actual pointing sequence. The actual pointing sequence records the actual azimuth and elevation angles of the antenna in the current control cycle and the previous four cycles, forming a set of historical actual pointing data containing five time points. In some embodiments, the antenna pointing residual sequence in the current control cycle is calculated based on the actual pointing sequence and the expected theoretical satellite pointing value. The theoretical satellite pointing value is calculated in real time based on the real-time position of the carrier, satellite orbit parameters, and antenna installation matrix. The antenna pointing residual sequence is obtained by performing vector difference operation between each azimuth and elevation angle data in the actual pointing sequence and the corresponding theoretical satellite pointing value. The antenna pointing residual sequence contains five residual points, and each residual point contains an azimuth error angle and an elevation error angle.

[0055] In some embodiments, the antenna pointing residual sequence is associated and aligned with the carrier motion sequence corresponding to the generation of disturbance-resistant optimized real-time control commands to construct feedback data pairs. The carrier motion sequence corresponding to the generation of disturbance-resistant optimized real-time control commands is inertial navigation data stored in a cache and synchronized with the command calculation time, including roll angle, pitch angle, yaw angle, and three-axis angular velocity. The association and alignment operation binds each residual point in the antenna pointing residual sequence with the motion state data of the carrier motion sequence at the same time according to a unified timestamp to construct feedback data pairs. Each feedback data pair is a data structure containing a timestamp, a two-dimensional pointing residual vector, and a six-dimensional carrier motion state vector. It can be understood that the feedback data pairs are passed through a feedback injection network, which consists of a three-layer fully connected neural network. The input layer dimension is eight, corresponding to the concatenation of the two-dimensional residual vector and the six-dimensional motion state vector in the feedback data pair. The output layer dimension is the same as the dimension of the state fusion mapping. The feedback injection network maps the feedback data pairs into feedback correction quantities with the same tensor structure as the state fusion mapping. The feedback correction quantity is a nine-row twelve-column real matrix.

[0056] It can be understood that in the state fusion layer, the feedback correction amount is weighted and fused with the current state fusion mapping. The weights are determined in real-time by an adaptive algorithm based on the confidence level of the feedback data. This algorithm calculates the confidence score by evaluating the noise level of the carrier motion sequence and the continuity of the residual sequence in the feedback data pair. The noise level is obtained by calculating the variance of the angular velocity components in the carrier motion sequence, and the continuity is obtained by calculating the mean of the differences between adjacent points in the residual sequence. The calculation method is the same as the weight calculation method for the feedback correction amount, specifically: Among them, symbols Represents the confidence score, symbol The mean variance of the three-axis angular velocity components of the carrier, with the sign... The average of the absolute values ​​of the differences between adjacent points in the antenna-pointing residual sequence, with the sign... and It is a positive adjustment parameter. In practice, the weighted fusion operation multiplies the feedback correction by the confidence score. Then, the current state is fused, mapped, and multiplied by... Finally, the two product matrices are added together to complete the online correction of the state fusion mapping. The updated state fusion mapping will be used for multi-scale feature extraction and subsequent processes in the next control cycle.

[0057] See Figure 5 , Figure 5This is a periodic comparison bar chart of the pointing residuals of a mobile satellite communication antenna, a professional data chart in the field of attitude control for mobile satellite communication systems. The fluctuation range of the azimuth residual is greater than that of the pitch residual (e.g., the azimuth residuals in periods 2 and 3 are below -0.1 degrees), reflecting stronger perturbations in the azimuth dimension caused by the vehicle's maneuver (consistent with attitude change patterns on rough terrain). The significant increase in the residual in period 5 indicates that the vehicle perturbations intensify during this stage, leading to a larger antenna pointing deviation. This chart is used to evaluate the pointing accuracy of the mobile satellite communication antenna; the magnitude of the residuals directly reflects the effectiveness of the control commands. The residual data is the core input of the closed-loop feedback flow and will be used subsequently to correct the state fusion mapping, improving the accuracy of the next round of control.

[0058] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0059] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A tracking method for a moving-mode antenna that integrates single-pulse phase comparison and inertial navigation information, characterized in that, The method includes: The satellite signals received by the mobile antenna and the navigation information output by the inertial navigation device are processed to generate an initial observation sequence; The initial observation sequence is processed in parallel through two channels, and a spatial state vector representing the antenna pointing state and a time correlation matrix representing the dynamic change trend are generated through the spatial spectrum decomposition channel and the time correlation channel, respectively. In the state fusion layer, the spatial state vector is cross-projected with the temporal correlation matrix to generate a state fusion mapping containing multi-dimensional state features; Multi-scale feature extraction is performed on the state fusion mapping to extract the steady-state feature components corresponding to the macroscopic motion trend and the dynamic feature components corresponding to the high-frequency disturbance, respectively. Based on the extracted steady-state feature components, an error evolution manifold is constructed to describe the changes in antenna-to-star error. The error evolution manifold and dynamic feature components are input into a preset compensation decision network. Based on the interaction between the two, the compensation decision network outputs a control command sequence with multi-step future prediction characteristics. Based on the control command sequence and the attitude change rate in the current navigation information, the control command sequence is dynamically compensated to generate real-time control commands optimized for disturbance resistance. The real-time control command is executed, and the actual antenna pointing data after the real-time control command is executed is collected to form a closed-loop feedback flow. The closed-loop feedback flow is injected into the state fusion layer to correct the state fusion mapping online.

2. The tracking method for a moving-mode antenna that integrates single-pulse phase comparison and inertial navigation information according to claim 1, characterized in that, The process of processing satellite signals received by the on-the-go antenna and navigation information output by the inertial navigation device to generate an initial observation sequence includes: The received satellite signals are processed by single-pulse phase comparison to calculate the phase difference between signals from different feed sources, thus obtaining the pointing error phase difference sequence. The carrier's attitude angle and angular velocity information output by the inertial navigation device are acquired simultaneously to obtain the carrier's motion sequence; The pointing error phase difference sequence and the carrier motion sequence are time-stamped and formatted to form a time-synchronized, formatted multidimensional data sequence, which is then used as the initial observation sequence.

3. The tracking method for a moving-mode antenna that integrates single-pulse phase comparison and inertial navigation information according to claim 2, characterized in that, The dual-channel parallel processing of the initial observation sequence includes: In the spatial spectrum decomposition channel, the phase difference data of different feed sources are spatially estimated to separate the principal component of the spatial spectrum related to the antenna main beam pointing and the sidelobe component reflecting environmental interference. The spatial state vector is constructed from the principal component of the spatial spectrum. In the time correlation channel, autocorrelation and cross-correlation analysis are performed on the continuous carrier motion sequence to identify the similarity and difference of the carrier motion pattern in different time segments, forming the time correlation matrix.

4. The tracking method for a moving-mode antenna that integrates single-pulse phase comparison and inertial navigation information according to claim 3, characterized in that, The step of cross-projecting the spatial state vector and the temporal correlation matrix in the state fusion layer includes: Project the spatial state vector onto the subspace spanned by the eigenvectors of the temporal correlation matrix, and calculate the projection coefficients; The time correlation matrix is ​​projected onto the space spanned by the basis vectors of the spatial state vector, and the projected energy distribution is calculated. Based on the projection coefficients and projection energy distribution, a tensor structure that integrates the spatiotemporal characteristics of both is constructed as the state fusion mapping.

5. The tracking method for a moving-mode antenna that integrates single-pulse phase comparison and inertial navigation information according to claim 4, characterized in that, The multi-scale feature extraction of the state fusion mapping includes: The state fusion mapping is downsampled and smoothed by applying a low-pass filter bank, and the low-frequency component reflecting the main trend of motion is extracted as the steady-state feature component. A high-pass filter bank is applied to perform differential processing on the state fusion mapping to extract high-frequency components that reflect instantaneous fluctuations as the dynamic feature components.

6. The tracking method for a moving-mode antenna that integrates single-pulse phase comparison and inertial navigation information according to claim 5, characterized in that, Based on the extracted steady-state feature components, an error evolution manifold is constructed to describe the changes in antenna-to-satellite error, including: Arrange the steady-state characteristic components over a continuous observation period to form an error characteristic trajectory; The error feature trajectory is reduced in dimension and its geometric structure is modeled using a manifold learning algorithm to obtain a low-dimensional smooth surface. Each point on the low-dimensional smooth surface represents an error state. The direction of the tangent at the point represents the evolution direction of the error state at the corresponding time, and the slope of the tangent represents the evolution rate of the error state. The low-dimensional smooth surface is used as the error evolution manifold.

7. The tracking method for a moving-mode antenna that integrates single-pulse phase comparison and inertial navigation information according to claim 6, characterized in that, The step of inputting the error evolution manifold and dynamic feature components into a preset compensation decision network includes: The compensation decision network is constructed, which includes a long short-term memory network layer, an attention mechanism layer, and a fully connected policy layer. The tangent vector sequence of the error evolution manifold at the most recent preset number of observation times is input into the long short-term memory network layer to learn the time-dependent pattern of error evolution. Simultaneously, the dynamic feature components at the current moment are input into the attention mechanism layer to calculate the weights between them and the hidden states output by the long short-term memory network layer, so as to focus on the historical error evolution information most relevant to the current perturbation. The weighted and fused feature vector from the attention mechanism layer is input into the fully connected strategy layer. The fully connected strategy layer maps a set of preset activation functions and outputs a sequence of antenna angle adjustment amounts for multiple future control cycles. The sequence of antenna angle adjustment amounts is used as the control command sequence.

8. The tracking method for a moving-mode antenna that integrates single-pulse phase comparison and inertial navigation information according to claim 7, characterized in that, The step of dynamically compensating the control command sequence based on the control command sequence and the attitude change rate in the current navigation information to generate real-time control commands optimized for disturbance resistance includes: Obtain the real-time roll rate, pitch rate, and azimuth rate of the vehicle from the current navigation information; The attitude change rate is input into a pre-trained perturbation prediction model, which predicts the additional pointing deviation caused by the vehicle maneuver within a future time window based on the time series of the change rate. The predicted additional pointing deviation is vector-superimposed with the command value at the corresponding time in the control command sequence to obtain the superimposed command sequence. The superimposed instruction sequence is subjected to amplitude limiting and filtering smoothing to generate real-time control instructions optimized for disturbance resistance.

9. The tracking method for a moving-mode antenna that integrates single-pulse phase comparison and inertial navigation information according to claim 8, characterized in that, The process of executing the real-time control command and collecting the actual antenna pointing data after the execution of the real-time control command forms a closed-loop feedback stream. This closed-loop feedback stream is then injected into the state fusion layer to perform online correction of the state fusion mapping, including: The disturbance-resistant optimized real-time control command is executed to drive the servo mechanism of the on-the-go antenna to make adjustments; After the real-time control command is executed, the actual pointing angle data of the antenna on the azimuth and elevation axes is collected by the antenna's angle sensor to obtain the actual pointing sequence. Based on the actual pointing sequence and the expected theoretical satellite pointing value, the antenna pointing residual sequence within the current control period is calculated. The antenna pointing residual sequence is associated and aligned with the carrier motion sequence corresponding to the generation of the real-time control command to construct a feedback data pair; The feedback data pairs are processed by a feedback injection network, which maps the feedback data pairs to feedback corrections that have the same tensor structure as the state fusion mapping. In the state fusion layer, the feedback correction amount is weighted and fused with the state fusion mapping at the current time. The weight of the feedback correction amount is determined in real time by an adaptive algorithm based on the confidence level of the feedback data, thereby completing the online correction of the state fusion mapping.

10. A tracking system for a moving antenna that integrates monopulse phase comparison and inertial navigation information, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the on-the-move antenna tracking method that integrates single-pulse phase ratio and inertial navigation information as described in any one of claims 1 to 9.

Images