A multi-source perception feedforward prediction compensation method for an inter-satellite laser link boresight

Abstract

The present application relates to the technical field of inter-satellite laser link, and particularly relates to a multi-source perception feedforward prediction compensation method for an inter-satellite laser link boresight, wherein an inter-satellite laser communication boresight alignment system is arranged on a satellite-borne laser terminal, and the system comprises: a multi-source sensing acquisition module connected to multiple sensors on a platform. A space-time registration module is connected to the multi-source sensing acquisition module. A prediction module is connected to the space-time registration module, and a trained time series prediction network is arranged in the prediction module. The prediction module outputs a feedforward prediction offset, and the output end of the prediction module is connected to a hybrid control calculation module. The hybrid control calculation module receives the feedforward amount of the prediction module at one end and receives the feedback amount calculated by a feedback controller from a light spot feedback detector at the other end, performs dynamic weighted fusion, and outputs a total control amount to a fast mirror driving module. The above modules are used to set a passive feedback correction to a hybrid closed-loop architecture combining feedforward prediction compensation and feedback fine-tuning.

Description

Technical Field

[0001] This invention relates to the technical field of inter-satellite laser links, and specifically to a multi-source sensing feedforward prediction and compensation method for the line of sight of an inter-satellite laser link. Background Technology

[0002] The establishment and maintenance of inter-satellite laser communication links rely on the high-precision alignment of the communication terminal with the line of sight (optical axis), a process known in the industry as Acquisition-Tracking-Pointing (ATP). A typical spaceborne laser terminal ATP system consists of two cascaded mechanisms: a coarse pointing mechanism (a gimbal-type two-dimensional turntable responsible for large-angle search and acquisition) and a fine pointing mechanism (centered on a fast reflector (FSM) responsible for precise correction at the micro-radian level). Its conventional operation involves the detector (CCD or four-quadrant detector QPD) detecting in real time the deviation of the beacon light or downlink spot relative to the detector center. The controller uses this error signal to drive the FSM motor to deflect, forming a closed-loop control loop of "detection error—feedback correction," ensuring the spot stably converges to the detector center.

[0003] For this field, existing publicly available technologies mainly follow the following routes: Firstly, optical signals are processed and optimized using algorithms to improve measurement accuracy. The core of this approach is to use improved wavelet thresholding and other image processing algorithms to denoise beacon light images affected by atmospheric turbulence and background light pollution during the optical signal reception and processing stage. Then, the gray-scale centroid method is used to calculate the centroid coordinates of the light spot, thereby improving centroid positioning accuracy. Essentially, this involves more accurately "measuring" deviations that have already occurred, representing an improvement in post-hoc measurement accuracy, and does not involve predicting or proactively offsetting the causes of deviations (platform vibration).

[0004] Secondly, there is the calibration benchmark fusion route based on orbital parameter exchange. This type of scheme exchanges orbital parameters (position vector, velocity vector, and covariance) between the two satellites in real time after the link is established. The relative orbital normal vector is calculated through coordinate transformation, and then dynamically weighted and fused with the static calibration results to generate a calibration benchmark corrected optical axis pointing. Its correction is based on the relative orbital geometry between the satellites, which can eliminate slow-changing pointing errors caused by orbital dynamics. However, its correction amount comes from orbital recursive data and it lacks the ability to sense and compensate for disturbances such as high-frequency micro-vibrations and attitude jitter of the satellite platform itself.

[0005] Thirdly, visual tracking combined with a two-stage feedback mechanism for alignment. This type of scheme uses a two-stage (coarse turntable + piezoelectric micro-motion stage) actuator to achieve coarse-to-fine pointing. At the receiving end, a four-quadrant detector outputs the position deviation of the light spot, which is then controlled by a position loop-velocity loop dual closed-loop (PID / PI) feedback control to drive the reflector to deflect and correct the deviation. Its control essence is based on feedback correction of measured errors, that is, "first detect the deviation, then correct the deviation."

[0006] While the above approaches each offer improvements, none have resolved the core challenge of line-of-sight alignment in high-dynamic platform environments for inter-satellite laser communication. Specifically: (1) It relies on feedback, is powerless against high-frequency disturbances, and has inherent phase lag. The correction action of existing feedback alignment (such as Chinese Invention Patent Publication No. CN122159963A and conventional FSM closed loop) occurs after the deviation is detected. The control loop itself has time delays in sampling, calculation, and execution. The flywheel, momentum wheel, solar panel drive mechanism, thermal deformation, etc. of the satellite platform can introduce micro-vibrations of tens to hundreds of hertz. When the disturbance frequency is close to or exceeds the feedback control bandwidth, the feedback loop cannot respond in time. A phase lag occurs between the correction amount and the actual deviation, resulting in large residual jitter, decreased alignment accuracy, or even link loss. The fundamental reason is that feedback control is "passive chasing" and cannot predict the arrival of disturbances.

[0007] (2) Rigidity depends on the FSM motor and mechanical precision, resulting in high engineering implementation costs and limited reliability. To pursue high precision within the feedback framework, traditional solutions can only continuously improve the resolution, stiffness, and response frequency of the FSM motor, imposing stringent requirements on the assembly and adjustment precision of the optomechanical structure. This leads to high payload weight, power consumption, and cost, and the gaps, hysteresis, and wear of mechanical components directly transmit pointing errors, making it difficult to meet the needs of low-cost, large-scale deployment of large-scale micro-nano satellite constellations. The reason is that existing solutions place almost all the burden of alignment accuracy on the "mechanical performance of the actuator" rather than the "intelligence of the control strategy".

[0008] (3) The error causes are perceived in a single way, lacking the fusion and prediction of multi-source information. For example, Chinese invention patent announcement number CN120495387B only uses spot images, and Chinese invention patent announcement number CN120834853B only uses orbital parameters. Both are single information sources. There are actually multiple sensors on the satellite that can sense the platform status, such as inertial measurement unit (IMU), star sensor, vibration sensor, temperature sensor, and spot detector. However, the existing technology does not fuse these multi-source heterogeneous data, nor does it use them to predict the line-of-sight offset at future moments, resulting in a large amount of status information that can be used for "advance compensation" being wasted.

[0009] (4) Low link establishment efficiency and poor stability maintenance. The above defects combined result in low acquisition probability, long link establishment cycle and slow recovery after link interruption in complex operating conditions such as rapid satellite maneuvering, multi-satellite coordination and no ground station coverage. This makes it difficult to meet the requirements of the new generation of low-Earth orbit constellations for high bandwidth, high reliability and strong autonomy of space-based networks.

[0010] Therefore, there is an urgent need for an alignment method that can proactively predict the impact of platform disturbances on the line of sight and compensate for deviations before they actually occur. This would break through the rigid dependence on the mechanical precision of the FSM at the control strategy level and achieve high-precision, robust inter-satellite laser alignment at the micro-radian level. Summary of the Invention

[0011] This invention proposes a multi-source sensing feedforward prediction and compensation method for the line-of-sight of an inter-satellite laser link. It sets passive feedback correction as a hybrid closed-loop architecture that combines feedforward prediction and compensation with feedback refinement. By fusing multi-source sensing data on the satellite, the method uses a prediction model to predict the line-of-sight offset caused by platform vibration / attitude disturbances in the next control cycle. It generates flexible compensation control quantities for the fast-reflecting mirror in advance and completes the compensation before the disturbance actually acts on the optical path, thereby breaking the rigid dependence of traditional laser communication on the mechanical precision of the motor.

[0012] A multi-source sensing feedforward prediction and compensation method for the line-of-sight of an inter-satellite laser link, designed for this purpose, includes the following steps: Step 1: Collect multi-source sensor data from the spaceborne laser terminal and its platform, align the multi-source sensor data with timestamps and coordinate system, and form a multi-channel time-series state vector. Step 2: Based on the multi-channel temporal state vector, use the pre-built lightweight temporal prediction model to perform feedforward prediction of the line-of-sight offset at future times, and generate the predicted line-of-sight offset. Step 3: Evaluate the prediction confidence based on the prediction residual of the lightweight time series prediction model within the recent window. The predicted feedforward compensation amount and the feedback correction amount calculated by the feedback controller based on the current measured error of the spot detector are weighted and fused. The feedforward weight and feedback weight are dynamically adjusted according to the prediction confidence to generate the total control amount of the fast-reflecting mirror. The total control amount drives the fast-reflecting mirror to complete the flexible compensation. Step 4: After completing the flexible compensation, new measured deviations are collected based on the data collected in Step 1. The measured deviations can enter the feedback loop of the next cycle and be compared with the predicted values ​​of the previous cycle. This is used for online evaluation and updating of the lightweight time series prediction model and weights, forming a rolling closed loop of data collection, prediction, hybrid compensation, and re-collection.

[0013] In step one, the multi-source sensor acquisition module synchronously acquires multi-source sensor data on the platform at a fixed period. The spatiotemporal registration module works in conjunction with the multi-source sensor acquisition module to perform timestamp alignment and coordinate transformation on the multi-source sensor data and output a multi-channel time-series state vector.

[0014] The multi-source sensing acquisition module includes: An inertial measurement unit, installed on the optomechanical base of the laser terminal, collects triaxial angular velocity and angular acceleration. Vibration sensors are placed along the critical vibration transmission path of the optomechanical structure to collect micro-vibrations of the platform. Star sensors provide satellite attitude quaternions; The gyroscope provides the rate of change of the satellite's attitude quaternion; The spot feedback detector is used to output the current line-of-sight deviation measurement value. In step four, the spot detector is used to collect new measured deviations, which are used for online evaluation and updating of the lightweight time-series prediction model and weights.

[0015] The input of the spatiotemporal registration module is connected to the multi-source sensor acquisition module 101, which performs timestamp alignment on the multi-source data, unifies it to the same clock reference, and uses interpolation to compensate for the sampling phase difference and coordinate transformation of each sensor. The coordinate transformation is performed by uniformly projecting the multi-source sensing data onto the azimuth and pitch axes of the optomechanical body coordinate system, and outputting a multi-channel time-series state vector. The output multi-channel timing state vector is: , where ω(k) is the triaxial angular velocity, avib(k) is the platform micro-vibration, q(k) is the satellite attitude quaternion, and Δx(k) and Δy(k) are the current spot deviations.

[0016] In step two, the registered multi-channel temporal state vector sequence over the past several sampling periods is input into the lightweight temporal prediction model. The lightweight temporal prediction model outputs the predicted offset of the line of sight in both azimuth and pitch directions. , ; in, This represents the predicted offset of the line of sight in the azimuth direction. Predict the offset of the line of sight in the pitch direction.

[0017] In step two, the method also includes connecting the spatiotemporal registration module through the input end of the prediction module 103, inputting the time window sequence into the deployed temporal prediction network, and outputting the predicted offset of the line of sight in the azimuth and pitch directions at future times. The predicted offset is converted into a feedforward compensation control quantity uff(k) through the calibrated offset-FSM control quantity mapping relationship. The output of the prediction module 103 is also connected to the hybrid control solution module.

[0018] In step two, the temporal prediction network is trained using historically acquired multi-source sensor sequences and corresponding measured line-of-sight offsets as sample pairs, with the mean square error between the predicted and measured offsets used as the training parameters. As a loss function; in-orbit real-time inference during the deployment phase, and incremental fine-tuning based on online residuals; in, Where N is the loss value and N is the sample size. This represents the line-of-sight offset predicted by the model. This represents the actual measured line-of-sight deviation.

[0019] In step three, one end of the hybrid control solution module receives the feedforward compensation control quantity uff(k) output by the prediction module, and the other end of the hybrid control solution module 104 receives the feedback correction quantity ufb(k) calculated by the feedback controller from the spot feedback detector. The total control quantity of the fast-reflection mirror is generated by fusing the control quantity in the following manner: ; according to Integrate two control signals; Dynamically adjust the feedforward weights based on the prediction confidence level. With feedback weights The total control quantity drives the fast-reflecting mirror to complete flexible compensation.

[0020] Step three also includes a fast-reflecting mirror drive module and a fast-reflecting mirror FSM; The hybrid control calculation module receives the feedforward from the prediction module at one end and the feedback from the spot feedback detector, calculated by the hybrid control calculation module, at the other end. These are then dynamically weighted and fused to output the total control quantity to the fast-reflection mirror drive module. In the feedback branch, the spot detector outputs the current measured deviation. The feedback correction amount is calculated by the hybrid control solution module. ; The fast-reflecting mirror drive module converts control signals into drive signals, which drive the fast-reflecting mirror of the FSM to deflect in azimuth and pitch dimensions, changing the output or receiving optical path and achieving line-of-sight compensation.

[0021] The evaluation method and weight adjustment method for the prediction confidence level described in step three include: Calculate the predicted residual within the recent window The sliding statistic, a smaller residual will improve When the residual exceeds the threshold, the reduction is applied. Revert to feedback-driven mode; Where wff is the feedforward weight, e prede (k) represents the prediction residual of the k-th control period, which is equal to the line-of-sight prediction offset within that period. Deviation from the measured line of sight of the spot detector The magnitude of the difference between them.

[0022] The beneficial technical effects of the present invention are as follows: 1. Overcoming the feedback lag bottleneck and significantly suppressing high-frequency micro-vibrations. Due to the use of feedforward predictive compensation, the correction action arrives before the disturbance, effectively increasing the control bandwidth and suppressing high-frequency platform jitter exceeding the bandwidth of traditional feedback control. The expected line-of-sight residual jitter (RMS) is significantly reduced compared to the pure feedback scheme, achieving alignment accuracy at the microradian level (µrad), meeting the requirements of inter-satellite laser link specifications of 2000–5000 km link establishment distance, communication rate ≥100 Gbps, and bit error rate (BER) < 1e-9.

[0023] 2. Breaking the rigid dependence on the mechanical precision of FSM motors reduces costs and weight. Because alignment accuracy is largely handled by "predictable control strategies" rather than simply relying on motor resolution and mechanical rigidity, high-precision alignment can be achieved on relatively low-cost actuators, which is beneficial for the lightweighting of micro- and nano-satellite payloads and the low-cost deployment of large-scale constellations.

[0024] 3. Multi-source information fusion, strong robustness and good adaptability. By fusing and predicting data from multiple sources such as IMU, star sensor, vibration sensor, and spot detector, and supplemented by a confidence-adaptive weighting mechanism, the system can maintain stable alignment even under complex conditions such as rapid satellite maneuvers, drastic attitude changes, and temporary failure of a single sensor, overcoming the shortcomings of single-source information schemes that are prone to failure.

[0025] 4. Improved link establishment efficiency and link reliability: High-precision, low-latency prediction and alignment shorten the acquisition and establishment time of inter-satellite links, improve the link acquisition probability and maintenance stability, and provide key support for autonomous operation in space-based intelligent networks, constellation networking and scenarios without ground station coverage.

[0026] 5. The method of this invention can be extended to other high-dynamic space pointing control fields such as high-precision directional antennas and inter-satellite measurement payloads, providing independent and controllable core technologies for low-orbit satellite internet, emergency communication, and maritime and aviation connections. Attached Figure Description

[0027] Figure 1 This is a block diagram of the overall structure of an inter-satellite laser communication line-of-sight alignment system according to an embodiment of the present invention.

[0028] Figure 2 This is a flowchart illustrating the overall process of a control method according to an embodiment of the present invention.

[0029] Figure 3 This is a block diagram illustrating the feedforward-feedback hybrid closed-loop control principle of an embodiment of the present invention.

[0030] Figure 4 This is a schematic diagram of the input and output structure of a prediction module according to an embodiment of the present invention.

[0031] Figure 5 This is a schematic diagram of the structure of a residual jitter time-domain comparison curve according to an embodiment of the present invention. Detailed Implementation

[0032] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. In order to make the above-mentioned objects, features and advantages of this application more apparent and understandable, many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0033] See Figures 1-5 The inter-satellite laser communication line-of-sight alignment system of this embodiment is deployed on a spaceborne laser terminal and includes: The multi-source sensing acquisition module 101 is connected to various sensors on the platform.

[0034] The spatiotemporal registration module 102 has its input terminal connected to the multi-source sensor acquisition module 101.

[0035] The prediction module 103 has its input connected to the spatiotemporal registration module, internally deploys a trained temporal prediction network, and outputs a feedforward prediction offset. Its output is connected to the hybrid control solution module 104.

[0036] The hybrid control calculation module 104 receives the feedforward quantity from the prediction module 103 at one end and the feedback quantity calculated by the feedback controller from the spot feedback detector 107 at the other end. It performs dynamic weighted fusion and outputs the total control quantity to the fast mirror drive module 105.

[0037] Fast-reflecting mirror drive module 105 and fast-reflecting mirror FSM106.

[0038] Each module can be implemented in real time by an onboard embedded computing unit (including processor, FPGA / GPU coprocessor, and memory), enabling prediction and control calculations to run on the unit.

[0039] The multi-source sensor acquisition module 101 acquires data at a fixed period. (e.g., 1 ms) synchronously acquire data from IMU, vibration sensor, star / gyroscope and spot detector.

[0040] The spatiotemporal registration module 102 performs time alignment and coordinate unification on the data from each channel to obtain a state vector sequence. ,in The length of the sliding time window (e.g., taking...) ).

[0041] Prediction module 103 inputs the above time window sequence into the temporal network and outputs the predicted line-of-sight offset for future time moments. , .

[0042] The prediction module 103 can employ a temporal convolutional network (TCN): taking multi-channel temporal data as input, extracting the temporal evolution features of perturbations through multiple layers of causal dilated convolutions, and outputting two-axis predicted offsets through fully connected layers. Alternatively, an LSTM network can be used instead.

[0043] The training phase of prediction module 103: using historically acquired multi-source sensor sequences and corresponding measured line-of-sight offsets as sample pairs, and using the mean square error between predicted and measured offsets as the loss function for training; in the deployment phase, the model performs real-time inference on-orbit and can be incrementally fine-tuned based on online residuals.

[0044] The predicted offset is converted into a feedforward compensation control quantity using the calibrated offset-FSM control quantity mapping relationship. .

[0045] Feedforward-feedback dynamic weighted fusion: In the feedback branch, the spot detector 107 outputs the current measured deviation. The feedback correction amount is calculated by the feedback controller (such as an incremental PID controller). Hybrid control solution module 104 The two control inputs are integrated. The dynamic weights are adjusted based on the prediction confidence level: the prediction residuals are calculated within the recent window. The sliding statistic, a smaller residual will improve (If taken) When the residual exceeds the threshold, reduce Revert to feedback-driven mode (to ensure safety in abnormal operating conditions).

[0046] FSM Flexible Compensation: Fast Reflective Mirror Drive Module 105 The signal is converted into a drive signal, which drives the fast-reflecting mirror 106 to deflect and complete the compensation. Since the compensation amount mainly comes from feedforward prediction and advance positioning, the dependence on the instantaneous mechanical response of the motor is greatly reduced, which is the so-called "flexible compensation".

[0047] Closed-loop iteration and online update: After compensation, the spot detector 107 collects new deviations, one path enters the next cycle feedback loop, and the other path is used to evaluate the prediction residuals, update the weights, and trigger online fine-tuning of the model. It then returns to the step of feedforward prediction of the next moment's line-of-sight offset by the multi-source data acquisition and prediction module 103, forming a rolling closed loop, continuously converging and stabilizing the line-of-sight alignment error at the microradian level.

[0048] See Figures 1-5 A multi-source sensing feedforward prediction and compensation method for the line-of-sight of an inter-satellite laser link includes the following steps: Step 1: Collect multi-source sensor data from the spaceborne laser terminal and its platform, align the multi-source sensor data with timestamps and coordinate system, and form a multi-channel time-series state vector. Step 2: Based on the multi-channel temporal state vector, use the pre-built lightweight temporal prediction model to perform feedforward prediction of the line-of-sight offset at future times, and generate the predicted line-of-sight offset. Step 3: Evaluate the prediction confidence based on the prediction residual of the lightweight time series prediction model within the recent window. The predicted feedforward compensation amount and the feedback correction amount calculated by the feedback controller based on the current measured error of the spot detector are weighted and fused. The feedforward weight and feedback weight are dynamically adjusted according to the prediction confidence to generate the total control amount of the fast-reflecting mirror. The total control amount drives the fast-reflecting mirror to complete the flexible compensation. Step 4: After completing the flexible compensation, new measured deviations are collected based on the data collected in Step 1. The measured deviations can enter the feedback loop of the next cycle and be compared with the predicted values ​​of the previous cycle. This is used for online evaluation and updating of the lightweight time series prediction model and weights, forming a rolling closed loop of data collection, prediction, hybrid compensation, and re-collection.

[0049] By using multi-source sensing and feedforward prediction, early compensation for line-of-sight offset is achieved, breaking the lag limitation of traditional feedback control and improving the ability to suppress high-frequency disturbances. At the same time, the confidence-adaptive weighting mechanism ensures the robustness of the system under abnormal operating conditions.

[0050] The lightweight temporal prediction model is a temporal convolutional network (TCN), a long short-term memory network (LSTM), or a pruned variant thereof. Its input is the sequence of the registered multi-channel temporal state vector described in point one over the past several sampling periods (sliding time window), and its output is the predicted offset of the line of sight in both azimuth and pitch directions over one or more future control periods. , The model is trained on the ground or in orbit using paired samples of historical "sensor data - measured line-of-sight offset" to learn the dynamic mapping relationship between platform disturbance signals and line-of-sight offset. It predicts deviations that are about to occur but have not yet been detected by the detector, which is the essential difference between it and all passive feedback schemes.

[0051] The predicted feedforward compensation amount The feedback correction amount is calculated by a feedback controller (such as a PID controller) based on the current measured error of the spot detector. Perform weighted fusion to generate the total control quantity of the fast mirror (FSM): ,in , The weights are dynamic. Furthermore, this invention sets up a confidence-adaptive weighting mechanism: the prediction confidence is evaluated based on the prediction residuals (the difference between predicted and measured values) of the prediction model within the recent window, and the prediction confidence is adjusted accordingly. The larger the value, the more the system relies on feedforward prediction; it automatically reduces this when the prediction residual increases (e.g., when encountering anomalous perturbations outside the training distribution). The system reverts to a feedback-based approach to ensure robustness and safety. The overall control input drives the FSM to perform flexible compensation. This flexibility is reflected in the fact that compensation no longer relies on the instantaneous mechanical response of the motor to catch up with deviations, but rather on predictive, pre-emptive adjustments, significantly reducing the requirements for motor response speed and mechanical precision.

[0052] The spot detector collects new measured deviations, which are then fed into the feedback loop of the next cycle. On the other hand, they are compared with the predicted values ​​of the previous cycle for online evaluation / updating of the prediction model and weights, forming a rolling closed loop of "acquisition-prediction-hybrid compensation-reacquisition". This enables continuous, stable, and high-precision tracking of dynamic line-of-sight offset, ultimately converging the alignment accuracy to the microradian level.

[0053] In step one, the multi-source sensor acquisition module 101 synchronously acquires multi-source sensor data on the platform at fixed intervals. The spatiotemporal registration module 102 works in conjunction with the multi-source sensor acquisition module 101 to perform timestamp alignment and coordinate transformation on the multi-source sensor data, outputting a multi-channel time-series state vector. Through the coordinated operation of the multi-source sensor acquisition module 101 and the spatiotemporal registration module 102, the prediction model is provided with full-dimensional, spatiotemporally consistent platform perturbation information input.

[0054] The multi-source sensing acquisition module 101 includes: An inertial measurement unit, installed on the optomechanical base of the laser terminal, collects triaxial angular velocity and angular acceleration. Vibration sensors are placed along the critical vibration transmission path of the optomechanical structure to collect micro-vibrations of the platform. Star sensors provide satellite attitude quaternions; The gyroscope provides the rate of change of the satellite's attitude quaternion; The light spot feedback detector 107 is used to output the current line-of-sight deviation measurement value. In step four, the light spot detector is used to collect new measured deviations, which are used for online evaluation and updating of the lightweight time-series prediction model and weights.

[0055] Specifically, the inertial measurement unit (IMU) includes a three-axis gyroscope and a three-axis accelerometer; vibration sensors (MEMS accelerometer arrays) are arranged along the critical vibration transmission path of the optomechanical structure to collect micro-vibrations of the platform. The star sensor / gyroscope combination provides satellite attitude quaternions. and its rate of change; the spot feedback detector 107 (four-quadrant detector QPD or high frame rate CCD) collects the current spot deviation of the downlink / beacon light on the detector. A sampling frequency of ≥1 kHz (IMU / vibration) is recommended to cover the platform's vibration frequency band. By fusing multi-source heterogeneous data from IMU, vibration, star sensor, gyroscope, and spot detector, comprehensive perception of multi-dimensional disturbances such as high-frequency vibration and attitude perturbation of the platform can be achieved.

[0056] The input terminal of the spatiotemporal registration module 102 is connected to the multi-source sensor acquisition module 101 to perform timestamp alignment on the multi-source data, unify it to the same clock reference, and use interpolation to compensate for the sampling phase difference and coordinate transformation of each sensor. The coordinate transformation is performed by uniformly projecting the multi-source sensing data onto the azimuth and pitch axes of the optomechanical body coordinate system, and outputting a multi-channel time-series state vector. The output multi-channel timing state vector is: , where ω(k) is the triaxial angular velocity, avib(k) is the platform micro-vibration, q(k) is the satellite attitude quaternion, and Δx(k) and Δy(k) are the current spot deviations.

[0057] The sampling phase difference between various sensors is eliminated by aligning the timestamps, and the multi-source data is unified into the optomechanical body coordinate system by coordinate transformation, so as to ensure the accuracy and consistency of the input of the subsequent prediction model.

[0058] In step two, the registered multi-channel temporal state vector sequence over the past several sampling periods is input into the lightweight temporal prediction model. The lightweight temporal prediction model outputs the predicted offset of the line of sight in both azimuth and pitch directions. , ; in, This represents the predicted offset of the line of sight in the azimuth direction. Predict the offset of the line of sight in the pitch direction.

[0059] By introducing a time-series prediction model, future line-of-sight offsets can be predicted using historical state sequences, thus upgrading from "passive feedback" to "active feedforward" and breaking through the limitations of passive feedback.

[0060] In step two, the method also includes connecting the spatiotemporal registration module 102 through the input end of the prediction module 103, inputting the time window sequence into the deployed temporal prediction network, and outputting the predicted offset of the line of sight in the azimuth and pitch directions at future times. The predicted offset is converted into a feedforward compensation control quantity uff(k) through the calibrated offset-FSM control quantity mapping relationship. The output of the prediction module 103 is also connected to the hybrid control solution module 104.

[0061] The prediction module generates the feedforward compensation amount in real time and maps the predicted offset to the FSM control quantity, providing a basis for advance compensation.

[0062] In step two, the temporal prediction network is trained using historically acquired multi-source sensor sequences and corresponding measured line-of-sight offsets as sample pairs, with the mean square error between the predicted and measured offsets used as the training parameters. As a loss function; in-orbit real-time inference during the deployment phase, and incremental fine-tuning based on online residuals; in, Where N is the loss value and N is the sample size. This represents the line-of-sight offset predicted by the model. This represents the actual measured line-of-sight deviation.

[0063] By combining ground training with on-orbit incremental fine-tuning, the model continuously adapts to changes in satellite on-orbit operating conditions, maintaining prediction accuracy.

[0064] In step three, one end of the hybrid control solution module 104 receives the feedforward compensation control quantity uff(k) output by the prediction module 103, and the other end of the hybrid control solution module 104 receives the feedback correction quantity ufb(k) calculated by the feedback controller from the spot feedback detector 107. The total control quantity of the fast-reflection mirror is generated by fusing the control quantity in the following manner: ; And in accordance with Integrate two control signals; Dynamically adjust the feedforward weights based on the prediction confidence level. With feedback weights The total control quantity drives the fast-reflecting mirror to complete flexible compensation.

[0065] By weighted fusion of feedforward and feedback, combined with an adaptive confidence weighting mechanism, the system primarily uses feedforward to improve accuracy when predictions are accurate, and primarily uses feedback to ensure safety when predictions are inaccurate.

[0066] Step three also includes a fast-reflecting mirror drive module 105 and a fast-reflecting mirror FSM 106; The hybrid control calculation module 104 receives the feedforward quantity from the prediction module 103 at one end and the feedback quantity calculated by the spot feedback detector 107 at the other end. It performs dynamic weighted fusion to output the total control quantity to the fast-reflection mirror drive module 105. In the feedback branch, the spot detector 107 outputs the current measured deviation. The feedback correction amount is calculated by the hybrid control calculation module 104. ;; The fast-reflecting mirror drive module 105 converts the control quantity into a drive signal, which drives the fast-reflecting mirror FSM 106 to deflect in both azimuth and pitch dimensions, thereby changing the output or receiving optical path and achieving line-of-sight compensation.

[0067] Through the collaboration between the fast-reflecting mirror drive module and the FSM, the total control quantity is converted into actual line-of-sight pointing adjustment, thus completing the physical compensation for platform disturbances.

[0068] The evaluation method and weight adjustment method for the prediction confidence level described in step three include: Calculate the predicted residual within the recent window The sliding statistic, a smaller residual will improve When the residual exceeds the threshold, the reduction is applied. Revert to feedback-driven mode; Where wff is the feedforward weight, e prede (k) represents the prediction residual of the k-th control period, which is equal to the line-of-sight prediction offset within that period. Deviation from the measured line of sight of the spot detector The magnitude of the difference between them.

[0069] The prediction confidence is evaluated in real time by using the sliding statistics of the predicted residuals, and the feedforward and feedback weights are dynamically adjusted accordingly to ensure that the system can still operate stably under abnormal disturbances.

[0070] The aforementioned prediction module can be replaced with any timing prediction structure such as a gated recurrent unit (GRU) or a lightweight variant of the Transformer; the input channels can be added or removed according to the actual onboard sensor configuration, but must include at least one sensor that reflects the platform motion status and a spot detector.

[0071] The fusion of feedforward and feedback can be extended from weighted summation to optimal fusion based on Kalman filtering / complementary filtering.

[0072] See Figure 5The horizontal axis represents the simulation time (0-2s), and the vertical axis represents the line-of-sight jitter error (μrad). The pure PID feedback curve exhibits the largest jitter amplitude and the most severe fluctuations; the track feedforward scheme can only suppress the slow variation trend, and high-frequency fluctuations are still obvious; the curve of the scheme in this invention is stable throughout, with a jitter amplitude significantly lower than the previous two schemes, and no obvious peak fluctuations.

[0073] The above are merely preferred embodiments of the present invention and are not intended to limit 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 multi-source perception feedforward prediction compensation of inter-satellite laser link boresight, characterized in that, Includes the following steps: Step 1: Collect multi-source sensor data from the spaceborne laser terminal and its platform, align the multi-source sensor data with timestamps and coordinate system, and form a multi-channel time-series state vector. Step 2: Based on the multi-channel temporal state vector, use the pre-built lightweight temporal prediction model to perform feedforward prediction of the line-of-sight offset at future times, and generate the predicted line-of-sight offset. Step 3: Evaluate the prediction confidence based on the prediction residual of the lightweight time series prediction model within the recent window. The predicted feedforward compensation amount and the feedback correction amount calculated by the feedback controller based on the current measured error of the spot detector are weighted and fused. The feedforward weight and feedback weight are dynamically adjusted according to the prediction confidence to generate the total control amount of the fast-reflecting mirror. The total control amount drives the fast-reflecting mirror to complete the flexible compensation. Step 4: After completing the flexible compensation, new measured deviations are collected based on the data collected in Step 1. The measured deviations can enter the feedback loop of the next cycle and be compared with the predicted values ​​of the previous cycle. This is used for online evaluation and updating of the lightweight time series prediction model and weights, forming a rolling closed loop of data collection, prediction, hybrid compensation, and re-collection.

2. The multi-source sensing feedforward prediction and compensation method for the line of sight of an inter-satellite laser link according to claim 1, characterized in that: In step one, the multi-source sensor acquisition module (101) synchronously acquires multi-source sensor data on the platform at a fixed period. The spatiotemporal registration module (102) works in conjunction with the multi-source sensor acquisition module (101) to perform timestamp alignment and coordinate transformation on the multi-source sensor data and output a multi-channel time-series state vector.

3. The multi-source sensing feedforward prediction and compensation method for the line of sight of an inter-satellite laser link according to claim 2, characterized in that, The multi-source sensing acquisition module (101) includes: An inertial measurement unit, installed on the optomechanical base of the laser terminal, collects triaxial angular velocity and angular acceleration. Vibration sensors are placed along the critical vibration transmission path of the optomechanical structure to collect micro-vibrations of the platform. Star sensors provide satellite attitude quaternions; The gyroscope provides the rate of change of the satellite's attitude quaternion; The spot feedback detector (107) is used to output the current line-of-sight deviation measurement value. In step four, the spot detector is used to collect new measured deviations, which are used for online evaluation and updating of the lightweight time-series prediction model and weights.

4. The multi-source sensing feedforward prediction and compensation method for the line of sight of an inter-satellite laser link according to claim 2, characterized in that: The input of the spatiotemporal registration module (102) is connected to the multi-source sensor acquisition module (101) to perform timestamp alignment on the multi-source data, unify it to the same clock reference, and use interpolation to compensate for the sampling phase difference and coordinate transformation of each sensor. The coordinate transformation is performed by uniformly projecting the multi-source sensing data onto the azimuth and pitch axes of the optomechanical body coordinate system, and outputting a multi-channel time-series state vector. The output multi-channel timing state vector is: , where ω(k) is the triaxial angular velocity, avib(k) is the platform micro-vibration, q(k) is the satellite attitude quaternion, and Δx(k) and Δy(k) are the current spot deviations.

5. The multi-source sensing feedforward prediction and compensation method for the line of sight of an inter-satellite laser link according to claim 2, characterized in that: In step two, the registered multi-channel temporal state vector sequence over the past several sampling periods is input into the lightweight temporal prediction model. The lightweight temporal prediction model outputs the predicted offset of the line of sight in both azimuth and pitch directions. , ; in, This represents the predicted offset of the line of sight in the azimuth direction. This is the predicted offset of the line of sight in the pitch direction.

6. The multi-source sensing feedforward prediction and compensation method for the line of sight of an inter-satellite laser link according to claim 5, characterized in that: In step two, the method also includes connecting the spatiotemporal registration module (102) to the input end of the prediction module (103), inputting the time window sequence into the deployed temporal prediction network, and outputting the predicted offset of the line of sight in the azimuth and pitch directions at future times. The predicted offset is converted into a feedforward compensation control quantity uff(k) through the calibrated offset-FSM control quantity mapping relationship. The output of the prediction module (103) is also connected to the hybrid control solution module (104).

7. The multi-source sensing feedforward prediction and compensation method for the line of sight of an inter-satellite laser link according to claim 6, characterized in that, In step two, the temporal prediction network is trained using historically acquired multi-source sensor sequences and corresponding measured line-of-sight offsets as sample pairs, with the mean square error between the predicted and measured offsets used as the training parameters. As a loss function; in-orbit real-time inference during the deployment phase, and incremental fine-tuning based on online residuals; in, Where N is the loss value and N is the sample size. This represents the line-of-sight offset predicted by the model. This represents the actual measured line-of-sight deviation.

8. The multi-source sensing feedforward prediction and compensation method for the line of sight of an inter-satellite laser link according to claim 6, characterized in that: In step three, one end of the hybrid control solution module (104) receives the feedforward compensation control quantity uff(k) output by the prediction module (103), and the other end of the hybrid control solution module (104) receives the feedback correction quantity ufb(k) calculated by the feedback controller from the spot feedback detector (107). The total control quantity of the fast-reflection mirror is generated by fusing the control quantity in the following manner: and in accordance with Integrate two control signals; Dynamically adjust the feedforward weights based on the prediction confidence level. With feedback weights The total control quantity drives the fast-reflecting mirror to complete flexible compensation.

9. The multi-source sensing feedforward prediction and compensation method for the line of sight of an inter-satellite laser link according to claim 8, characterized in that: Step 3 also includes a fast-reflecting mirror drive module (105) and a fast-reflecting mirror FSM (106). One end of the hybrid control calculation module (104) receives the feedforward quantity from the prediction module (103), and the other end receives the feedback quantity calculated by the hybrid control calculation module (104) from the spot feedback detector (107). The two ends are dynamically weighted and fused to output the total control quantity to the fast-reflection mirror drive module (105). On the feedback branch, the spot detector (107) outputs the current measured deviation. The feedback correction amount is calculated by the hybrid control calculation module (104). ; The fast-reflecting mirror drive module (105) converts the control quantity into a drive signal, which drives the fast-reflecting mirror of the fast-reflecting mirror FSM (106) to deflect in the azimuth and pitch dimensions, changing the output or receiving optical path and realizing line-of-sight compensation.

10. The multi-source sensing feedforward prediction and compensation method for the line of sight of an inter-satellite laser link according to claim 1, characterized in that, The evaluation method and weight adjustment method for the prediction confidence level described in step three include: Calculate the predicted residual within the recent window The sliding statistic, a smaller residual will improve When the residual exceeds the threshold, the reduction is applied. Revert to feedback-driven mode; Where wff is the feedforward weight, e prede (k) represents the prediction residual of the k-th control period, which is equal to the line-of-sight prediction offset within that period. Deviation from the measured line of sight of the spot detector The magnitude of the difference between them.

Images