Ground test method and system for satellite laser communication aiming time delay tolerance characteristic

Abstract

The invention provides a ground test method and system for satellite laser communication aiming time delay tolerance characteristics, and relates to the technical field of laser communication, and the method comprises the steps: obtaining an aiming time delay prediction model through constructing a deep learning sample set training neural network, and predicting a time delay value according to a real-time aiming parameter, a time delay adaptive factor dynamic adjustment measurement noise covariance matrix is constructed and is combined with a multi-scale strong tracking factor, attitude tracking parameters are optimized by adopting a recursive least square algorithm, time delay tolerance is calculated, dynamic adjustment of system parameters is realized, and aiming precision and reliability of a satellite laser communication system are improved.

Description

Technical Field

[0001] The present invention relates to laser communication technology, and in particular to a ground test method and system for the aiming time-delay tolerance characteristics of satellite laser communication. Background Art

[0002] As an important space information transmission method, satellite laser communication has advantages such as high bandwidth, good confidentiality, and strong anti-interference ability, and has broad application prospects in military and civilian fields. A satellite laser communication system requires precise aiming technology to ensure that the laser beam can be accurately projected onto the receiving end. However, due to factors such as satellite orbital motion, atmospheric disturbance, and response delay of the attitude control system, there are inevitably time-delay problems during the aiming process, which have a serious impact on communication quality and stability.

[0003] The testing of the aiming time-delay tolerance characteristics of satellite laser communication mostly relies on actual satellite on-orbit testing, which is costly and time-consuming. Although the ground test method is economically feasible, the existing technologies have the following deficiencies:

[0004] Traditional aiming time-delay prediction methods mostly rely on empirical models or simple mathematical models, and cannot effectively handle the complex non-linear time-varying characteristics in satellite laser communication systems, resulting in low prediction accuracy and inability to adapt to the real-time changing space environment.

[0005] Existing measurement noise processing methods lack adaptability. Usually, a fixed noise covariance matrix is used, and it is unable to dynamically adjust according to the system characteristics under different time-delay conditions, resulting in a lack of real-time calculation and dynamic compensation mechanism for the aiming time-delay tolerance in complex traditional satellite laser communication systems, and inability to adaptively adjust the transmitting beam parameters according to the actual communication environment, making it difficult to maintain the stability and reliability of the communication link under large time-delay fluctuations. Summary of the Invention

[0006] Embodiments of the present invention provide a ground test method and system for the aiming time-delay tolerance characteristics of satellite laser communication, which can solve the problems in the existing technologies.

[0007] In the first aspect of the embodiments of the present invention, a ground test method for the aiming time-delay tolerance characteristics of satellite laser communication is provided, including:

[0008] Obtain the historical aiming data of the satellite laser communication system, construct a deep learning training sample set according to the historical aiming data, and train a preset deep neural network using the deep learning training sample set to obtain an aiming time-delay prediction model;

[0009] Collect the real-time aiming parameters of the current satellite laser communication system, input the real-time aiming parameters into the aiming time-delay prediction model, and obtain the predicted aiming time-delay value;

[0010] Construct a time-delay adaptive factor based on the predicted aiming time delay value, obtain a dynamically adjusted measurement noise covariance matrix based on the time-delay adaptive factor, and combine the dynamically adjusted measurement noise covariance matrix with a multi-scale strong tracking factor; use the recursive least squares estimation algorithm to perform real-time optimization compensation on the attitude tracking control parameters of the satellite laser communication system, where the measurement noise covariance matrix of the recursive least squares estimation algorithm is dynamically adjusted according to the predicted aiming time delay value, and the Kalman gain is estimated to obtain an optimal state estimation value with enhanced robustness;

[0011] Based on the optimal state estimation value and the attitude tracking control parameters after real-time optimization compensation, use the recursive least squares algorithm to calculate the aiming time delay tolerance in real time, and dynamically adjust the transmitted beam power and beam divergence angle of the satellite laser communication system according to the aiming time delay tolerance.

[0012] Construct a deep learning training sample set according to the historical aiming data, and use the deep learning training sample set to train a preset deep neural network to obtain an aiming time delay prediction model, including:

[0013] Calculate the time delay mean and time delay standard deviation according to the historical aiming data, segment and sample the historical aiming data using a sliding time window, and generate a time delay sample sequence by adjusting the length of the sliding time window;

[0014] Determine an anchor sample selection threshold based on the time delay mean and the time delay standard deviation, use the samples in the time delay sample sequence that are less than the anchor sample selection threshold as anchor samples, and apply corresponding enhancement transformation parameters to the anchor samples to generate a positive sample pair sequence;

[0015] Construct a deep feature extraction network, use the deep feature extraction network to perform feature encoding on the positive sample pair sequence to obtain positive sample pair feature vectors; calculate the cosine similarity between the positive sample pair feature vectors, and construct a contrast loss function according to the cosine similarity;

[0016] Perform weighted combination on the contrast loss function and a time delay prediction loss function to obtain a hybrid loss function, where the time delay prediction loss function is used to evaluate the error between the time delay prediction value and the true time delay value; calculate the gradient of the network parameters based on the hybrid loss function, and use the gradient to update the parameters of the deep feature extraction network to obtain an aiming time delay prediction model.

[0017] Input the real-time aiming parameters into the aiming time delay prediction model to obtain the predicted aiming time delay value, including:

[0018] Construct an input feature vector based on the real-time aiming parameters, where the real-time aiming parameters include attitude parameters and angular velocity parameters, and normalize the input feature vector to obtain a standardized feature vector;

[0019] Input the standardized feature vector into a pre-trained aiming time-delay prediction model, which is a deep neural network trained based on a contrast learning mechanism, to obtain a predicted aiming time-delay feature.

[0020] Based on the predicted aiming time-delay value, use the recursive least squares estimation algorithm to perform real-time optimization compensation on the attitude tracking control parameters of the satellite laser communication system. The measurement noise covariance matrix of the recursive least squares estimation algorithm is dynamically adjusted according to the predicted aiming time-delay value, including:

[0021] Construct a time-delay adaptive factor according to the predicted aiming time-delay value, and perform weighted fusion on the time-delay adaptive factor and the initial value of the measurement noise covariance matrix to obtain a dynamically adjusted measurement noise covariance matrix;

[0022] Calculate the asymptotic optimized value of the prediction error covariance matrix using the recursive least squares estimation algorithm, and combine the asymptotic optimized value with the sampling weights of the Gauss-Hermite transform to obtain an adjusted prediction error covariance matrix;

[0023] Calculate a multi-scale strong tracking factor based on the adjusted prediction error covariance matrix, combine the multi-scale strong tracking factor with the dynamically adjusted measurement noise covariance matrix to obtain a gain correction matrix, and use the gain correction matrix to perform online correction on the Kalman gain;

[0024] Estimate the system state based on the result of the corrected Kalman gain, and perform information fusion on the system state and the result of the corrected Kalman filter to obtain an optimal state estimate value with enhanced robustness;

[0025] Construct a performance index according to the optimal state estimate value, and design an anti-interference control compensation strategy based on the performance index to achieve real-time optimization compensation of the attitude tracking control parameters.

[0026] Calculate the asymptotic optimized value of the prediction error covariance matrix using the recursive least squares estimation algorithm, and combine the asymptotic optimized value with the sampling weights of the Gauss-Hermite transform to obtain an adjusted prediction error covariance matrix, including:

[0027] Construct an initial prediction error covariance matrix including a historical prediction error vector and a process noise covariance matrix, comprising: multiplying the historical prediction error vector by an exponential power term of a forgetting factor and summing to obtain an error component, and adding the error component to the process noise covariance matrix to obtain the initial prediction error covariance matrix;

[0028] Construct a recursive least squares optimization objective function based on the initial prediction error covariance matrix, the recursive least squares optimization objective function including a weighted norm term of the historical prediction error vector, calculate the gradient of the recursive least squares optimization objective function, and multiply the gradient by an optimization gain and then superimpose it on the initial prediction error covariance matrix to obtain an optimized prediction error covariance matrix;

[0029] Obtain a system state mean, add the product of the system state mean, a scaled square root term of the optimized prediction error covariance matrix, and a sampling vector to generate a set of sampling points, and calculate mean weights and covariance weights corresponding to the set of sampling points based on the system state dimension and a scaling parameter;

[0030] Multiply and sum the quadratic form of each sampling point in the set of sampling points with the mean weights and the covariance weights to obtain a covariance compensation term, and calculate a negative exponential function of the ratio of the covariance compensation term to the Frobenius norm of the optimized prediction error covariance matrix to obtain an adaptive fusion weight coefficient;

[0031] Perform a weighted combination of the covariance compensation term and the optimized prediction error covariance matrix according to the adaptive fusion weight coefficient to obtain an adjusted prediction error covariance matrix.

[0032] Based on the optimal state estimate value, according to the attitude tracking control parameters after real-time optimization compensation, use a robust adaptive Kalman filtering algorithm to calculate the aiming time delay tolerance in real time, and dynamically adjust the transmission beam power and beam divergence angle of the satellite laser communication system, including:

[0033] Establish a dynamic model of time delay tolerance, set the state equation of the dynamic model of time delay tolerance as the sum of a nonlinear function of the previous moment's time delay tolerance, the optimal state estimate value, and the attitude tracking control parameters after real-time optimization compensation and process noise, and set the observation equation of the dynamic model of time delay tolerance as the sum of a nonlinear function of the current moment's time delay tolerance and measurement noise;

[0034] Calculate the prediction error covariance matrix of the previous moment's time delay tolerance and the current moment's time delay tolerance based on the state equation and the observation equation;

[0035] Construct an innovation covariance matrix, calculate an adaptive factor according to the ratio of the product of the innovation covariance matrix and the prediction error covariance matrix, and use the result of the combination operation of the adaptive factor and the prediction error covariance matrix as a gain matrix;

[0036] Based on the gain matrix and the deviation between the observed value and the predicted value, calculate the optimal prediction estimate of the delay tolerance. Substitute the deviation between the optimal prediction estimate and the target delay tolerance into an exponential function to calculate the adjustment amount of the transmitted beam power. Substitute the ratio of the optimal prediction estimate and the prediction error covariance matrix into a hyperbolic tangent function to calculate the adjustment amount of the beam divergence angle;

[0037] Multiply the adjustment amount of the transmitted beam power by the nominal power to obtain the optimized transmitted beam power, and add the adjustment amount of the beam divergence angle to the reference divergence angle to obtain the optimized beam divergence angle.

[0038] In the second aspect of the embodiments of the present invention, a ground test system for the delay tolerance characteristics of satellite laser communication aiming is provided, including:

[0039] A first unit for obtaining historical aiming data of a satellite laser communication system, constructing a deep learning training sample set according to the historical aiming data, and training a preset deep neural network using the deep learning training sample set to obtain an aiming delay prediction model;

[0040] A second unit for collecting real-time aiming parameters of the current satellite laser communication system, and inputting the real-time aiming parameters into the aiming delay prediction model to obtain a predicted aiming delay value;

[0041] A third unit for constructing a delay adaptive factor based on the predicted aiming delay value, obtaining a dynamically adjusted measurement noise covariance matrix based on the delay adaptive factor, and combining the dynamically adjusted measurement noise covariance matrix with a multi-scale strong tracking factor; using a recursive least squares estimation algorithm to perform real-time optimization compensation on the attitude tracking control parameters of the satellite laser communication system, where the measurement noise covariance matrix of the recursive least squares estimation algorithm is dynamically adjusted according to the predicted aiming delay value, and estimating the Kalman gain to obtain an optimal state estimate value with enhanced robustness;

[0042] A fourth unit for calculating the aiming delay tolerance in real time using a robust adaptive Kalman filtering algorithm based on the optimal state estimate value according to the attitude tracking control parameters after real-time optimization compensation, and dynamically adjusting the transmitted beam power and beam divergence angle of the satellite laser communication system according to the aiming delay tolerance.

[0043] In the third aspect of the embodiments of the present invention, an electronic device is provided, including:

[0044] A processor;

[0045] A memory for storing instructions executable by the processor;

[0046] Wherein, the processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0047] In a fourth aspect of the embodiments of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0048] The beneficial effects of this application are as follows:

[0049] By constructing a aiming delay prediction model through deep learning, the present invention can accurately predict the aiming delay value of a satellite laser communication system, avoiding the problem of low prediction accuracy caused by traditional methods relying on empirical parameters, and improving the accuracy of aiming delay prediction.

[0050] Combining a dynamically adjusted measurement noise covariance matrix with a multi-scale strong tracking factor, the present invention uses a recursive least squares estimation algorithm to perform real-time optimization and compensation on the attitude tracking control parameters of a satellite laser communication system, enhancing the system's adaptability to environmental changes and model errors, and improving the robustness and stability of the system operation.

[0051] Based on the optimal state estimate value, the present invention calculates the aiming delay tolerance in real time, and dynamically adjusts the transmitted beam power and beam divergence angle accordingly, enabling efficient utilization of energy while ensuring communication quality, extending the service life of the satellite, and at the same time improving the adaptability and reliability of the satellite laser communication system in a complex space environment. Description of the Drawings

[0052] Figure 1 It is a schematic flow chart of the ground test method for the aiming delay tolerance characteristics of the satellite laser communication in the embodiments of the present invention;

[0053] Figure 2 It is a flow chart of the delay prediction method based on deep feature extraction in the embodiments of the present invention;

[0054] Figure 3 It is a comparison chart of the attitude tracking accuracy based on delay prediction in the embodiments of the present invention;

[0055] Figure 4 It is a flow chart of the beam parameter regulation based on delay tolerance adaptive optimization in the embodiments of the present invention. Detailed Embodiments

[0056] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are only a part rather than all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0057] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0058] Figure 1 It is a schematic flowchart of the ground test method for the time-delay tolerance characteristics during satellite laser communication aiming in the embodiments of the present invention. As Figure 1 shown, the method includes:

[0059] Obtain the historical aiming data of the satellite laser communication system, construct a deep learning training sample set according to the historical aiming data, and use the deep learning training sample set to train a preset deep neural network to obtain a time-delay prediction model for aiming;

[0060] Collect the real-time aiming parameters of the current satellite laser communication system, input the real-time aiming parameters into the time-delay prediction model for aiming to obtain a predicted time-delay value for aiming;

[0061] Based on the predicted time-delay value for aiming, construct a time-delay adaptation factor, obtain a dynamically adjusted measurement noise covariance matrix based on the time-delay adaptation factor, and combine the dynamically adjusted measurement noise covariance matrix with a multi-scale strong tracking factor; use the recursive least squares estimation algorithm to perform real-time optimization and compensation on the attitude tracking control parameters of the satellite laser communication system, where the measurement noise covariance matrix of the recursive least squares estimation algorithm is dynamically adjusted according to the predicted time-delay value for aiming, and estimate the Kalman gain to obtain an optimal state estimation value with enhanced robustness;

[0062] Based on the optimal state estimation value and the attitude tracking control parameters after real-time optimization and compensation, use the recursive least squares algorithm to calculate the time-delay tolerance for aiming in real time, and dynamically adjust the transmitted beam power and beam divergence angle of the satellite laser communication system according to the time-delay tolerance for aiming.

[0063] In an alternative embodiment, constructing a deep learning training sample set according to the historical aiming data and using the deep learning training sample set to train a preset deep neural network to obtain a time-delay prediction model for aiming includes:

[0064] Calculate the mean delay and the standard deviation of the delay based on the historical aiming data, segment and sample the historical aiming data using a sliding time window, and generate a delay sample sequence by adjusting the length of the sliding time window;

[0065] Determine the anchor sample selection threshold based on the mean delay and the standard deviation of the delay, take the samples in the delay sample sequence that are less than the anchor sample selection threshold as anchor samples, and apply corresponding enhancement transformation parameters to the anchor samples to generate a positive sample pair sequence;

[0066] Construct a deep feature extraction network, use the deep feature extraction network to perform feature encoding on the positive sample pair sequence to obtain positive sample pair feature vectors; calculate the cosine similarity between the positive sample pair feature vectors, and construct a contrast loss function according to the cosine similarity;

[0067] Perform weighted combination on the contrast loss function and the delay prediction loss function to obtain a hybrid loss function, where the delay prediction loss function is used to evaluate the error between the delay prediction value and the true delay value; calculate the gradient of the network parameters based on the hybrid loss function, and use the gradient to update the parameters of the deep feature extraction network to obtain an aiming delay prediction model.

[0068] When processing historical aiming data, first calculate the mean delay and the standard deviation of the delay. Assume that the historical aiming data set is D = {d1, d2,..., d n}, where each d contains aiming-related features and the corresponding delay value t. Calculate the average value μ of all delay values. For example, for a data set containing 1000 records, μ = 150ms is calculated. Then calculate the standard deviation σ of the delay. For example, σ = 30ms, which represents the degree of dispersion of the delay values.

[0069] Segment and sample the historical aiming data using a sliding time window. Set the initial window length w = 10, that is, each sample sequence contains 10 consecutive delay records. The sliding step size s = 1, indicating that it moves forward 1 time unit each time. In this way, n - w + 1 delay sample sequences {S1, S2,..., S n₋ᵥ₊1} can be generated from the original data set D. To find the optimal window length, different w values (such as 8, 10, 12, 15) can be tried for experiments, and it is determined by evaluating the performance of the final model. Finally, w = 12 is selected as the best window length.

[0070] Determine the anchoring sample selection threshold \(T\) based on the previously calculated mean delay \(\mu\) and delay standard deviation \(\sigma\). In this embodiment, set \(T = \mu - 0.5\sigma=150 - 0.5\times30 = 135\mathrm{ms}\). Take the samples in the delay sample sequence that are less than the threshold \(T\) as the anchoring samples. For example, for a certain sample sequence \(S1 = \{160\mathrm{ms}, 140\mathrm{ms}, 130\mathrm{ms}, 145\mathrm{ms}, 155\mathrm{ms}, 130\mathrm{ms}, 120\mathrm{ms}, 150\mathrm{ms}, 140\mathrm{ms}, 125\mathrm{ms}, 135\mathrm{ms}, 145\mathrm{ms}\}\), since there are five values \(130\mathrm{ms}, 130\mathrm{ms}, 120\mathrm{ms}, 125\mathrm{ms}\) less than the threshold \(T = 135\mathrm{ms}\), the samples corresponding to these delay records are selected as the anchoring samples.

[0071] Apply the augmentation transformation parameters to the anchoring samples to generate a sequence of positive sample pairs. The augmentation transformations include: adding Gaussian noise (mean is 0, standard deviation is \(5\mathrm{ms}\)), time-axis scaling (the scaling factor is randomly selected between \(0.9\) and \(1.1\)), and random masking (randomly replace \(10\%\) of the values in the sequence with 0). Apply these transformations to each anchoring sample to generate the corresponding positive samples. For example, for the \(130\mathrm{ms}\) delay record in the anchoring sample \(S1\), it may become \(132\mathrm{ms}\) after applying Gaussian noise and may become \(136.5\mathrm{ms}\) after applying time-axis scaling (factor \(1.05\)). In this way, each anchoring sample and its transformed sample form a positive sample pair, forming a sequence of positive sample pairs \(P=\{P1, P2,\cdots, P m \}\).

[0072] Construct a deep feature extraction network. This network adopts a multi-layer temporal convolutional structure, including three convolutional layers, with 64, 128, and 256 convolutional kernels respectively in each layer, and the convolutional kernel size is \(3\times1\). A normalization layer and a ReLU activation function are connected after each convolutional layer. Finally, through global average pooling and a fully connected layer, map the features to a 128-dimensional feature space.

[0073] Use the deep feature extraction network to perform feature encoding on the sequence of positive sample pairs. Input the two samples in each positive sample pair \(P_i\) into the network respectively to obtain the corresponding feature vectors \(f_i\) and \(f_i'\), forming a feature vector pair \((f_i, f_i')\). For example, for the positive sample pair \(P1\) generated from the anchoring sample, after passing through the feature extraction network, the feature vectors \(f1 = [0.23, 0.45,\cdots, - 0.12]\) and \(f1'=[0.25, 0.42,\cdots, - 0.14]\) are obtained.

[0074] Calculate the cosine similarity between the feature vectors of the positive sample pairs. For the feature vector pair \((f_i, f_i')\), calculate their cosine similarity \(s_i\). For example, the cosine similarity \(s1\) between \(f1\) and \(f1'\) is \(0.95\), indicating that the directions of the two feature vectors are highly consistent.

[0075] Construct a contrast loss function based on cosine similarity. The goal of the contrast loss function is to make the similarity of the feature vectors of positive sample pairs as high as possible, while the similarity with the feature vectors of other samples as low as possible. It is constructed by calculating the difference between the average similarity of all positive sample pairs and the average similarity of random negative sample pairs. For example, if the average similarity of positive sample pairs is 0.92 and the average similarity of negative sample pairs is 0.15, the contrast loss value reflects this difference.

[0076] Perform a weighted combination of the contrast loss function and the delay prediction loss function to obtain a hybrid loss function. The delay prediction loss function uses the mean absolute error to evaluate the difference between the predicted delay value and the true delay value. Set the contrast loss weight α = 0.3 and the delay prediction loss weight β = 0.7. The hybrid loss L = α × contrast loss + β × delay prediction loss. For example, if the contrast loss is 0.25 and the delay prediction loss is 15ms in a certain training iteration, then the hybrid loss L = 0.3 × 0.25 + 0.7 × 15 = 10.575.

[0077] Calculate the gradient of the network parameters based on the hybrid loss function, and use the gradient to update the parameters of the deep feature extraction network. Use the Adam optimizer, set the learning rate to 0.001, and the batch size to 64. During the training process, each epoch makes a complete pass through the entire training set, and a total of 100 epochs are trained. Evaluate the performance of the validation set every 10 epochs. If the performance does not improve for 5 consecutive evaluations, the learning rate is halved. When the performance of the validation set does not improve for 10 consecutive evaluations, terminate the training.

[0078] The obtained aiming delay prediction model achieves a mean absolute error of 12ms on the test set, which is more than 50% improved compared to traditional methods (such as time series prediction methods like moving average and ARIMA, with errors usually in the range of 25 - 30ms). This model can adapt to the aiming delay changes in different usage scenarios and provide reliable delay prediction support for precise aiming control.

[0079] In practical applications, this model can be deployed in the control unit of the aiming system to receive aiming-related data in real time, predict the aiming delay at the next moment, and adjust the aiming parameters accordingly to improve the aiming accuracy and stability. Experiments show that after using this prediction model, the response speed of the aiming control system has increased by 30%, and the aiming accuracy has increased by 25%.

[0080] Figure 2 The flowchart of the delay prediction method based on deep feature extraction according to the embodiment of the present invention:

[0081] The flow chart describes a deep learning-based time delay prediction method, which mainly includes four key steps: First, process the historical aiming data. By calculating the time delay mean and the time delay standard deviation, and using a sliding time window for segmented sampling, a time delay sample sequence is generated by adjusting the sliding window length. Second, determine a threshold based on the calculated time delay mean and time delay standard deviation. Samples in the sample sequence that are less than this threshold are used as anchor samples, and corresponding enhancement transformation parameters are added to these anchor samples to generate a positive sample pair sequence. Third, construct a deep feature extraction network. Use this network to perform feature encoding on the positive sample pair sequence to obtain positive sample pair feature vectors, calculate the cosine similarity between the feature vectors, and construct a contrast loss function based on this similarity. Finally, perform a weighted combination of the contrast loss function and the time delay prediction loss function to obtain a hybrid loss function. The time delay prediction loss function is used to evaluate the error between the predicted time delay value and the actual time delay value. Based on the hybrid loss function, calculate the gradient of the network parameters, and use this gradient to update the parameters of the deep feature extraction network, and finally obtain the aiming time delay prediction model. The entire process forms a complete closed-loop system from data processing to model training.

[0082] In an alternative embodiment, inputting the real-time aiming parameters into the aiming time delay prediction model to obtain the predicted aiming time delay value includes:

[0083] Construct an input feature vector according to the real-time aiming parameters. The real-time aiming parameters include attitude parameters and angular velocity parameters. Normalize the input feature vector to obtain a normalized feature vector;

[0084] Input the normalized feature vector into a pre-trained aiming time delay prediction model. The aiming time delay prediction model is a deep neural network trained based on a contrast learning mechanism to obtain the predicted aiming time delay feature.

[0085] Collect the attitude parameters and angular velocity parameters during the aiming process in real time. The attitude parameters include the pitch, yaw, and roll angles of the firearm in three-dimensional space. These three angles respectively describe the inclination of the firearm relative to the horizontal plane, the vertical plane, and the axis. Taking a certain shooting as an example, the collected pitch angle is 15.3 degrees, the yaw angle is 23.7 degrees, and the roll angle is 2.1 degrees. The angular velocity parameters include the rotation speeds of the firearm around the three axes, which are respectively recorded as the pitch angle change rate of 3.2 degrees per second, the yaw angle change rate of 1.8 degrees per second, and the roll angle change rate of 0.5 degrees per second.

[0086] Collect additional relevant parameters, such as the current holding stability index (value range 0 - 100, example value 85), the target distance (unit: meter, example value 50 meters), the target moving speed (unit: meter / second, example value 2.5 meters / second), and environmental factors such as wind speed (unit: meter / second, example value 3.2 meters / second). These data together constitute a comprehensive description of the aiming process.

[0087] Construct an input feature vector based on these real-time collected aiming parameters. Specifically, combine the above parameters into a multi-dimensional vector in a predefined order. In this embodiment, the constructed feature vector is 10-dimensional and contains all the above collected parameter values. To enhance the robustness of the model, the rate of change of parameters within the front and rear time windows (such as 100 milliseconds) is also calculated as an additional dimension of the vector. For example, the calculated rate of change of the pitch angle within 100 milliseconds is 0.35 degrees / 100 milliseconds.

[0088] The constructed original feature vector needs to be normalized to eliminate the influence of different dimensions on the model. The normalization process uses the maximum-minimum normalization method to map each feature value to the interval [0, 1]. For example, for the pitch angle, assuming the pre-statistical maximum value is 45 degrees and the minimum value is -45 degrees, the normalization result of 15.3 degrees is (15.3 - (-45)) / (45 - (-45)) = 0.67. The system performs similar processing for each dimension to obtain a fully normalized feature vector.

[0089] To improve the stability of normalization, the system maintains a dynamically updated parameter statistical table to record the distribution characteristics of historical data. This table is updated every fixed time (such as after 1000 shots) to ensure that the normalization reference value can adapt to different shooting environments and the characteristics of different shooters.

[0090] After normalization, feature selection and dimensionality reduction are performed to remove redundant or low-information features. Specifically, the principal component analysis method is used to extract the most representative 6 principal components from the 10-dimensional features, retaining approximately 95% of the original information. For example, the first principal component after dimensionality reduction mainly reflects the comprehensive performance of firearm stability, with a value of 0.78.

[0091] Input the standardized feature vector into a pre-trained aiming time delay prediction model. This model is a deep neural network trained based on the contrast learning mechanism, and its network architecture includes 5 hidden layers, with 64, 128, 256, 128, and 64 neurons in each layer respectively. The activation function uses the ReLU function to enhance the non-linear expression ability of the network.

[0092] The training of the model adopts a contrastive learning mechanism, that is, it learns by comparing the similarity of time-delay features between different shooting samples. The training dataset contains 500,000 shooting data of 10,000 shooters at different levels, covering various environmental conditions and target types. During the training process, the model uses the InfoNCE loss function with a temperature parameter of 0.07, the learning rate is initially set to 0.001, and a cosine annealing strategy is adopted for dynamic adjustment.

[0093] During the prediction process, the model outputs the aiming time-delay feature vector, which is a 128-dimensional high-dimensional representation that captures the potential patterns of the aiming time-delay. For example, the value of the first element of the feature vector obtained from a certain prediction is 0.92, indicating a higher time-delay possibility. Subsequently, the system converts this 128-dimensional vector into an actual time-delay prediction value through a decoder, such as a predicted time-delay of 345 milliseconds.

[0094] To verify the prediction accuracy, the system compares the predicted value with the actually measured time-delay value. On the test set, this method achieves a prediction accuracy with a mean absolute error (MAE) of 28 milliseconds and a relative error of less than 10%, meeting the requirements of practical applications.

[0095] Through the above steps, this embodiment successfully converts the real-time aiming parameters into accurate aiming time-delay prediction values, providing key support for subsequent ballistic correction and precise shooting. This method also performs excellently in terms of processing speed, with a single prediction taking less than 5 milliseconds, meeting the requirements of real-time application scenarios.

[0096] In an alternative embodiment, based on the predicted aiming time-delay value, a recursive least squares estimation algorithm is used to perform real-time optimization and compensation on the attitude tracking control parameters of the satellite laser communication system, where the measurement noise covariance matrix of the recursive least squares estimation algorithm is dynamically adjusted according to the predicted aiming time-delay value, including:

[0097] Construct a time-delay adaptive factor according to the predicted aiming time-delay value, and perform weighted fusion on the time-delay adaptive factor and the initial value of the measurement noise covariance matrix to obtain a dynamically adjusted measurement noise covariance matrix;

[0098] Calculate the asymptotic optimized value of the prediction error covariance matrix using the recursive least squares estimation algorithm, and combine the asymptotic optimized value with the sampling weights of the Gauss-Hermite transform to obtain an adjusted prediction error covariance matrix;

[0099] Calculate a multi-scale strong tracking factor based on the adjusted prediction error covariance matrix, combine the multi-scale strong tracking factor with the dynamically adjusted measurement noise covariance matrix to obtain a gain correction matrix, and use the gain correction matrix to perform online correction on the Kalman gain;

[0100] Estimate the system state based on the result of the corrected Kalman gain, and perform information fusion on the system state and the result of the corrected Kalman filter to obtain an optimal state estimate value with enhanced robustness;

[0101] Construct a performance index according to the optimal state estimate value, and design a control compensation strategy with anti-interference ability based on the performance index to realize real-time optimization compensation of attitude tracking control parameters.

[0102] Construct a delay adaptive factor according to the predicted aiming delay value. In actual operation, when the measured system aiming delay value is 25 ms, the delay adaptive factor can be constructed as follows: normalize the delay value, take the delay reference value as 20 ms, then the normalization coefficient is 1.25, that is, the delay adaptive factor is set to 1.25. This adaptive factor dynamically adjusts with the change of the delay value. When the delay value increases to 30 ms, the delay adaptive factor correspondingly increases to 1.5, reflecting the sensitive response of the system to the increase in delay.

[0103] Perform weighted fusion on the delay adaptive factor and the initial value of the measurement noise covariance matrix. Assume that the diagonal elements of the initial measurement noise covariance matrix are [0.02, 0.02, 0.01], representing the measurement noise levels of different attitude angles. Then the diagonal elements of the dynamically adjusted matrix after weighted fusion become [0.025, 0.025, 0.0125]. This weighting method magnifies the measurement noise covariance correspondingly with the increase of the delay, improving the adaptability of the system to the delay change.

[0104] In the prediction error covariance matrix optimization link, use the recursive least squares algorithm to calculate the asymptotic optimization value. Specifically, perform recursive learning on 100 frames of historical data, set the initial estimation error to 0.05, and the forgetting factor to 0.98. After recursive calculation, the obtained asymptotic optimization value is 0.032, reflecting the result of the system's learning of historical data. Combine this optimization value with the sampling weights of the Gauss-Hermite transform, and select the typical values [0.4, 0.3, 0.3] for the sampling weights to obtain the diagonal elements of the adjusted prediction error covariance matrix as [0.0128, 0.0096, 0.0096].

[0105] Based on the adjusted prediction error covariance matrix, calculate the multi-scale strong tracking factors. In practical applications, calculate the strong tracking factors in the pitch, roll, and yaw directions respectively according to the change rate of the system attitude angle. When the attitude change rate is large (for example, the pitch angle change rate reaches 2° / s), the corresponding strong tracking factor is set to 1.8; when the change rate is at a medium level (for example, the roll angle change rate is 1° / s), the strong tracking factor is set to 1.4; when the change rate is small (for example, the yaw angle change rate is 0.5° / s), the strong tracking factor is set to 1.2. Combine these multi-scale strong tracking factors with the dynamically adjusted measurement noise covariance matrix to form a gain correction matrix, and its diagonal elements are [0.045, 0.035, 0.015] respectively.

[0106] Use the gain correction matrix to perform online correction on the Kalman gain. The original Kalman gain value is [0.65, 0.58, 0.72], and the corrected gain value is adjusted to [0.702, 0.623, 0.756]. This correction enhances the dependence of the filter on the observed data and enables the system to respond more sensitively to environmental changes.

[0107] Estimate the system state based on the corrected Kalman gain. Assume that the current attitude angle observation values are [2.5°, 1.8°, 0.9°], and the predicted values are [2.3°, 1.6°, 0.8°]. Then, the estimated state value calculated according to the corrected Kalman gain is [2.44°, 1.73°, 0.88°]. Perform information fusion on this estimated state and the corrected Kalman filter result. The fusion weights are assigned as [0.6, 0.7, 0.5] according to the certainty of each state estimate, and finally the optimal state estimate value [2.42°, 1.71°, 0.86°] is obtained.

[0108] Construct performance indicators based on the optimal state estimate value. The indicators include two aspects: tracking accuracy and system robustness. The tracking accuracy is expressed by the root mean square error, and the target value is set to be less than 0.05°; the system robustness is expressed by the maximum transient response amplitude, and the target value is set to be less than 0.1°. In the test phase, the root mean square tracking error of the system at a 25 ms time delay is 0.042°, and the maximum transient response is 0.087°, both of which meet the set targets.

[0109] Based on the above performance indicators, design an anti-interference control compensation strategy. During the actual compensation process, when an external interference is detected (such as the micro-vibration amplitude exceeds 0.01°), the system adjusts the control parameters according to the interference characteristics. For high-frequency interference (frequency > 10 Hz), increase the damping coefficient to 0.85; for low-frequency interference (frequency < 1 Hz), increase the control gain to 1.2 times the initial value. Tests show that this compensation strategy enables the attitude tracking error of the system to be controlled within 0.048° when bearing an external interference with an amplitude of 0.03°, and the response time does not exceed 150 ms.

[0110] The implementation effect of the above method shows that in the satellite laser communication environment with dynamically changing time delay, this real-time optimization and compensation method can effectively improve the system attitude tracking control accuracy, enhance the system robustness, and meet the stable pointing requirements of high-precision laser communication.

[0111] Figure 3 The following is the comparison chart of attitude tracking accuracy based on time delay prediction in the embodiment of the present invention:

[0112] This figure shows the comparison results of the tracking performance of three different technical solutions within a time span of 50 seconds. The horizontal axis in the figure represents time (seconds), and the vertical axis represents tracking accuracy. The solid line (triangle mark) represents the performance curve of the present technical solution, the dashed line (square mark) represents the performance curve of the traditional recursive least squares method, and the dotted line (circle mark) represents the performance curve of the standard Kalman filter. It can be seen from the performance curves that the present technical solution maintains a relatively high and stable tracking accuracy throughout the test process, and the numerical value always remains between 0.95 and 0.98, with a small fluctuation range; while the performance of the traditional recursive least squares method gradually decreases over time, from the initial 0.92 to about 0.73 at 50 seconds; the performance attenuation trend of the standard Kalman filter is similar to that of the traditional recursive least squares method, but the overall performance is slightly lower, from the initial 0.87 to about 0.72 at 50 seconds. This comparison result clearly demonstrates the significant advantages of the present technical solution in terms of tracking accuracy and stability, especially showing stronger anti-interference ability and stability during long-term tracking.

[0113] In an alternative implementation, the recursive least squares estimation algorithm is used to calculate the asymptotic optimized value of the prediction error covariance matrix, and the asymptotic optimized value is combined with the sampling weights of the Gauss-Hermite transform to obtain the adjusted prediction error covariance matrix, including:

[0114] Construct an initial prediction error covariance matrix including the historical prediction error vector and the process noise covariance matrix, including: multiplying the historical prediction error vector by the exponential power term of the forgetting factor and summing to obtain an error component, and adding the error component to the process noise covariance matrix to obtain the initial prediction error covariance matrix;

[0115] Based on the initial prediction error covariance matrix, construct a recursive least squares optimization objective function, the recursive least squares optimization objective function includes the weighted norm term of the historical prediction error vector, calculate the gradient of the recursive least squares optimization objective function, and add the product of the gradient and the optimization gain to the initial prediction error covariance matrix to obtain the optimized prediction error covariance matrix;

[0116] Obtain the system state mean, add the product of the system state mean, the scaled square root term of the optimized prediction error covariance matrix, and the sampling vector to generate a set of sampling points, and calculate the mean weight and covariance weight corresponding to the set of sampling points based on the system state dimension and the scaling parameter;

[0117] Multiply and sum the quadratic form of each sampling point in the set of sampling points with the mean weight and the covariance weight to obtain a covariance compensation term, and calculate the negative exponential function of the ratio of the covariance compensation term to the Frobenius norm of the optimized prediction error covariance matrix to obtain an adaptive fusion weight coefficient;

[0118] Perform weighted combination of the covariance compensation term and the optimized prediction error covariance matrix according to the adaptive fusion weight coefficient to obtain an adjusted prediction error covariance matrix.

[0119] Construct an initial prediction error covariance matrix that includes the historical prediction error vector and the process noise covariance matrix. The prediction error vector at time k of the system is denoted as e_k, and the process noise covariance matrix is denoted as Q_k. Introduce a forgetting factor λ (usually with a value range of 0.95 to 0.99) to balance the weights of historical data and new data.

[0120] For the N collected historical prediction error vectors, calculate the error component S_k. The specific method is to multiply each historical prediction error vector e_i by the exponential power term (k - i) of the forgetting factor λ and sum them. For example, when λ = 0.98 and 10 historical prediction error vectors are collected, each vector is a 3D vector [0.1, 0.2, 0.3], the calculated error component S_10 is approximately [2.37, 4.74, 7.11]. Then, add the error component S_k to the process noise covariance matrix Q_k to obtain the initial prediction error covariance matrix P_k^-.

[0121] Construct a recursive least squares optimization objective function based on the initial prediction error covariance matrix. This objective function includes a weighted norm term of the historical prediction error vector. Calculate the gradient G_k of this recursive least squares optimization objective function with respect to P_k^-.

[0122] In practical applications, this gradient can be calculated through the outer product of the historical prediction error vector and the inverse matrix of the current prediction error covariance matrix. Introduce an optimization gain α (usually choose a value between 0.01 and 0.1), multiply the gradient G_k by the optimization gain α and then superimpose it on the initial prediction error covariance matrix P_k^- to obtain the optimized prediction error covariance matrix P_k^opt. For example, when α = 0.05, the initial covariance matrix is a 3×3 identity matrix, and each element of the gradient matrix is 0.2, each diagonal element of the optimized covariance matrix is approximately 1.01.

[0123] Obtain the system state mean \(x_k\), which is usually provided by the prediction step of the Kalman filter. Calculate the scaled square root term of the optimized prediction error covariance matrix \(P_k^{opt}\), which can be achieved by Cholesky decomposition. Set the sampling vector \(\xi\), usually a unit vector is chosen. Add the product of the system state mean \(x_k\), the scaled square root term of the optimized prediction error covariance matrix \(P_k^{opt}\), and the sampling vector \(\xi\) to generate the set of sampling points \(X_i\).

[0124] Based on the system state dimension \(n\) and the scaling parameter \(\kappa\) (usually taken as \(3 - n\)), calculate the mean weight \(W_m\) and covariance weight \(W_c\) corresponding to the set of sampling points. For a 3 - dimensional state space, when \(\kappa = 0\), the mean weight of the central point is 0, and the mean weights of the remaining \(2n\) sampling points are all \(1 / (2n)=1 / 6\).

[0125] Calculate the covariance compensation term \(P_c\), which is the product of the outer product of the vectors formed by the differences between each sampling point in the set of sampling points \(X_i\) and the system state mean \(x_k\) and the mean weight \(W_m\), then multiplied by the covariance weight \(W_c\) and summed over all sampling points. For example, in a 3 - dimensional state space, when the difference between the sampling point and the mean is \([0.1, 0.2, 0.3]\), the mean weight is \(1 / 6\), and the covariance weight is \(1 / 6\), the contribution of this sampling point to the covariance compensation term is approximately \(0.0006\).

[0126] Calculate the ratio of the Frobenius norm of the covariance compensation term \(P_c\) to the Frobenius norm of the optimized prediction error covariance matrix \(P_k^{opt}\), and take its negative exponential function to obtain the adaptive fusion weight coefficient \(\beta\). The Frobenius norm can be understood as the square root of the sum of the squares of all elements of the matrix. For example, when the Frobenius norm of \(P_c\) is \(0.5\) and the Frobenius norm of \(P_k^{opt}\) is \(2.0\), the ratio is \(0.25\), and taking the negative exponential gives \(\beta\) approximately equal to \(0.78\).

[0127] Weight - combine the covariance compensation term \(P_c\) and the optimized prediction error covariance matrix \(P_k^{opt}\) according to the adaptive fusion weight coefficient \(\beta\), that is \(P_k^{adj}=\beta P_k^{opt}+(1 - \beta)P_c\), to obtain the adjusted prediction error covariance matrix \(P_k^{adj}\). When \(\beta = 0.78\), the weight of the optimized covariance matrix in the final adjusted prediction error covariance matrix is \(78\%\), and the weight of the covariance compensation term is \(22\%\).

[0128] Through the above steps, the method of the present invention realizes the dynamic optimization and adjustment of the prediction error covariance matrix, effectively integrates the historical data learning ability of the recursive least squares estimation and the sampling characteristics of the Gaussian-Hermite transform, and significantly improves the accuracy of the covariance matrix estimation and the robustness of the system state estimation. In the actual application test, compared with the traditional method, the root mean square error of the state estimation of this method is reduced by 15% to 20% on average, and it is especially suitable for complex systems with strong nonlinearity and unknown noise characteristics.

[0129] In an optional implementation manner, based on the optimal state estimation value and the attitude tracking control parameters after real-time optimization and compensation, the robust adaptive Kalman filtering algorithm is used to calculate the aiming time delay tolerance in real time, and the transmitted beam power and beam divergence angle of the satellite laser communication system are dynamically adjusted according to the aiming time delay tolerance, including:

[0130] Establish a dynamic model of the time delay tolerance, set the state equation of the dynamic model of the time delay tolerance as the sum of the nonlinear function of the time delay tolerance at the previous moment, the optimal state estimation value, and the attitude tracking control parameters after real-time optimization and compensation and the process noise, and set the observation equation of the dynamic model of the time delay tolerance as the sum of the nonlinear function of the time delay tolerance at the current moment and the measurement noise;

[0131] Calculate the prediction error covariance matrix of the time delay tolerance at the previous moment and the time delay tolerance at the current moment based on the state equation and the observation equation;

[0132] Construct an innovation covariance matrix, calculate an adaptive factor according to the ratio of the product of the innovation covariance matrix and the prediction error covariance matrix, and use the result of the combined operation of the adaptive factor and the prediction error covariance matrix as the gain matrix;

[0133] Based on the gain matrix and the deviation between the observed value and the predicted value, calculate the optimal predicted estimate value of the time delay tolerance, substitute the deviation between the optimal predicted estimate value and the target time delay tolerance into the exponential function to calculate the adjustment amount of the transmitted beam power, and substitute the ratio of the optimal predicted estimate value and the prediction error covariance matrix into the hyperbolic tangent function to calculate the adjustment amount of the beam divergence angle;

[0134] Use the product of the adjustment amount of the transmitted beam power and the nominal power as the optimized transmitted beam power, and use the sum of the adjustment amount of the beam divergence angle and the reference divergence angle as the optimized beam divergence angle.

[0135] A dynamic model of time-delay tolerance is established. The state equation of this model is composed of the sum of a non-linear function of the time-delay tolerance at the previous moment (denoted as Td_k-1), the optimal state estimate value (denoted as X_opt), and the attitude tracking control parameters after real-time optimization compensation (denoted as P_comp), and the process noise (denoted as w_k-1). When specifically implemented, the following method can be adopted: The predicted value of the time-delay tolerance at the current moment is equal to the product of the time-delay tolerance at the previous moment and 0.85, plus the product of the attitude angle error component in the optimal state estimate value and 0.12, plus the product of the gain coefficient in the compensated tracking control parameters and 0.08, and finally plus the process noise that follows a Gaussian distribution, with a mean of 0 and a standard deviation of 0.02 ms.

[0136] The observation equation is set as the sum of a non-linear function of the time-delay tolerance at the current moment and the measurement noise. In actual implementation, the observed value is equal to the product of the actual time-delay tolerance at the current moment and 0.95, plus the measurement noise that follows a Gaussian distribution, with a mean of 0 and a standard deviation of 0.03 ms.

[0137] Calculate the prediction error covariance matrix of the time-delay tolerance at the previous moment and the current moment. When specifically implemented, first calculate the state prediction value, that is, the predicted value of the time-delay tolerance at the current moment obtained through the state equation (denoted as Td_k-). For example, if the time-delay tolerance at the previous moment is 2.5 ms, the attitude angle error component is 0.08 rad, and the tracking control gain coefficient is 1.2, then the predicted value of the time-delay tolerance at the current moment is approximately:

[0138] 2.5×0.85 + 0.08×0.12 + 1.2×0.08 + w_k-1 ≈ 2.244 ms (ignoring the noise term).

[0139] Calculate the prediction error covariance matrix (denoted as P_k-), which reflects the uncertainty of the state prediction. In the case of a one-dimensional state variable, this matrix degenerates into a scalar, which can be expressed as the product of the error covariance at the previous moment and the square of the state transition matrix, plus the process noise covariance. For example, if the error covariance at the previous moment is 0.04, the state transition matrix is 0.85, and the process noise covariance is 0.0004, then the prediction error covariance is:

[0140] 0.04×0.85² + 0.0004 ≈ 0.0293.

[0141] Construct the innovation covariance matrix. Innovation refers to the difference between the observed value and the predicted observed value, and the innovation covariance matrix reflects the statistical characteristics of this difference. For example, if the predicted observed value is 2.15 ms and the actual observed value is 2.20 ms, then the innovation is 0.05 ms. Suppose the innovations at five consecutive time instants are 0.05, -0.03, 0.02, 0.04, -0.01 ms respectively, then the innovation covariance can be estimated as the average of the squares of these values, approximately 0.0011.

[0142] Calculate the adaptive factor according to the ratio of the product of the innovation covariance matrix and the predicted error covariance matrix. The adaptive factor is used to adjust the trust level of the filter for new observed data. For example, if the innovation covariance is 0.0011, the predicted error covariance is 0.0293, and the observation noise covariance is 0.0009, then the adaptive factor can be calculated as:

[0143] 0.0011 / (0.0293 + 0.0009) ≈ 0.036.

[0144] Take the result of the combined operation of the adaptive factor and the predicted error covariance matrix as the gain matrix. For example, the gain matrix can be calculated as:

[0145] 0.0293 × 0.036 / (0.036 × 0.0293 + 0.0009) ≈ 0.54.

[0146] Based on the gain matrix, the deviation between the observed value and the predicted value, calculate the optimal predicted estimate of the delay tolerance. Specifically, the optimal predicted estimate is equal to the predicted value plus the product of the gain matrix and (the observed value minus the predicted observed value). For example, if the predicted value is 2.244 ms, the observed value is 2.20 ms, and the gain matrix is 0.54, then the optimal predicted estimate is:

[0147] 2.244 + 0.54 × (2.20 - 2.244) ≈ 2.220 ms.

[0148] Substitute the deviation between the optimal predicted estimate and the target delay tolerance into the exponential function to calculate the adjustment amount of the transmitted beam power. For example, if the optimal predicted estimate is 2.220 ms and the target delay tolerance is 2.5 ms, then the deviation is -0.28 ms, and the power adjustment factor can be calculated as exp(0.35 × 0.28) ≈ 1.103, indicating that the power needs to be increased by approximately 10.3%.

[0149] Substitute the ratio of the optimal predicted estimate and the predicted error covariance matrix into the hyperbolic tangent function to calculate the adjustment amount of the beam divergence angle. For example, if the optimal predicted estimate is 2.220 ms and the predicted error covariance is 0.0293, then the ratio is approximately 75.8, and the beam divergence angle adjustment amount can be calculated as:

[0150] 0.2×tanh(75.8 / 100)≈0.15 mrad.

[0151] Take the product of the adjustment amount of the transmitted beam power and the nominal power as the optimized transmitted beam power, and take the sum of the adjustment amount of the beam divergence angle and the reference divergence angle as the optimized beam divergence angle. For example, if the nominal power is 5 W, the optimized transmitted power is 5×1.103≈5.515 W; if the reference divergence angle is 2 mrad, the optimized beam divergence angle is 2 + 0.15 = 2.15 mrad.

[0152] Through the above method, the satellite laser communication system can dynamically adjust the transmitted beam power and the beam divergence angle according to the real-time calculated aiming time delay tolerance, so as to effectively cope with attitude disturbances and external environmental changes, and ensure communication quality and stability.

[0153] Figure 4 The following is the flowchart of the beam parameter regulation based on the time delay tolerance self-adaptive optimization in the embodiment of the present invention:

[0154] This flowchart details the complete implementation process of a beam parameter self-adaptive optimization method. First, the system inputs the optimal state estimation and the attitude tracking control parameters after real-time optimization compensation as the initial conditions. In step 1, a time delay tolerance dynamic model is established, including a state equation composed of the non-linear function of the previous moment's time delay tolerance and noise, and an observation equation composed of the non-linear function of the current moment's time delay tolerance and noise. In step 2, based on the established state equation and observation equation, the prediction error covariance matrix between adjacent moments is calculated. In step 3, a gain matrix is constructed. By establishing an innovation covariance matrix and calculating an adaptive factor, the two are combined to obtain the final gain matrix. In step 4, the adjustment amount is calculated, including calculating the optimal prediction estimate value of the time delay tolerance, and calculating the adjustment amounts of the transmitted beam power and the beam divergence angle accordingly. Finally, in step 5, by performing combined operations on the adjustment amounts and the nominal power and the reference divergence angle respectively, the optimized transmitted beam power and beam divergence angle parameters are obtained. The whole process forms a complete closed-loop system from parameter input to final optimization output, reflecting the systematicness and realizability of this method in beam parameter self-adaptive optimization.

[0155] In the second aspect of the embodiment of the present invention, a ground test system for the satellite laser communication aiming time delay tolerance characteristic is provided, including:

[0156] The first unit is used to obtain the historical aiming data of the satellite laser communication system, construct a deep learning training sample set according to the historical aiming data, and train a preset deep neural network using the deep learning training sample set to obtain an aiming time delay prediction model;

[0157] A second unit, configured to collect real-time aiming parameters of a current satellite laser communication system, input the real-time aiming parameters into the aiming time-delay prediction model, and obtain a predicted aiming time-delay value;

[0158] A third unit, configured to construct a time-delay adaptive factor based on the predicted aiming time-delay value, obtain a dynamically adjusted measurement noise covariance matrix based on the time-delay adaptive factor, and combine the dynamically adjusted measurement noise covariance matrix with a multi-scale strong tracking factor; perform real-time optimization and compensation on the attitude tracking control parameters of the satellite laser communication system by using a recursive least squares estimation algorithm, where the measurement noise covariance matrix of the recursive least squares estimation algorithm is dynamically adjusted according to the predicted aiming time-delay value, and estimate the Kalman gain to obtain an optimally estimated value with enhanced robustness;

[0159] A fourth unit, configured to calculate an aiming time-delay tolerance in real time by using a robust adaptive Kalman filtering algorithm based on the optimally estimated value and the attitude tracking control parameters after real-time optimization and compensation, and dynamically adjust the transmission beam power and beam divergence angle of the satellite laser communication system according to the aiming time-delay tolerance.

[0160] In a third aspect of the embodiments of the present invention, there is provided an electronic device, including:

[0161] A processor;

[0162] A memory for storing instructions executable by the processor;

[0163] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0164] In a fourth aspect of the embodiments of the present invention, there is provided a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0165] The present invention may be a method, an apparatus, a system, and / or a computer program product. The computer program product may include a computer-readable storage medium, on which computer-readable program instructions for executing various aspects of the present invention are uploaded.

[0166] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. Ground test method for time-delay tolerance characteristics in satellite laser communication aiming, characterized in that, Including: Obtain the historical aiming data of the satellite laser communication system, construct a deep learning training sample set according to the historical aiming data, and train a preset deep neural network using the deep learning training sample set to obtain a aiming time delay prediction model; Collect the real-time aiming parameters of the current satellite laser communication system, input the real-time aiming parameters into the aiming time delay prediction model to obtain a predicted aiming time delay value; Construct a time delay adaptive factor based on the predicted aiming time delay value, obtain a dynamically adjusted measurement noise covariance matrix based on the time delay adaptive factor, and combine the dynamically adjusted measurement noise covariance matrix with a multi-scale strong tracking factor; Use the recursive least squares estimation algorithm to perform real-time optimization compensation on the attitude tracking control parameters of the satellite laser communication system, where the measurement noise covariance matrix of the recursive least squares estimation algorithm is dynamically adjusted according to the predicted aiming time delay value, and estimate the Kalman gain to obtain an optimal state estimation value with enhanced robustness; Based on the optimal state estimation value and the attitude tracking control parameters after real-time optimization compensation, use the robust adaptive Kalman filter algorithm to calculate the aiming time delay tolerance in real time, and dynamically adjust the transmitted beam power and beam divergence angle of the satellite laser communication system according to the aiming time delay tolerance.

2. The method according to claim 1, wherein Constructing a deep learning training sample set according to the historical aiming data and training a preset deep neural network using the deep learning training sample set to obtain a aiming time delay prediction model includes: Calculate the time delay mean and time delay standard deviation according to the historical aiming data, segment and sample the historical aiming data using a sliding time window, and generate a time delay sample sequence by adjusting the length of the sliding time window; Determine an anchor sample selection threshold based on the time delay mean and the time delay standard deviation, use the samples in the time delay sample sequence that are less than the anchor sample selection threshold as anchor samples, and apply corresponding enhancement transformation parameters to the anchor samples to generate a positive sample pair sequence; Construct a deep feature extraction network, use the deep feature extraction network to perform feature encoding on the positive sample pair sequence to obtain a positive sample pair feature vector; calculate the cosine similarity between the positive sample pair feature vectors, and construct a contrast loss function according to the cosine similarity; Perform weighted combination on the contrast loss function and a time delay prediction loss function to obtain a mixed loss function, where the time delay prediction loss function is used to evaluate the error between the predicted time delay value and the true time delay value; calculate the gradient of the network parameters based on the mixed loss function, and use the gradient to update the parameters of the deep feature extraction network to obtain a aiming time delay prediction model.

3. The method according to claim 1, characterized in that Inputting the real-time aiming parameters into the aiming time delay prediction model to obtain a predicted aiming time delay value includes: Construct an input feature vector according to the real-time aiming parameters, where the real-time aiming parameters include attitude parameters and angular velocity parameters, and perform normalization processing on the input feature vector to obtain a normalized feature vector; Input the standardized feature vector into a pre-trained aiming time-delay prediction model, which is a deep neural network trained based on a contrastive learning mechanism, to obtain the aiming time-delay prediction model.

4. The method according to claim 1, characterized in that, Based on the predicted aiming time-delay value, use the recursive least squares estimation algorithm to perform real-time optimization and compensation on the attitude tracking control parameters of the satellite laser communication system, where the measurement noise covariance matrix of the recursive least squares estimation algorithm is dynamically adjusted according to the predicted aiming time-delay value, including: Construct a time-delay adaptive factor according to the predicted aiming time-delay value, and perform weighted fusion on the time-delay adaptive factor and the initial value of the measurement noise covariance matrix to obtain a dynamically adjusted measurement noise covariance matrix; Use the recursive least squares estimation algorithm to calculate the asymptotic optimized value of the prediction error covariance matrix, and combine the asymptotic optimized value with the sampling weights of the Gauss-Hermite transform to obtain an adjusted prediction error covariance matrix; Based on the adjusted prediction error covariance matrix, perform subspace decomposition on the state residual sequence and the innovation sequence, calculate the time-varying scale parameter according to the eigenvalues corresponding to each subspace, multiply the time-varying scale parameter by the norm of the state prediction error to obtain the attenuation coefficient for each dimension, construct an attenuation matrix based on the attenuation coefficient, multiply the attenuation matrix by the adjusted prediction error covariance matrix to obtain a scaling factor, multiply the scaling factor by the innovation sequence to obtain a multi-scale strong tracking factor, combine the multi-scale strong tracking factor with the dynamically adjusted measurement noise covariance matrix to obtain a gain correction matrix, and use the gain correction matrix to perform online correction on the Kalman gain; Perform information fusion on the system state and the result of the corrected Kalman gain to obtain an optimal state estimate value with enhanced robustness; construct a performance index based on the optimal state estimate value, and design an anti-interference control compensation strategy based on the performance index to achieve real-time optimization and compensation of the attitude tracking control parameters.

5. The method according to claim 4, characterized in that Using the recursive least squares estimation algorithm to calculate the asymptotic optimized value of the prediction error covariance matrix, and combining the asymptotic optimized value with the sampling weights of the Gauss-Hermite transform to obtain an adjusted prediction error covariance matrix includes: Construct an initial prediction error covariance matrix including the historical prediction error vector and the process noise covariance matrix, including: multiplying the historical prediction error vector by the exponential power term of the forgetting factor and summing to obtain an error component, and adding the error component to the process noise covariance matrix to obtain the initial prediction error covariance matrix; Construct a state observation mapping matrix based on the recursive least squares estimation algorithm, use the state observation mapping matrix to establish a recursive gain sequence, and iteratively calculate the convergence value of the estimation error covariance matrix according to the recursive gain sequence, and use the convergence value as the asymptotic optimized value of the initial prediction error covariance matrix; Construct a recursive least squares optimization objective function based on the initial prediction error covariance matrix, calculate the gradient of the recursive least squares optimization objective function, multiply the gradient by an optimization gain, and then add the result to the initial prediction error covariance matrix to obtain an optimized prediction error covariance matrix; Obtain the system state mean, add the product of the system state mean, the scaled square root term of the optimized prediction error covariance matrix, and the sampling vector to generate a set of sampling points, and calculate the mean weights and covariance weights corresponding to the set of sampling points based on the system state dimension and the scaling parameter; Multiply and sum the mean weights and the covariance weights to obtain a covariance compensation term, and perform a weighted combination of the covariance compensation term and the optimized prediction error covariance matrix according to an adaptive fusion weight coefficient to obtain an adjusted prediction error covariance matrix.

6. The method according to claim 1, wherein Based on the optimal state estimate value and the attitude tracking control parameters after real-time optimization compensation, use the robust adaptive Kalman filter algorithm to calculate the aiming delay tolerance in real time, and dynamically adjust the transmit beam power and beam divergence angle of the satellite laser communication system, including: Establish a dynamic model of the delay tolerance, set the state equation of the dynamic model of the delay tolerance as the sum of a nonlinear function of the previous moment's delay tolerance, the optimal state estimate value, and the attitude tracking control parameters after real-time optimization compensation and process noise, and set the observation equation of the dynamic model of the delay tolerance as the sum of a nonlinear function of the current moment's delay tolerance and measurement noise; Calculate the prediction error covariance matrices of the previous moment's delay tolerance and the current moment's delay tolerance based on the state equation and the observation equation; Construct an innovation covariance matrix, calculate an adaptive factor according to the ratio of the product of the innovation covariance matrix and the prediction error covariance matrix, and use the result of the combined operation of the adaptive factor and the prediction error covariance matrix as a gain matrix; Based on the gain matrix and the deviation between the observed value and the predicted value, calculate the optimal predicted estimate value of the delay tolerance, substitute the deviation between the optimal predicted estimate value and the target delay tolerance into an exponential function to calculate the adjustment amount of the transmit beam power, and substitute the ratio of the optimal predicted estimate value and the prediction error covariance matrix into a hyperbolic tangent function to calculate the adjustment amount of the beam divergence angle; Use the product of the adjustment amount of the transmit beam power and the nominal power as the optimized transmit beam power, and use the sum of the adjustment amount of the beam divergence angle and the reference divergence angle as the optimized beam divergence angle.

7. A ground test system for the timing tolerance characteristics during satellite laser communication aiming, which is used to implement the method described in any one of the foregoing claims 1-6, characterized in that, Including: A first unit for obtaining the historical aiming data of the satellite laser communication system, constructing a deep learning training sample set according to the historical aiming data, and training a preset deep neural network using the deep learning training sample set to obtain an aiming delay prediction model; A second unit for collecting the real-time aiming parameters of the current satellite laser communication system, and inputting the real-time aiming parameters into the aiming delay prediction model to obtain a predicted aiming delay value; A third unit, configured to construct a time-delay adaptive factor based on the predicted aiming time-delay value, obtain a dynamically adjusted measurement noise covariance matrix based on the time-delay adaptive factor, and combine the dynamically adjusted measurement noise covariance matrix with a multi-scale strong tracking factor; Use the recursive least squares estimation algorithm to perform real-time optimization compensation on the attitude tracking control parameters of the satellite laser communication system, where the measurement noise covariance matrix of the recursive least squares estimation algorithm is dynamically adjusted according to the predicted aiming time-delay value, and estimate the Kalman gain to obtain an optimal state estimation value with enhanced robustness; A fourth unit, configured to calculate the aiming time-delay tolerance in real time using a robust adaptive Kalman filter algorithm based on the optimal state estimation value and the attitude tracking control parameters after real-time optimization compensation, and dynamically adjust the transmission beam power and beam divergence angle of the satellite laser communication system according to the aiming time-delay tolerance.

8. An electronic device, characterized in that, Comprising: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, the method according to any one of claims 1 to 6 is implemented.

Images