Inter-satellite laser communication and signal scheduling method based on adaptive polarization modulation

Abstract

The invention provides a self-adaptive polarization modulation-based inter-satellite laser communication and signal scheduling method, which relates to the field of laser communication, and comprises the following steps of: acquiring polarization state information of emitted and received light beams, calculating a mismatch degree, generating a compensation parameter, constructing a polarization state prediction model by using a deep learning neural network for pre-correction, and calculating the mismatch degree; a polarization modulator is adopted to perform real-time modulation and establish a link quality evaluation model, carrier synchronization and channel compensation are realized based on an adaptive optical carrier recovery technology of coherent detection, a link state is monitored in real time, and a training data set is updated.

Description

Technical Field

[0001] The present invention relates to the technical field of laser communication, and in particular, to an inter-satellite laser communication and signal scheduling method based on adaptive polarization modulation. Background Art

[0002] Inter-satellite laser communication is a key technology for realizing a space information network, which has advantages such as high bandwidth, high security, and low power consumption. In an inter-satellite laser communication system, the polarization state matching between the transmitting end and the receiving end has an important impact on the communication quality. The space environment is complex and changeable, and factors such as satellite attitude changes and atmospheric disturbances will cause the polarization state of the light beam to drift, affecting the stability and reliability of the communication link;

[0003] At present, the polarization state control in inter-satellite laser communication mainly adopts fixed compensation or simple feedback regulation methods. Traditional polarization state control methods usually rely on real-time measurement and feedback, and cannot effectively cope with the rapidly changing space environment. At the same time, existing polarization state compensation technologies lack the ability to predict the future polarization state change trend, and it is difficult to achieve advance compensation and optimized regulation;

[0004] Therefore, there is an urgent need for a solution to solve the problems existing in the prior art. Summary of the Invention

[0005] An embodiment of the present invention provides an inter-satellite laser communication and signal scheduling method based on adaptive polarization modulation, which can at least solve some of the problems existing in the prior art.

[0006] In a first aspect of an embodiment of the present invention, an inter-satellite laser communication and signal scheduling method based on adaptive polarization modulation is provided, including:

[0007] Obtain the polarization state information of the transmitting light beam of the first inter-satellite laser communication terminal and the polarization state information of the receiving light beam of the second inter-satellite laser communication terminal, calculate the polarization state mismatch degree according to the light beam polarization state information and the receiving light beam polarization state information, and generate an initial polarization state compensation parameter based on the polarization state mismatch degree;

[0008] Construct a polarization state prediction model based on a deep learning neural network, the input of the polarization state prediction model includes atmospheric disturbance data, satellite attitude data, and historical polarization state data, the output includes the predicted polarization state drift trend, and pre-correct the polarization state compensation parameter according to the polarization state drift trend;

[0009] Use a polarization modulator to perform real-time polarization state modulation on the transmitting light beam, adjust the modulation parameter of the polarization modulator according to the pre-corrected polarization state compensation parameter, establish an inter-satellite laser communication link quality evaluation model based on the modulation parameter of the polarization modulator, and calculate the link quality parameter;

[0010] Based on the link quality parameters, an adaptive optical carrier recovery technique based on coherent detection uses an optical frequency comb as the local oscillator light source, achieves carrier synchronization through a digital phase-locked loop, and uses an adaptive equalization algorithm to compensate for channel dispersion and nonlinear effects, dynamically adjusting the equalizer parameters according to the channel state information;

[0011] Transmit the demodulated communication signal through the inter-satellite laser communication link, and monitor the link state parameters in real time. Update the training data set of the deep learning neural network according to the link state parameters to realize the online optimization of the polarization state prediction model.

[0012] In an alternative embodiment,

[0013] Calculate the polarization state mismatch degree based on the beam polarization state information and the received beam polarization state information, and generate initial polarization state compensation parameters based on the polarization state mismatch degree, including:

[0014] Obtain the transmitted beam polarization state information and the received beam polarization state information, where the polarization state information includes the total light intensity, the difference between the horizontal and vertical polarization components, the difference between the positive and negative forty-five degree polarization components, and the difference between the right-handed and left-handed circular polarization components;

[0015] Construct a fourth-order Mueller matrix, which is composed of a phase delay matrix, a rotation matrix, and a depolarization matrix. Perform eigenvalue decomposition on the fourth-order Mueller matrix to obtain the contribution components of each matrix, and calculate the polarization state mismatch degree based on the contribution components. The polarization state mismatch degree is determined by the ratio of the square root of each contribution component to the sum of the square roots of all contribution components;

[0016] Construct a compensation parameter optimization objective function based on the polarization state mismatch degree. The compensation parameters include a phase compensation amount and a rotation angle compensation amount. Use the gradient iteration method to solve the optimization objective function to obtain the initial polarization state compensation parameters, and compensate and modulate the polarization state of the transmitted beam through a polarization modulator according to the initial polarization state compensation parameters.

[0017] In an alternative embodiment,

[0018] Construct a polarization state prediction model based on a deep learning neural network. The input of the polarization state prediction model includes atmospheric disturbance data, satellite attitude data, and historical polarization state data, and the output includes the predicted polarization state drift trend. Pre-correct the polarization state compensation parameters according to the polarization state drift trend, including:

[0019] Collect the atmospheric temperature distribution, atmospheric pressure distribution, humidity distribution, and wind speed distribution to form atmospheric disturbance data, collect the roll angle, pitch angle, yaw angle, and angular velocity vector to form satellite attitude data, and obtain the Stokes parameter vectors at multiple sampling moments to form a historical polarization state sequence, where the sampling interval of the historical polarization state sequence is a preset time interval;

[0020] Input the atmospheric disturbance data, the satellite attitude data, the historical polarization state sequence, and the initial polarization state compensation parameters into a recurrent neural network with a hybrid attention mechanism. The recurrent neural network calculates the hidden layer state at the current moment based on long short-term memory units, generates attention weights for the atmospheric disturbance data and the satellite attitude data according to the hidden layer state, and performs weighted fusion of the attention weights with the atmospheric disturbance data and the satellite attitude data respectively to obtain a first feature vector and a second feature vector. Then, perform feature fusion on the first feature vector, the second feature vector, and the hidden layer state to obtain feature fusion data;

[0021] Input the feature fusion data into a pre-trained neural network mapping function to predict the Stokes parameters after the next preset time interval. Calculate the difference between the Stokes parameters and the Stokes parameters at the latest moment in the historical polarization state sequence to obtain a polarization state difference, divide the polarization state difference by the preset time interval to obtain a polarization state drift velocity vector, construct a pre-compensation objective function including a second-order regularization term of the pre-compensation parameters according to the polarization state drift velocity vector, and use the gradient descent method in combination with a deep reinforcement learning framework to iteratively optimize the pre-compensation objective function to obtain pre-compensation parameters, and use the pre-compensation parameters to pre-compensate the Stokes parameters at the latest moment in the historical polarization state sequence.

[0022] In an alternative embodiment,

[0023] Constructing a pre-compensation objective function including a second-order regularization term of the pre-compensation parameters according to the polarization state drift velocity vector, and using the gradient descent method in combination with a deep reinforcement learning framework to iteratively optimize the pre-compensation objective function to obtain pre-compensation parameters includes:

[0024] Construct a system state vector including the current polarization state, polarization state drift velocity vector, compensation error, and compensation parameters, and construct an action vector from the rotation angle parameter adjustment amount and the phase parameter adjustment amount;

[0025] Calculate the square term of the difference between the product of the compensation matrix and the predicted polarization state and the target polarization state based on the system state vector, and construct an immediate reward function in combination with the square terms of the rotation angle parameter adjustment amount and the phase parameter adjustment amount. Use a discount factor to accumulate the immediate rewards within the future prediction step count to obtain a pre-compensation objective function;

[0026] Construct a deep neural network as the value network and the policy network based on the pre-compensation objective function. The value network takes the system state vector and the action vector as inputs and outputs a value evaluation. The policy network takes the system state vector as an input and outputs an action policy. Store the state transition samples in the experience replay buffer;

[0027] Construct a temporal difference objective function based on the immediate reward function and the value network, construct a policy gradient objective function based on the value network and the policy network, and introduce the policy distribution entropy to construct an entropy regularization objective function;

[0028] Use the gradient descent method to optimize and update the network parameters of the value network for the temporal difference objective function, optimize and update the network parameters of the policy network for the weighted sum of the policy gradient objective function and the entropy regularization objective function, and use the soft update method to update the target network parameters;

[0029] Inject Gaussian noise into the action policy output by the policy network to obtain the actual executed action, calculate the variance of the Gaussian noise based on the compensation error, and use the actual executed action as the update amount of the compensation parameter to iteratively optimize the compensation parameter to obtain the pre-compensation parameter.

[0030] In an alternative embodiment,

[0031] Use a polarization modulator to perform real-time polarization state modulation on the transmitted beam, adjust the modulation parameters of the polarization modulator according to the pre-calibrated polarization state compensation parameters, establish an inter-satellite laser communication link quality evaluation model based on the modulation parameters of the polarization modulator, and calculate the link quality parameters including:

[0032] Use a polarization modulator to perform real-time polarization state modulation on the transmitted beam, adjust the modulation parameters of the polarization modulator according to the pre-calibrated polarization state compensation parameters, and the modulation parameters include the polarization state rotation angle and the polarization state ellipticity;

[0033] Establish an inter-satellite laser communication link quality evaluation model based on the modulation parameters of the polarization modulator, and the link quality evaluation model includes the influence factor of atmospheric disturbance on the polarization state, the polarization state analysis factor at the receiving end, and the system polarization extinction ratio factor;

[0034] Calculate the link quality parameters according to the link quality evaluation model, and update the polarization state compensation parameters in real time according to the link quality parameters.

[0035] In an alternative embodiment,

[0036] Based on the link quality parameters, for the adaptive optical carrier recovery technology based on coherent detection, an optical frequency comb is used as the local oscillator light source, carrier synchronization is achieved through a digital phase-locked loop, and an adaptive equalization algorithm is used to compensate for channel dispersion and nonlinear effects. Dynamically adjusting the equalizer parameters according to the channel state information includes:

[0037] Obtain the link quality parameters, use an optical frequency comb as the local oscillator light source, generate frequency comb spectral lines according to the preset center frequency and comb tooth interval, and the frequency comb spectral lines are determined by the amplitude, frequency and initial phase of the frequency components. Monitor the temperature drift and driving voltage fluctuation, calculate the frequency drift according to the product of the temperature sensitivity coefficient and the temperature drift and the product of the voltage sensitivity coefficient and the driving voltage fluctuation, and perform real-time compensation on the frequency of the optical frequency comb;

[0038] Perform coherent detection on the received optical signal and the optical frequency comb to obtain a beat signal, perform complex conjugate multiplication on the beat signal and the local reference signal to obtain a phase error, input the phase error into a digital phase-locked loop, and use a proportional-integral controller to filter the phase error. Calculate the control voltage according to the weighted sum of the product of the phase error at the current moment and the proportional coefficient and the cumulative sum of the historical phase errors multiplied by the integral coefficient;

[0039] Calculate the output phase of the numerically controlled oscillator according to the control voltage, and the output phase is obtained by adding the output phase at the previous moment, the product of the center frequency and the sampling period, and the product of the control voltage and the gain coefficient. Use the output phase as the phase parameter of the local reference signal to achieve carrier synchronization;

[0040] Establish a channel transmission model, which includes a dispersion compensation model and a nonlinear effect compensation model. The dispersion compensation model calculates the dispersion compensation coefficient according to the optical signal wavelength, transmission distance and the speed of light. The nonlinear effect compensation model calculates the nonlinear compensation coefficient according to the input signal, nonlinear coefficient, signal power and effective length. Perform feature extraction and optimize the loss function design through deep learning technology, and combine with the adaptive equalization algorithm to compensate for the channel dispersion and the nonlinear effect. Multiply multiple delayed signals of the input signal by the equalizer coefficients respectively and sum them to obtain the output signal of the adaptive equalization algorithm;

[0041] Calculate the adaptive step factor according to the link quality parameters, use the product of the adaptive step factor, the error signal and the conjugate signal of the input signal delayed signal as the update amount to update the equalizer coefficients, and dynamically adjust the size of the adaptive step factor based on the link quality parameters to achieve adaptive optimization of the equalizer coefficients and output the demodulated communication signal.

[0042] In an optional implementation manner,

[0043] Feature extraction is performed through deep learning techniques and the loss function design is optimized. Combining with an adaptive equalization algorithm to compensate for the channel dispersion and the nonlinear effect. Multiplying multiple delayed signals of the input signal by the equalizer coefficients respectively and summing them to obtain the output signal of the adaptive equalization algorithm, which includes:

[0044] Performing time-domain feature extraction on the input signal using a sliding time window with a preset length to obtain the signal amplitude, signal phase change, and signal instantaneous power. Performing a fast Fourier transform on the signal data within the sliding time window to obtain a frequency-domain signal, and extracting the spectral amplitude, channel dispersion characteristics, and nonlinear effect characteristics of the frequency-domain signal. Combining the signal amplitude, the signal phase change, the signal instantaneous power, the spectral amplitude, the channel dispersion characteristics, and the nonlinear effect characteristics to form a feature vector;

[0045] Inputting the feature vector into a first neural network for feature dimensionality reduction to obtain a first feature. The first neural network includes a weight matrix and a bias vector. Inputting the first feature into a long short-term memory network to obtain a hidden state feature. The long short-term memory network is used to model temporal characteristics. Inputting the hidden state feature into a fully connected neural network to obtain the equalizer coefficients;

[0046] Compensating for the channel dispersion and the nonlinear effect. Multiplying multiple delayed signals of the input signal by the equalizer coefficients respectively and summing them to obtain the output signal of the adaptive equalization algorithm. Subtracting the output signal of the adaptive equalization algorithm from the desired output signal to obtain an error signal. Updating the equalizer coefficients according to the error signal, an adaptive step factor, the conjugate signal of the delayed signal of the input signal, and a fusion factor;

[0047] Constructing a comprehensive loss function including a mean square error loss term, a bit error rate loss term, and a peak-to-average power ratio loss term. Using the gradient descent method to optimize the network parameters of the first neural network, the network parameters of the long short-term memory network, and the network parameters of the fully connected neural network according to the comprehensive loss function;

[0048] Calculating a channel state index according to the signal-to-noise ratio, the error vector magnitude, and the bit error rate. Dynamically adjusting the adaptive step factor according to the channel state index and the error signal. Updating the fusion factor according to the channel state index and the comprehensive loss function.

[0049] In an optional implementation manner,

[0050] Transmit the demodulated communication signal through the inter-satellite laser communication link, and monitor the link status parameters in real time. Update the training data set of the deep learning neural network according to the link status parameters to achieve online optimization of the polarization state prediction model, including:

[0051] Transmit the demodulated communication signal through the inter-satellite laser communication link, and use a polarization state analyzer to monitor the link status parameters of the inter-satellite laser communication link in real time. The link status parameters include signal-to-noise ratio, bit error rate, and polarization extinction ratio;

[0052] Form training sample pairs by combining the link status parameters with the corresponding polarization state parameters, and update the training data set of the deep learning neural network according to the training sample pairs. The deep learning neural network is used to establish a polarization state prediction model;

[0053] Train the deep learning neural network with the training data set, adjust the network learning rate according to the signal-to-noise ratio in the link status parameters, adjust the weight coefficient of the network optimization objective function according to the bit error rate in the link status parameters, and dynamically adjust the network structure parameters according to the polarization extinction ratio in the link status parameters to achieve online optimization of the polarization state prediction model.

[0054] In the present invention, through the method combining adaptive polarization modulation and deep learning prediction, the polarization state mismatch in inter-satellite laser communication can be effectively compensated, the stability and reliability of the communication link can be significantly improved, the probability of communication interruption can be reduced, an optical frequency comb is used as the local oscillator light source, combined with a digital phase-locked loop and an adaptive equalization algorithm, high-precision carrier synchronization and channel compensation are achieved, the influence of atmospheric disturbance and nonlinear effects is effectively overcome, the anti-interference ability and signal quality of the communication system are improved, and by monitoring the link status in real time and updating the training data of the deep learning model, dynamic optimization and adaptive adjustment of the polarization state prediction model are achieved, so that the system has strong environmental adaptability and self-learning ability, can continuously improve the communication performance, and extend the effective communication time. Description of the Drawings

[0055] Figure 1 It is a schematic flowchart of the inter-satellite laser communication and signal scheduling method based on adaptive polarization modulation according to an embodiment of the present invention;

[0056] Figure 2 It is a compensation error convergence curve graph of the inter-satellite laser communication and signal scheduling method based on adaptive polarization modulation according to an embodiment of the present invention;

[0057] Figure 3 It is a system convergence performance comparison graph of the inter-satellite laser communication and signal scheduling method based on adaptive polarization modulation according to an embodiment of the present invention. Detailed Embodiments

[0058] 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.

[0059] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments may be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0060] Figure 1 It is a schematic flowchart of an inter-satellite laser communication and signal scheduling method based on adaptive polarization modulation according to an embodiment of the present invention. As Figure 1 shown, the method includes:

[0061] Obtain the polarization state information of the transmitting beam of the first inter-satellite laser communication terminal and the polarization state information of the receiving beam of the second inter-satellite laser communication terminal. Calculate the polarization state mismatch degree according to the beam polarization state information and the receiving beam polarization state information, and generate an initial polarization state compensation parameter based on the polarization state mismatch degree;

[0062] Construct a polarization state prediction model based on a deep learning neural network. The input of the polarization state prediction model includes atmospheric disturbance data, satellite attitude data, and historical polarization state data, and the output includes the predicted polarization state drift trend. Pre-correct the polarization state compensation parameter according to the polarization state drift trend;

[0063] Use a polarization modulator to perform real-time polarization state modulation on the transmitting beam. Adjust the modulation parameter of the polarization modulator according to the pre-corrected polarization state compensation parameter. Establish an inter-satellite laser communication link quality evaluation model based on the modulation parameter of the polarization modulator, and calculate the link quality parameter;

[0064] According to the link quality parameter, based on the coherent detection-based adaptive optical carrier recovery technology, use an optical frequency comb as the local oscillator light source, achieve carrier synchronization through a digital phase-locked loop, and use an adaptive equalization algorithm to compensate for channel dispersion and nonlinear effects, and dynamically adjust the equalizer parameter according to the channel state information;

[0065] Transmit the demodulated communication signal through the inter-satellite laser communication link, and monitor the link state parameter in real time. Update the training data set of the deep learning neural network according to the link state parameter to realize the online optimization of the polarization state prediction model.

[0066] In an alternative embodiment,

[0067] Based on the emitted beam polarization state information and the received beam polarization state information, calculate the polarization state mismatch degree, and generate an initial polarization state compensation parameter based on the polarization state mismatch degree, including:

[0068] Obtain the emitted beam polarization state information and the received beam polarization state information, where the polarization state information includes the total light intensity, the difference between the horizontal and vertical polarization components, the difference between the positive and negative forty-five degree polarization components, and the difference between the right-handed and left-handed circular polarization components;

[0069] Construct a fourth-order Mueller matrix, which is composed of a phase delay matrix, a rotation matrix, and a depolarization matrix. Perform eigenvalue decomposition on the fourth-order Mueller matrix to obtain the contribution components of each matrix, and calculate the polarization state mismatch degree based on the contribution components. The polarization state mismatch degree is determined by the ratio of the square root of each contribution component to the sum of the square roots of all contribution components;

[0070] Based on the polarization state mismatch degree, construct an optimization objective function for the compensation parameter. The compensation parameter includes a phase compensation amount and a rotation angle compensation amount. Use the gradient iteration method to solve the optimization objective function to obtain the initial polarization state compensation parameter, and compensate and modulate the polarization state of the emitted beam through a polarization modulator according to the initial polarization state compensation parameter.

[0071] Obtain the polarization state information of the emitted beam and the received beam. Separate the incident light into two polarization components, horizontal and vertical, through a polarization beam splitter, and measure the light intensities of the two components using a photodetector. At the same time, use a wave plate and a polarization analyzer to obtain the polarization components of positive and negative forty-five degrees, and then obtain the right-handed and left-handed circular polarization components through a quarter-wave plate. Based on the measurement results, obtain the total light intensity, the difference between the horizontal and vertical polarization components, the difference between the positive and negative forty-five degree polarization components, and the difference between the right-handed and left-handed circular polarization components, which constitute the complete polarization state information.

[0072] After obtaining the polarization state information, construct a fourth-order Mueller matrix. This matrix is obtained by multiplying three basic matrices: a phase delay matrix, a rotation matrix, and a depolarization matrix. Among them, the phase delay matrix describes the phase delay characteristics of light waves during transmission, the rotation matrix characterizes the rotational change of the polarization state, and the depolarization matrix reflects the depolarization effect during the channel transmission process. Perform eigenvalue decomposition on the constructed fourth-order Mueller matrix to obtain the contribution components of these three basic matrices. By calculating the square root of each contribution component and taking the ratio of it to the sum of the square roots of all contribution components, an accurate polarization state mismatch degree can be obtained.

[0073] Based on the calculated polarization state mismatch degree, an optimization objective function for compensation parameters is constructed. This function takes the phase compensation amount and the rotation angle compensation amount as optimization variables, and the goal is to minimize the polarization state mismatch degree. The gradient iteration method is used to solve this optimization objective function. The specific process is as follows: First, calculate the gradients of the objective function with respect to the two compensation parameters, and then iteratively update the compensation parameters according to the gradient direction until the algorithm converges to obtain the optimal compensation parameter values. Finally, these optimized compensation parameters are input into the polarization modulator to perform real-time compensation modulation on the polarization state of the transmitted beam.

[0074] Exemplarily, in a certain satellite laser communication experiment, first, the polarization state information of the transmitted beam is measured: the total light intensity is 1.0 milliwatt, the difference between the horizontal and vertical polarization components is 0.3 milliwatt, the difference between the positive and negative forty-five-degree polarization components is -0.2 milliwatt, and the difference between the right-handed and left-handed circular polarization components is 0.1 milliwatt. Based on these measurement data, a fourth-order Mueller matrix describing the channel characteristics is constructed.

[0075] Perform eigenvalue decomposition on this fourth-order Mueller matrix, and the four eigenvalues obtained are 0.95, 0.85, 0.80, and 0.75 respectively. Calculate the square roots of each eigenvalue to obtain the contribution components: 0.975, 0.922, 0.894, and 0.866. The sum of these contribution components is 3.657. Further calculate the normalized contribution degrees to be 0.267, 0.252, 0.244, and 0.237 respectively.

[0076] Based on these data, an optimization objective function is constructed, and the optimal compensation parameters are obtained by solving through the gradient iteration method: the phase compensation amount is 0.35π, and the rotation angle compensation amount is 0.28π. After applying these compensation parameters to the polarization modulator, the polarization state mismatch degree of the system is reduced from the original 0.45 to 0.12.

[0077] In this embodiment, by introducing complete polarization state characterization parameters, the system's perception ability of the channel polarization characteristics is improved. Through the Mueller matrix decomposition method, the decoupling analysis of complex channel effects is realized. By solving the compensation parameters through the optimization algorithm, the optimality of the compensation effect is ensured. The systematic improvement significantly improves the accuracy of polarization state compensation, enhances the system's adaptability to channel perturbations, and effectively reduces the degradation of communication performance caused by polarization state mismatch.

[0078] In an alternative embodiment,

[0079] A polarization state prediction model is constructed based on a deep learning neural network. The input of the polarization state prediction model includes atmospheric disturbance data, satellite attitude data, and historical polarization state data, and the output includes the predicted polarization state drift trend. Pre-correcting the polarization state compensation parameters according to the polarization state drift trend includes:

[0080] Collect the atmospheric temperature distribution, atmospheric pressure distribution, humidity distribution, and wind speed distribution to form atmospheric disturbance data, collect the roll angle, pitch angle, yaw angle, and angular velocity vector to form satellite attitude data, and obtain the Stokes parameter vectors at multiple sampling moments to form a historical polarization state sequence, where the sampling interval of the historical polarization state sequence is a preset time interval;

[0081] Input the atmospheric disturbance data, the satellite attitude data, the historical polarization state sequence, and the initial polarization state compensation parameters into a recurrent neural network with a hybrid attention mechanism. The recurrent neural network calculates the hidden layer state at the current moment based on long short-term memory units, generates attention weights for the atmospheric disturbance data and the satellite attitude data according to the hidden layer state, and performs weighted fusion of the attention weights with the atmospheric disturbance data and the satellite attitude data respectively to obtain a first feature vector and a second feature vector. Then, perform feature fusion on the first feature vector, the second feature vector, and the hidden layer state to obtain feature fusion data;

[0082] Input the feature fusion data into a pre-trained neural network mapping function to predict the Stokes parameters after the next preset time interval. Calculate the difference between the Stokes parameters and the Stokes parameters at the latest moment in the historical polarization state sequence to obtain a polarization state difference. Divide the polarization state difference by the preset time interval to obtain a polarization state drift velocity vector. Construct a pre-compensation objective function containing a second-order regularization term of the pre-compensation parameters, and use the gradient descent method to iteratively optimize the pre-compensation objective function in combination with a deep reinforcement learning framework to obtain pre-compensation parameters, and use the pre-compensation parameters to pre-compensate the Stokes parameters at the latest moment in the historical polarization state sequence.

[0083] Perform the data collection and preprocessing stage. The system collects complete atmospheric disturbance data including atmospheric temperature distribution, atmospheric pressure distribution, humidity distribution, and wind speed distribution through a distributed atmospheric sensor network. At the same time, obtain real-time data of the roll angle, pitch angle, and yaw angle from the satellite attitude control system, as well as the angular velocity vector information of the satellite body to form a satellite attitude data set. In addition, continuously collect the Stokes parameter vectors at multiple moments at a preset fixed time interval to form a historical polarization state sequence database.

[0084] The collected multi-source data is input into a recurrent neural network with a hybrid attention mechanism for processing. This network uses long short-term memory units as the basic computational units, which can effectively capture the long-term dependencies in the data sequence, calculate the hidden layer state at the current moment, and then calculate the attention weights for the atmospheric disturbance data and satellite attitude data respectively based on this hidden layer state. This attention mechanism enables the network to adaptively focus on the importance of different data features. Subsequently, the system performs weighted fusion of the calculated attention weights with the atmospheric disturbance data and satellite attitude data respectively to obtain two feature vectors. Finally, these feature vectors are subjected to deep feature fusion with the hidden layer state to generate feature fusion data containing multi-dimensional information.

[0085] The feature fusion data is input into a pre-trained neural network mapping function. Through learning from historical data, this mapping function can predict the Stokes parameters after the next time interval. The system calculates the difference between the predicted Stokes parameters and the latest recorded Stokes parameters in the historical polarization state sequence, and divides the difference by the preset time interval to obtain the drift velocity vector reflecting the change trend of the polarization state. Based on the drift velocity vector, an objective function containing the second-order regularization term of the pre-compensation parameters is constructed, and the gradient descent method combined with the deep reinforcement learning framework is used to optimize the objective function to obtain the optimal pre-compensation parameters, and these parameters are used to pre-compensate the current polarization state.

[0086] Exemplarily, assume that in a certain satellite laser communication mission, the following data is collected: the atmospheric temperature distribution fluctuates in the range of -60°C to 30°C, the pressure distribution drops from 1 atmosphere to 0.1 atmosphere, the relative humidity distribution varies between 0 - 80%, and the wind speed distribution shows that the maximum wind speed reaches 15 m / s. The satellite attitude data shows that the roll angle fluctuates within the range of ±2 degrees, the pitch angle remains at about 45 degrees, the yaw angle is controlled within the range of ±1 degree, and the maximum component of the angular velocity vector does not exceed 0.1 degree / s. The system continuously collects 1000 groups of Stokes parameter vectors at a time interval of 1 millisecond.

[0087] The parameter vector is input into the hybrid attention recurrent neural network, which contains 128 LSTM units, and the dimension of the attention mechanism layer is 64. The network assigns a weight of 0.6 to the atmospheric disturbance data and a weight of 0.4 to the attitude data. After feature fusion, the data dimension is compressed to 32 dimensions. The pre-trained neural network mapping function predicts the Stokes parameters at the next time point. By taking the difference with the current parameters and dividing by the time interval of 1 millisecond, the drift trend of the polarization state is obtained. Finally, through 500 iterations of optimization, the optimal pre-compensation parameters are obtained, realizing the pre-compensation of the polarization state.

[0088] In this embodiment, through the fusion analysis of multi-source data, the system's perception ability of the communication environment is improved. By introducing a hybrid attention mechanism, the system's ability to extract key features is enhanced. Through the establishment of a prediction model, an accurate prediction of the change trend of the polarization state is achieved.

[0089] In an alternative embodiment,

[0090] Construct a pre-compensation objective function including a second-order regularization term of pre-compensation parameters based on the polarization state drift velocity vector. Using the gradient descent method and combining with a deep reinforcement learning framework, iteratively optimize the pre-compensation objective function to obtain pre-compensation parameters, including:

[0091] Construct a system state vector including the current polarization state, polarization state drift velocity vector, compensation error, and compensation parameters, and construct the rotation angle parameter adjustment amount and phase parameter adjustment amount as an action vector;

[0092] Calculate the square term of the difference between the product of the compensation matrix and the predicted polarization state and the target polarization state based on the system state vector, and construct an immediate reward function in combination with the square terms of the rotation angle parameter adjustment amount and the phase parameter adjustment amount. Use a discount factor to accumulate the immediate rewards within the future prediction steps to obtain the pre-compensation objective function;

[0093] Construct a deep neural network as a value network and a policy network based on the pre-compensation objective function. The value network inputs the system state vector and the action vector and outputs a value evaluation. The policy network inputs the system state vector and outputs an action policy, and store the state transition samples in an experience replay buffer;

[0094] Construct a temporal difference objective function based on the immediate reward function and the value network, construct a policy gradient objective function based on the value network and the policy network, and introduce a policy distribution entropy to construct an entropy regularization objective function;

[0095] Use the gradient descent method to optimize and update the network parameters of the value network for the temporal difference objective function, optimize and update the network parameters of the policy network for the weighted sum of the policy gradient objective function and the entropy regularization objective function, and use a soft update method to update the target network parameters;

[0096] Inject Gaussian noise into the action policy output by the policy network to obtain the actual executed action. Calculate the variance of the Gaussian noise based on the compensation error, and use the actual executed action as the update amount of the compensation parameter to iteratively optimize the compensation parameter to obtain the pre-compensation parameter.

[0097] Construct a complete system state vector, which consists of four key components: the currently measured polarization state, the calculated polarization state drift velocity vector, the compensated error monitored in real time, and the currently used compensation parameters. At the same time, form an action vector from the adjustment amounts of the rotation angle parameter and the phase parameter to be optimized, and output it as the control quantity.

[0098] Construct an immediate reward function, which consists of two parts: the first part is the square term of the difference between the product of the compensation matrix and the predicted polarization state and the target polarization state, used to measure the compensation effect; the second part is the square term of the adjustment amounts of the rotation angle parameter and the phase parameter, used to constrain the parameter change range. By introducing a discount factor, the immediate rewards within the future prediction step count are weighted and accumulated to form a pre-compensation objective function.

[0099] Under the deep reinforcement learning framework, construct a value network and a policy network respectively. The value network receives the system state vector and the action vector as inputs and outputs the value evaluation of the current state-action combination. The policy network receives the system state vector as input and outputs the corresponding action policy. During the training process, store the experience samples of state transitions in the experience replay buffer for subsequent offline learning.

[0100] Based on the constructed immediate reward function and value network, establish a temporal difference objective function to evaluate and optimize the prediction ability of the value network. At the same time, construct a policy gradient objective function based on the value network and the policy network, and introduce the policy distribution entropy to construct an entropy regularization objective function to improve the exploration efficiency.

[0101] Use the gradient descent method to optimize these objective functions respectively: the optimization of the temporal difference objective function is used to update the value network parameters; the weighted sum of the policy gradient objective function and the entropy regularization objective function is optimized to update the policy network parameters. At the same time, use the soft update method to update the target network parameters to ensure the stability of the training process.

[0102] Inject Gaussian noise into the action policy output by the policy network to generate the actual executed action. The variance of the Gaussian noise is dynamically adjusted according to the compensated error, increasing the exploration intensity when the compensated error is large and decreasing the exploration range vice versa. Use the actual executed action as the update amount of the compensation parameters to iteratively optimize the compensation parameters and finally obtain the pre-compensation parameters.

[0103] Exemplarily, assume that during a certain polarization state compensation process, the initial values of the state vectors are: the current polarization state [1.0, 0.5, -0.3, 0.2], the polarization state drift velocity vector [0.01, -0.02, 0.015, -0.005], the compensation error 0.1, and the current compensation parameters [0.3, 0.4]. The initial value of the action vector is set as: the rotation angle adjustment amount 0.05, the phase parameter adjustment amount 0.03, and the constructed immediate reward function is: R = -||M(θ, φ)S_pred - S_target||² - α(θ² + φ²), where M is the compensation matrix, S_pred is the predicted polarization state, S_target is the target polarization state, and α is the weight factor. The discount factor γ is set to 0.95, and the number of prediction steps is 10 steps;

[0104] The value network adopts a three-layer fully connected network structure. The number of nodes in the input layer is the sum of the dimensions of the state vector and the action vector, the number of nodes in the hidden layer is 64, and the output layer is a single node. The policy network also adopts a three-layer fully connected structure. The number of nodes in the input layer is the dimension of the state vector, the number of nodes in the hidden layer is 64, and the number of nodes in the output layer is the dimension of the action vector.

[0105] The capacity of the experience replay buffer is set to 10000, and the batch size for each sampling is 128. The temporal difference objective function is constructed using the TD(0) algorithm, and the policy gradient objective function is estimated using the advantage function. The entropy regularization coefficient is set to 0.01.

[0106] During the training process, the learning rates of both the value network and the policy network are set to 0.001, and the soft update coefficient of the target network is 0.005. Gaussian noise with a mean of 0 is injected into the action policy output by the policy network. The initial variance is set to 0.1 and decays as the compensation error decreases.

[0107] After 1000 rounds of iterative optimization, the compensation error is reduced from the initial 0.1 to 0.02. The finally obtained pre-compensation parameters are: the rotation angle compensation amount 0.42, and the phase compensation amount 0.35. These parameters significantly reduce the mismatch between the actual polarization state and the target polarization state, achieving an accurate pre-compensation effect.

[0108] In this embodiment, by considering the long-term benefits, it breaks through the limitation of traditional methods that only focus on immediate effects, obtains a better compensation effect, realizes the accurate modeling of complex non-linear mapping relationships with the help of the powerful expression ability of deep neural networks, enables the compensation strategy to be continuously optimized and adapted to environmental changes through the online learning of the policy network, and enhances the adaptability of the algorithm to unknown states by introducing an exploration mechanism;

[0109] The polarization state compensation parameter optimization method in the prior art mainly uses a single gradient descent or heuristic algorithm for solution. The optimization process only considers the compensation error of the current state, lacks consideration of long-term benefits, is prone to falling into local optimal solutions, has a fixed optimization strategy and lacks exploration ability, and is difficult to adapt to complex and changeable space channel environments, resulting in unstable compensation effects. In addition, traditional methods often regard the optimization of compensation parameters as an independent optimization problem and ignore the continuity characteristics of state transitions during the compensation process, making it difficult for the optimization results to meet the actual application requirements. In this embodiment, a complete state vector including the current polarization state, drift velocity vector, compensation error, and compensation parameters is constructed, and the adjustment amount of the compensation parameters is used as the action vector to establish a mapping relationship from the state space to the action space. By designing an immediate reward function including a compensation effect term and a parameter constraint term, and accumulating future benefits in combination with a discount factor, the optimization of long-term compensation effects is achieved;

[0110] This embodiment adopts a dual-network architecture, separates value evaluation and policy generation. The value network is responsible for evaluating the long-term value of state-action combinations, and the policy network is responsible for generating the optimal compensation parameter adjustment strategy. Through the experience replay mechanism, the correlation between samples is broken, and the learning efficiency is improved. The policy distribution entropy is introduced as a regularization term to maintain an appropriate exploration ability while ensuring the convergence of the policy. In summary, this embodiment significantly improves the overall performance of the satellite laser communication system, provides a new technical path for high-precision polarization state compensation, and is of great significance for improving the reliability of space laser communication.

[0111] In an alternative embodiment,

[0112] A polarization modulator is used to perform real-time polarization state modulation on the transmitted beam. The modulation parameters of the polarization modulator are adjusted according to the pre-calibrated polarization state compensation parameters. Based on the modulation parameters of the polarization modulator, a quality evaluation model for inter-satellite laser communication links is established, and the link quality parameters calculated include:

[0113] A polarization modulator is used to perform real-time polarization state modulation on the transmitted beam. The modulation parameters of the polarization modulator are adjusted according to the pre-calibrated polarization state compensation parameters. The modulation parameters include the polarization state rotation angle and the polarization state ellipticity;

[0114] Based on the modulation parameters of the polarization modulator, a quality evaluation model for inter-satellite laser communication links is established. The link quality evaluation model includes an influence factor of atmospheric disturbance on the polarization state, a polarization state analysis factor at the receiving end, and a system polarization extinction ratio factor;

[0115] The link quality parameters are calculated according to the link quality evaluation model, and the polarization state compensation parameters are updated in real time according to the link quality parameters.

[0116] Enter the polarization state modulation stage. According to the polarization state compensation parameters obtained in advance through calibration, the modulation parameters of the polarization modulator are accurately set, including two key parameters: the polarization state rotation angle and the polarization state ellipticity. After receiving these parameters, the polarization modulator performs real-time polarization state modulation on the emitted beam, and realizes precise control of the polarization state of the emitted beam by adjusting the relative intensity and phase difference of the beam in different polarization directions.

[0117] Establish a quality evaluation model for the inter-satellite laser communication link. This model contains three core evaluation factors: the first is the influence factor of atmospheric disturbance on the polarization state, which considers the disturbance effects of atmospheric turbulence, scattering, absorption, etc. on the polarization state; the second is the polarization state analysis factor at the receiving end, which describes the analysis ability of the receiving end optical system for the polarization state of the incident light; the third is the system polarization extinction ratio factor, which reflects the isolation degree between different polarization states in the communication system. These three factors are combined through a mathematical model to form a complete link quality evaluation model.

[0118] Perform link quality parameter calculation and compensation parameter update. According to the established link quality evaluation model, calculate the quality parameters of the current communication link. These quality parameters reflect the quality of the current polarization state modulation effect. Based on the calculated link quality parameters, the polarization state compensation parameters are updated in real time. When the link quality parameters are lower than the preset threshold, the update process of the compensation parameters is triggered. The new compensation parameter values are calculated through an optimization algorithm, and the updated parameters are re-input into the polarization modulator to form a closed-loop control.

[0119] Exemplarily, assume that during a certain inter-satellite laser communication process, the rotation angle in the initial polarization state compensation parameters is set to π / 4, and the ellipticity is set to 0.5. The polarization modulator modulates the emitted beam according to these parameters, so that the beam obtains a specific polarization state distribution;

[0120] In the link quality evaluation model, the expression of the atmospheric disturbance influence factor is exp(-σ²h / h0), where σ² represents the atmospheric turbulence intensity, h represents the transmission distance, and h0 is the atmospheric equivalent height. The polarization state analysis factor at the receiving end is represented by the Mueller matrix, which includes the polarization state response characteristics of the receiving optical system. The system polarization extinction ratio factor is defined as the logarithm of the power ratio between orthogonal polarization states.

[0121] Through real-time monitoring, it is found that the link quality parameters gradually decrease from the initial value of 0.9 to 0.7, which is lower than the preset threshold of 0.8. At this time, the compensation parameter update mechanism is triggered. After algorithm optimization calculation, the rotation angle is adjusted to π / 3, and the ellipticity is adjusted to 0.6. The updated compensation parameters are input into the polarization modulator to re-adjust the polarization state of the emitted beam, so that the link quality parameters are restored to the ideal level.

[0122] Through this closed-loop control of real-time modulation and evaluation, the stable operation of the inter-satellite laser communication link is ensured. The introduction of the evaluation model provides a reliable basis for the update of compensation parameters, improving the adaptability and reliability of the communication system.

[0123] In this embodiment, through precise polarization state modulation, the quality of the polarization state of the transmitted beam is improved. Through the multi-factor evaluation model, the accurate evaluation of the link quality is realized. Through the real-time parameter update mechanism, the continuous optimization ability of the system is ensured. In summary, this embodiment not only improves the accuracy of polarization state modulation, but also enhances the adaptive ability of the system, effectively solving various problems existing in traditional methods, providing a new technical path for high-quality inter-satellite laser communication, and having important significance for improving the overall performance of the space communication system.

[0124] In an alternative embodiment,

[0125] According to the link quality parameters, based on the adaptive optical carrier recovery technology of coherent detection, an optical frequency comb is used as the local oscillator light source, carrier synchronization is achieved through a digital phase-locked loop, and an adaptive equalization algorithm is used to compensate for channel dispersion and nonlinear effects. Dynamically adjusting the equalizer parameters according to the channel state information includes:

[0126] Obtain the link quality parameters, use an optical frequency comb as the local oscillator light source, generate frequency comb spectral lines according to the preset center frequency and comb tooth interval, and the frequency comb spectral lines are determined by the amplitude, frequency and initial phase of the frequency components. Monitor the temperature drift and drive voltage fluctuation, calculate the frequency drift according to the product of the temperature sensitivity coefficient and the temperature drift and the product of the voltage sensitivity coefficient and the drive voltage fluctuation, and perform real-time compensation on the frequency of the optical frequency comb;

[0127] The received optical signal is coherently detected with the optical frequency comb to obtain a beat signal. The beat signal is multiplied by the complex conjugate of the local reference signal to obtain a phase error. The phase error is input into a digital phase-locked loop, and a proportional-integral controller is used to filter the phase error. Calculate the control voltage according to the weighted sum of the product of the phase error at the current moment and the proportional coefficient and the cumulative sum of the historical phase errors multiplied by the integral coefficient;

[0128] Calculate the output phase of the numerically controlled oscillator according to the control voltage. The output phase is obtained by adding the output phase at the previous moment, the product of the center frequency and the sampling period, and the product of the control voltage and the gain coefficient. The output phase is used as the phase parameter of the local reference signal to achieve carrier synchronization;

[0129] A channel transmission model is established. The channel transmission model includes a dispersion compensation model and a non - linear effect compensation model. The dispersion compensation model calculates the dispersion compensation coefficient according to the optical signal wavelength, transmission distance, and the speed of light. The non - linear effect compensation model calculates the non - linear compensation coefficient according to the input signal, non - linear coefficient, signal power, and effective length. Feature extraction and optimized loss function design are carried out through deep learning technology. Combining with an adaptive equalization algorithm, the channel dispersion and the non - linear effect are compensated. Multiple delayed signals of the input signal are multiplied by the equalizer coefficients respectively and summed to obtain the output signal of the adaptive equalization algorithm;

[0130] The adaptive step - size factor is calculated according to the link quality parameter. The product of the adaptive step - size factor, the error signal, and the conjugate signal of the input signal delayed signal is used as the update amount to update the equalizer coefficients. Based on the link quality parameter, the size of the adaptive step - size factor is dynamically adjusted to achieve the adaptive optimization of the equalizer coefficients, and the demodulated communication signal is output.

[0131] The initial working state of the optical frequency comb is set according to the preset center frequency and comb - tooth interval parameters. The amplitude, frequency, and initial phase of each frequency component need to be precisely controlled. The working - environment temperature change is continuously monitored by a temperature sensor, and the temperature drift amount is recorded. At the same time, a high - precision voltage detection circuit is used to collect the fluctuation value of the driving voltage in real - time. The temperature - induced frequency offset is obtained by multiplying the temperature drift amount by the temperature - sensitive coefficient, and the voltage - induced frequency offset is obtained by multiplying the driving - voltage fluctuation amount by the voltage - sensitive coefficient. The two are added to obtain the total frequency drift amount. According to the calculated frequency drift amount, the driving parameters of the optical frequency comb are adjusted to achieve real - time frequency compensation and ensure that the output characteristics of the frequency comb remain stable.

[0132] The received optical signal and the output light of the compensated optical frequency comb are subjected to coherent detection, and a beat - frequency signal is obtained at the output end of the photodetector. The beat - frequency signal is subjected to a complex - conjugate multiplication operation with a digitally generated local reference signal to extract the phase - error information. After the phase - error signal is input into the digital phase - locked loop, it is processed by a proportional - integral controller. The controller adopts a two - way parallel structure: the proportional path multiplies the phase error at the current moment by the proportional coefficient to obtain the immediate control quantity; the integral path accumulates the historical phase errors and multiplies by the integral coefficient to obtain the cumulative control quantity. The two control quantities are added to obtain the final control voltage value.

[0133] After receiving the control voltage input, the numerically controlled oscillator first reads the output phase value of the previous moment, takes the product of the center frequency and the sampling period as the reference phase increment, and then takes the product of the control voltage and the oscillator gain coefficient as the correction phase amount. The sum of the three gives the output phase at the current moment. This output phase is used to generate a new local reference signal, completing a phase update process. Through continuous phase tracking, carrier synchronization is achieved.

[0134] When establishing a channel transmission model, first construct a dispersion compensation model. This model inputs the working wavelength of the optical signal, the actual transmission distance, and the speed of light in vacuum, and calculates the dispersion compensation coefficient. At the same time, construct a nonlinear effect compensation model. This model comprehensively considers the input signal characteristics, the fiber nonlinear coefficient, the signal transmission power, and the effective fiber length, and calculates the nonlinear compensation coefficient. Use a multi-layer convolutional neural network to extract features of the signal, design a loss function based on the mean square error, and combine a recurrent neural network to establish a channel response prediction model. Input the extracted features into an adaptive equalization algorithm, perform multi-stage delay expansion on the input signal, multiply them with the corresponding equalizer coefficients respectively and sum them up to obtain the equalized output signal.

[0135] Based on the real-time obtained link quality parameters, dynamically calculate the adaptive step size factor. Multiply this step size factor with the error signal and the conjugate signal of the input signal delay component to obtain the update amount of the equalizer coefficient. When the link quality parameter is low, increase the step size factor to accelerate the convergence speed; when the link quality parameter is high, decrease the step size factor to improve the stability. Through this adaptive update mechanism, the optimization adjustment of the equalizer coefficient is realized, and finally a stable demodulated communication signal is output.

[0136] Exemplarily, assume that the center frequency of the optical frequency comb is set to 193.1 THz and the comb tooth interval is 25 GHz. A temperature drift of 0.1 °C is monitored, and the temperature sensitivity coefficient is -1.5 GHz / °C; the drive voltage fluctuates by 0.01 V, and the voltage sensitivity coefficient is 2 GHz / V. The total frequency drift is calculated to be -0.13 GHz, and based on this, the drive parameters of the frequency comb are compensated and adjusted.

[0137] In coherent detection, the beat frequency signal has a frequency of 2 GHz and an amplitude of 0.5 V. The proportional coefficient of the digital phase-locked loop is set to 0.2, and the integral coefficient is 0.05. When the detected phase error is 0.1 rad, the calculated control voltage is 0.025 V.

[0138] The gain coefficient of the numerically controlled oscillator is set to 1 MHz / V, and the sampling period is 0.1 ns. When the output phase of the previous moment is 1.5 rad, the output phase at the current moment is calculated to be 1.5126 rad.

[0139] In the channel model, the wavelength of the optical signal is 1550 nm, the transmission distance is 100 km, and the calculated dispersion compensation coefficient is 17 ps / nm / km. The nonlinear coefficient is set to 2.6 W^(-1)km^(-1), the signal power is 1 mW, and the nonlinear compensation coefficient is calculated.

[0140] The equalization algorithm adopts an 11-tap structure, and the initial equalizer coefficients are [0.1, 0.2, 0.3, 0.4, 0.5, 1, 0.5, 0.4, 0.3, 0.2, 0.1]. When the link quality parameter is 0.9, the step size factor is set to 0.001; when the quality parameter drops to 0.7, the step size factor increases to 0.005 to accelerate the convergence process of the equalizer coefficients.

[0141] In this embodiment, through the frequency compensation mechanism of multi-factor coupling, the frequency stability of the optical frequency comb is significantly improved. Through the application of the digital phase-locked loop, faster and more accurate carrier synchronization is achieved. Through the deep learning-enhanced channel model, the accuracy of channel compensation is improved. Through the adaptive equalization algorithm, the adaptability of the system to channel changes is enhanced;

[0142] The existing inter-satellite laser communication systems usually adopt single temperature control or simple feedback regulation to maintain frequency stability, lacking comprehensive consideration of multi-source interference. In terms of channel compensation, conventional methods often deal with dispersion effects and nonlinear effects separately, and use equalization algorithms with fixed parameters, making it difficult to adapt to the complex and changing space channel environment. Traditional carrier synchronization technologies mainly rely on hardware phase-locked loops, with problems such as slow response speed and poor adaptability;

[0143] In this embodiment, the effects of temperature drift and driving voltage fluctuation are decoupled and analyzed. An accurate frequency drift compensation model is established through their respective sensitivity coefficients, realizing high-precision stable control of the optical frequency comb, achieving the transformation from single compensation to multi-dimensional collaboration, the improvement from fixed parameters to adaptive optimization, and the leap from independent processing to system integration. This not only improves the stability of the communication system but also enhances the environmental adaptability of the system, effectively solving various problems existing in traditional methods and realizing the overall improvement of the performance of the communication system, providing an innovative solution for the development of space laser communication systems.

[0144] In an alternative embodiment,

[0145] Through deep learning technology for feature extraction and optimization of loss function design, combined with the adaptive equalization algorithm to compensate for the channel dispersion and the nonlinear effect, multiplying the multiple delayed signals of the input signal by the equalizer coefficients and summing them to obtain the output signal of the adaptive equalization algorithm includes:

[0146] Perform time-domain feature extraction on the input signal using a sliding time window with a preset length to obtain the signal amplitude, signal phase change, and signal instantaneous power. Perform a fast Fourier transform on the signal data within the sliding time window to obtain a frequency-domain signal, and extract the spectral amplitude, channel dispersion characteristics, and nonlinear effect characteristics of the frequency-domain signal. Combine the signal amplitude, the signal phase change, the signal instantaneous power, the spectral amplitude, the channel dispersion characteristics, and the nonlinear effect characteristics to form a feature vector;

[0147] Input the feature vector into a first neural network for feature dimensionality reduction to obtain a first feature. The first neural network includes a weight matrix and a bias vector. Input the first feature into a long short-term memory network to obtain a hidden state feature. The long short-term memory network is used to model temporal features. Input the hidden state feature into a fully connected neural network to obtain an equalizer coefficient;

[0148] Compensate for the channel dispersion and the nonlinear effect. Multiply each of the multiple delayed signals of the input signal by the equalizer coefficient and sum them to obtain the output signal of the adaptive equalization algorithm. Subtract the output signal of the adaptive equalization algorithm from the desired output signal to obtain an error signal. Update the equalizer coefficient according to the error signal, the adaptive step factor, the conjugate signal of the delayed signal of the input signal, and the fusion factor;

[0149] Construct a comprehensive loss function that includes a mean square error loss term, a bit error rate loss term, and a peak-to-average power ratio loss term. Use the gradient descent method to optimize the network parameters of the first neural network, the network parameters of the long short-term memory network, and the network parameters of the fully connected neural network according to the comprehensive loss function;

[0150] Calculate the channel state index according to the signal-to-noise ratio, the error vector magnitude, and the bit error rate. Dynamically adjust the adaptive step factor according to the channel state index and the error signal. Update the fusion factor according to the channel state index and the comprehensive loss function.

[0151] Perform signal feature extraction. Set a sliding time window with a preset length, perform time-domain analysis on the input signal, and extract three basic features: obtain the signal amplitude by calculating the signal envelope; obtain the signal phase change by the phase difference between adjacent sampling points; obtain the signal instantaneous power by the square of the signal amplitude. Perform a fast Fourier transform on the signal data within the sliding time window, and extract three features in the frequency domain: obtain the spectral amplitude by calculating the magnitude of the frequency-domain signal; obtain the channel dispersion characteristics by the product of the dispersion coefficient and the signal bandwidth; obtain the nonlinear effect characteristics by the product of the nonlinear coefficient and the signal power. Combine these six features in sequence to form a feature vector.

[0152] The eigenvector is input into the first neural network for dimensionality reduction. This network consists of multiple layers, each layer containing a weight matrix and a bias vector. The features are transformed through a non-linear activation function to obtain the first reduced-dimensional feature. The first feature is input into the long short-term memory network in chronological order. This network contains three control units: an input gate, a forget gate, and an output gate, which model the temporal information through a gating mechanism and output the hidden state features. The hidden state features are processed by a fully-connected neural network, which maps the features to the same dimension as the number of equalizer taps to obtain the equalizer coefficients.

[0153] The signal is compensated using the obtained equalizer coefficients. The input signal passes through multiple different delay units to obtain a sequence of delayed signals. Each delayed signal is multiplied by the corresponding equalizer coefficient, and all the products are added to obtain the output signal of the equalization algorithm. The output signal is subtracted from the pre-determined desired output signal to obtain the error signal. Based on the error signal, the current adaptive step size factor, the conjugate signal of the input signal delay sequence, and the fusion factor, the update amount of the equalizer coefficients is calculated to update the equalizer coefficients.

[0154] A comprehensive loss function is constructed to optimize the neural network. The loss function consists of three terms: the mean square error loss term between the output signal and the desired signal, the loss term of the system bit error rate, and the loss term of the peak-to-average power ratio of the output signal. The gradient descent method is used to calculate the gradients of the loss function with respect to each network parameter, and the weight matrix and bias vector of the first neural network, the gating parameters of the long short-term memory network, and the connection weights of the fully-connected neural network are updated respectively.

[0155] The channel state metrics are calculated in real time. The signal-to-noise ratio is obtained through the signal-to-noise power ratio, the error vector magnitude is obtained through the deviation between the actual constellation points and the ideal constellation points, and the bit error rate is obtained through the ratio of the number of demodulated error bits to the total number of bits. These three metrics are weighted to obtain the channel state metric. According to the change trend of the channel state metric and the magnitude of the error signal, the size of the adaptive step size factor is dynamically adjusted. At the same time, according to the channel state metric and the value of the comprehensive loss function, the fusion factor used for updating the equalizer coefficients is updated.

[0156] Exemplarily, the sliding time window length is set to 64 sampling points to extract features from the input 16QAM modulated signal. In the time domain features, the signal amplitude range is [0, 1], the phase change range is [-π, π], and the instantaneous power range is [0, 1]. A 2048-point FFT is performed to obtain the frequency domain signal, and the spectral amplitude, dispersion features, and non-linear features are extracted to form a feature vector with a dimension of 6.

[0157] The first neural network adopts a three-layer structure, with the number of hidden layer neurons being 32 and 16 respectively. The ReLU activation function is used to reduce the 6D feature vector to an 8D feature. The long short-term memory network sets 16 memory cells to establish a temporal dependence relationship. The fully connected network maps the 16D feature to an 11D output, corresponding to an 11-tap equalizer.

[0158] The initial coefficients of the equalizer are set to [0.1, 0.2, 0.3, 0.4, 0.5, 1, 0.5, 0.4, 0.3, 0.2, 0.1]. The input signal passes through 10 delay units to obtain 11 delayed signals. The initial adaptive step size factor is set to 0.001, and the fusion factor is set to 0.5.

[0159] The weights of the three loss terms in the comprehensive loss function are 0.5, 0.3, and 0.2 respectively. The Adam optimizer is used for parameter update, and the learning rate is set to 0.001. When the channel state index is less than 0.8, the step size factor is increased; when the value of the loss function is greater than the threshold, the fusion factor is decreased to achieve adaptive adjustment.

[0160] In this embodiment, through the fusion extraction of time-frequency domain features, the ability to characterize signal features is improved. Through the processing of deep neural networks, the ability to compensate for nonlinear distortion is enhanced. Through multi-objective optimization design, the overall equalization performance is improved. Through parameter adaptive adjustment, the adaptability of the algorithm to channel changes is improved.

[0161] The adaptive equalization algorithms in the prior art often only focus on the features of a single domain, either time domain features or frequency domain features, and it is difficult to comprehensively reflect signal characteristics. Conventional equalization algorithms adopt fixed update strategies and lack the adaptive ability to channel states. Traditional methods are too single in the design of optimization objectives and usually only consider the mean square error, which cannot meet the multi-dimensional optimization requirements in complex communication environments. In this embodiment, time domain features and frequency domain features are fused. Through a sliding time window, dynamic extraction of time domain features such as signal amplitude, phase change, and instantaneous power is achieved. At the same time, the fast Fourier transform is used to obtain spectral amplitude, channel dispersion characteristics, and nonlinear effect characteristics, constructing a complete feature vector. In summary, this embodiment not only improves the equalization accuracy but also enhances the robustness of the algorithm, effectively solves various problems existing in traditional methods, realizes the overall improvement of the performance of the equalization algorithm, and provides an innovative solution for the performance optimization of communication systems.

[0162] In an alternative embodiment,

[0163] Transmitting the demodulated communication signal through the inter-satellite optical communication link and real-time monitoring of the link state parameters, and updating the training data set of the deep learning neural network according to the link state parameters to realize the online optimization of the polarization state prediction model includes:

[0164] Transmit the demodulated communication signal through the inter-satellite laser communication link, and use a polarization state analyzer to monitor the link state parameters of the inter-satellite laser communication link in real time. The link state parameters include signal-to-noise ratio, bit error rate, and polarization extinction ratio;

[0165] Form training sample pairs by combining the link state parameters with the corresponding polarization state parameters, and update the training data set of the deep learning neural network according to the training sample pairs. The deep learning neural network is used to establish a polarization state prediction model;

[0166] Train the deep learning neural network using the training data set, adjust the network learning rate according to the signal-to-noise ratio in the link state parameters, adjust the weight coefficient of the network optimization objective function according to the bit error rate in the link state parameters, and dynamically adjust the network structure parameters according to the polarization extinction ratio in the link state parameters to achieve online optimization of the polarization state prediction model.

[0167] Send the demodulated communication signal through the inter-satellite laser communication link. During the transmission process, the polarization state analyzer continuously collects the link state parameters. The signal-to-noise ratio is obtained by calculating the ratio of the received signal power to the noise power, the bit error rate is obtained by statistically calculating the ratio of the number of error bits to the total number of bits, and the polarization extinction ratio is obtained by measuring the power ratio between the orthogonal polarization states. These three parameters together reflect the transmission quality of the link.

[0168] For each sampling moment, record the current polarization state parameters, including characteristic parameters such as the ellipticity and azimuth angle of the polarization state. Pair these polarization state parameters with the link state parameters measured at the corresponding moment to form training sample pairs. Each training sample pair contains input features (link state parameters) and output labels (polarization state parameters). Continuously add the newly collected training sample pairs to the training data set to achieve dynamic update of the data set.

[0169] Construct a deep learning neural network with a multi-layer structure, including an input layer, multiple hidden layers, and an output layer. The input layer receives the link state parameters, and after non-linear transformation by the hidden layers, finally generates the predicted polarization state parameters at the output layer. Each hidden layer contains multiple neurons, and the neurons are connected by a weight matrix and equipped with an activation function to achieve non-linear mapping.

[0170] Adjust the network learning rate according to the change of signal-to-noise ratio: when the signal-to-noise ratio is relatively high, it indicates that the channel condition is good, and a larger learning rate can be used to accelerate convergence; when the signal-to-noise ratio decreases, reduce the learning rate to improve the training stability. Secondly, adjust the weight coefficients of each item in the optimization objective function according to the change of bit error rate: when the bit error rate increases, increase the weight of the corresponding loss item to strengthen the learning of samples with larger errors. Finally, dynamically adjust the network structure parameters according to the change of polarization extinction ratio: when the polarization extinction ratio decreases, appropriately increase the number of network layers or neurons to improve the expression ability of the network.

[0171] Exemplarily, assume that during a certain inter-satellite laser communication process, the demodulated 16QAM modulated signal is transmitted through the link. The polarization state analyzer samples once every millisecond, and measures link state parameters such as signal-to-noise ratio, bit error rate, and polarization extinction ratio. At the same time, record the polarization state parameters at the corresponding moment, including ellipticity and azimuth angle.

[0172] The initial configuration of the deep neural network is: a four-layer structure, with 3 nodes in the input layer (corresponding to three link state parameters), 64 and 32 nodes in the two hidden layers respectively, and 2 nodes in the output layer (corresponding to polarization state parameters). The initial learning rate is set to 0.001, the optimization objective function includes a mean square error term and a regularization term, and the initial weight ratio is 1:0.1.

[0173] When it is monitored that the signal-to-noise ratio drops from 20 dB to 15 dB, adjust the learning rate to 0.0005; when the bit error rate rises from 1e-6 to 1e-4, adjust the error term weight to 2:0.1; when the polarization extinction ratio drops from 30 dB to 20 dB, add 16 neurons to the second hidden layer. Retrain after each adjustment until the network performance meets the requirements.

[0174] Use the Adam optimizer to update the parameters, and use 128 samples for batch training each time. Evaluate the model performance through cross-validation. When the prediction error on the validation set is less than the threshold, complete one round of training. Continuously perform online optimization to enable the model to adapt to the dynamic changes of the link state.

[0175] In this embodiment, through the real-time monitoring of the link state and the dynamic update of the dataset, the response speed of the prediction model to link changes is improved. Through the adaptive adjustment of multi-dimensional parameters, the environmental adaptability of the model is enhanced. Through the online optimization mechanism, the continuous improvement of the prediction performance is achieved, realizing the transformation from a static model to a dynamic model, the improvement from single-parameter to multi-dimensional parameters, and the leap from offline training to online optimization. This not only improves the prediction accuracy but also enhances the adaptability of the algorithm, effectively solving various problems existing in traditional methods.

[0176] Figure 2This is the compensation error convergence curve graph of the inter-satellite laser communication and signal scheduling method based on adaptive polarization modulation in the embodiments of the present invention. As Figure 2 shown, in the first 100 iterations, the proposed technical solution (circular markers) shows the fastest convergence rate, with the error rapidly decreasing from 0.1 to 0.035; the adaptive filtering algorithm (square markers) converges more slowly and requires 300 iterations to reach an error level of 0.045; the traditional LMS algorithm (triangular markers) performs moderately and reaches an error level of 0.04 after 250 iterations. Figure 2 This demonstrates the dual advantages of the proposed technical solution in terms of convergence rate and final accuracy, with the final error stabilizing at 0.02.

[0177] Figure 3 This is the system convergence performance comparison graph of the inter-satellite laser communication and signal scheduling method based on adaptive polarization modulation in the embodiments of the present invention, showing the performance comparison of the proposed technical solution with the MMSE algorithm and the CMA algorithm during the convergence process of the equalizer coefficients. As Figure 3 shown, the proposed technical solution shows a faster convergence rate and lower steady-state error, reflecting the advantages of the feature extraction method based on deep learning in adaptive equalization. Although the MMSE algorithm and the CMA algorithm can also achieve system convergence, their convergence rates are slower and the final steady-state errors are larger, indicating that the performance of traditional algorithms is limited in complex channel environments.

[0178] The present invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having thereon computer-readable program instructions for performing various aspects of the present invention.

[0179] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended 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 for some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An inter-satellite laser communication and signal scheduling method based on adaptive polarization modulation, characterized in that Including: Obtain the polarization state information of the transmitting beam of the first inter-satellite laser communication terminal and the polarization state information of the receiving beam of the second inter-satellite laser communication terminal. Calculate the polarization state mismatch degree according to the beam polarization state information and the receiving beam polarization state information, and generate an initial polarization state compensation parameter based on the polarization state mismatch degree; Construct a polarization state prediction model based on a deep learning neural network. The input of the polarization state prediction model includes atmospheric disturbance data, satellite attitude data, and historical polarization state data, and the output includes the predicted polarization state drift trend. Pre-correct the polarization state compensation parameter according to the polarization state drift trend; Use a polarization modulator to perform real-time polarization state modulation on the transmitting beam. Adjust the modulation parameter of the polarization modulator according to the pre-corrected polarization state compensation parameter. Establish a link quality evaluation model for the inter-satellite laser communication link based on the modulation parameter of the polarization modulator, and calculate the link quality parameter; According to the link quality parameter, based on the adaptive optical carrier recovery technology of coherent detection, use an optical frequency comb as the local oscillator light source, achieve carrier synchronization through a digital phase-locked loop, and use an adaptive equalization algorithm to compensate for channel dispersion and nonlinear effects. Dynamically adjust the equalizer parameter according to the channel state information; Transmit the demodulated communication signal through the inter-satellite laser communication link, and monitor the link state parameter in real time. Update the training data set of the deep learning neural network according to the link state parameter to realize the online optimization of the polarization state prediction model.

2. The method according to claim 1, wherein Calculating the polarization state mismatch degree according to the beam polarization state information and the receiving beam polarization state information, and generating an initial polarization state compensation parameter based on the polarization state mismatch degree includes: Obtain the polarization state information of the transmitting beam and the polarization state information of the receiving beam. The polarization state information includes the total light intensity, the difference between the horizontal and vertical polarization components, the difference between the positive and negative forty-five degree polarization components, and the difference between the right-handed and left-handed circular polarization components; Construct a fourth-order Mueller matrix, which is composed of a phase delay matrix, a rotation matrix, and a depolarization matrix. Perform eigenvalue decomposition on the fourth-order Mueller matrix to obtain the contribution components of each matrix. Calculate the polarization state mismatch degree based on the contribution components. The polarization state mismatch degree is determined by the ratio of the square root of each contribution component to the sum of the square roots of all contribution components; Construct a compensation parameter optimization objective function based on the polarization state mismatch degree. The compensation parameter includes a phase compensation amount and a rotation angle compensation amount. Use the gradient iteration method to solve the optimization objective function to obtain the initial polarization state compensation parameter. Compensate and modulate the polarization state of the transmitting beam through a polarization modulator according to the initial polarization state compensation parameter.

3. The method according to claim 1, characterized in that, Constructing a polarization state prediction model based on a deep learning neural network. The input of the polarization state prediction model includes atmospheric disturbance data, satellite attitude data, and historical polarization state data, and the output includes the predicted polarization state drift trend. Pre-correcting the polarization state compensation parameter according to the polarization state drift trend includes: Collect the atmospheric temperature distribution, atmospheric pressure distribution, humidity distribution, and wind speed distribution to form atmospheric disturbance data, collect the roll angle, pitch angle, yaw angle, and angular velocity vector to form satellite attitude data, and obtain the Stokes parameter vectors at multiple sampling moments to form a historical polarization state sequence, where the sampling interval of the historical polarization state sequence is a preset time interval; Input the atmospheric disturbance data, the satellite attitude data, the historical polarization state sequence, and the initial polarization state compensation parameters into a recurrent neural network with a hybrid attention mechanism. The recurrent neural network calculates the hidden layer state at the current moment based on long short-term memory units, generates attention weights for the atmospheric disturbance data and the satellite attitude data according to the hidden layer state, and performs weighted fusion of the attention weights with the atmospheric disturbance data and the satellite attitude data respectively to obtain a first feature vector and a second feature vector. Feature fusion data is obtained by performing feature fusion on the first feature vector, the second feature vector, and the hidden layer state; Input the feature fusion data into a pre-trained neural network mapping function to predict the Stokes parameters after the next preset time interval. Calculate the difference between the Stokes parameters and the Stokes parameters at the latest moment in the historical polarization state sequence to obtain a polarization state difference. Divide the polarization state difference by the preset time interval to obtain a polarization state drift velocity vector. Construct a pre-compensation objective function containing a second-order regularization term of the pre-compensation parameter based on the polarization state drift velocity vector. Use the gradient descent method and combine it with a deep reinforcement learning framework to iteratively optimize the pre-compensation objective function to obtain a pre-compensation parameter, and use the pre-compensation parameter to perform pre-compensation on the Stokes parameters at the latest moment in the historical polarization state sequence.

4. The method according to claim 3, wherein Constructing a pre-compensation objective function containing a second-order regularization term of the pre-compensation parameter and using the gradient descent method and combining it with a deep reinforcement learning framework to iteratively optimize the pre-compensation objective function to obtain a pre-compensation parameter includes: Construct a system state vector including the current polarization state, polarization state drift velocity vector, compensation error, and compensation parameter, and construct an action vector from the rotation angle parameter adjustment amount and the phase parameter adjustment amount; Calculate the square term of the difference between the product of the compensation matrix and the predicted polarization state and the target polarization state based on the system state vector, and construct an immediate reward function in combination with the square terms of the rotation angle parameter adjustment amount and the phase parameter adjustment amount. Use a discount factor to accumulate the immediate rewards within the future prediction step count to obtain a pre-compensation objective function; Construct a deep neural network as a value network and a policy network based on the pre-compensation objective function. The value network inputs the system state vector and the action vector and outputs a value evaluation. The policy network inputs the system state vector and outputs an action policy, and stores the state transition samples in an experience replay buffer; Construct a temporal difference objective function based on the immediate reward function and the value network, construct a policy gradient objective function based on the value network and the policy network, and introduce a policy distribution entropy to construct an entropy regularization objective function; The gradient descent method is used to optimize and update the network parameters of the value network for the temporal difference objective function, and the weighted sum of the policy gradient objective function and the entropy regularization objective function is optimized to update the network parameters of the policy network. The target network parameters are updated in a soft update manner; Gaussian noise is injected into the action policy output by the policy network to obtain the actual executed action. The variance of the Gaussian noise is calculated based on the compensation error, and the actual executed action is used as the update amount of the compensation parameter to iteratively optimize the compensation parameter to obtain the pre-compensation parameter.

5. The method according to claim 1, characterized in that, A polarization modulator is used to perform real-time polarization state modulation on the transmitted beam. The modulation parameters of the polarization modulator are adjusted according to the pre-calibrated polarization state compensation parameters. Based on the modulation parameters of the polarization modulator, a model for evaluating the quality of the inter-satellite laser communication link is established, and the link quality parameters calculated include: A polarization modulator is used to perform real-time polarization state modulation on the transmitted beam. The modulation parameters of the polarization modulator are adjusted according to the pre-calibrated polarization state compensation parameters. The modulation parameters include the polarization state rotation angle and the polarization state ellipticity; Based on the modulation parameters of the polarization modulator, a model for evaluating the quality of the inter-satellite laser communication link is established. The link quality evaluation model includes the influence factor of atmospheric disturbance on the polarization state, the polarization state analysis factor at the receiving end, and the system polarization extinction ratio factor; The link quality parameters are calculated according to the link quality evaluation model, and the polarization state compensation parameters are updated in real time according to the link quality parameters.

6. The method according to claim 1, wherein According to the link quality parameters, based on the adaptive optical carrier recovery technology of coherent detection, an optical frequency comb is used as the local oscillator light source, carrier synchronization is achieved through a digital phase-locked loop, and an adaptive equalization algorithm is used to compensate for channel dispersion and nonlinear effects. The equalizer parameters are dynamically adjusted according to the channel state information, including: The link quality parameters are obtained. An optical frequency comb is used as the local oscillator light source. The frequency comb spectrum lines are generated according to the preset center frequency and comb tooth interval. The frequency comb spectrum lines are determined by the amplitude, frequency, and initial phase of the frequency components. The temperature drift amount and the driving voltage fluctuation amount are monitored. The frequency drift amount is calculated according to the product of the temperature sensitivity coefficient and the temperature drift amount and the product of the voltage sensitivity coefficient and the driving voltage fluctuation amount, and the frequency of the optical frequency comb is compensated in real time; The received optical signal and the optical frequency comb are coherently detected to obtain a beat signal. The beat signal is multiplied by the complex conjugate of the local reference signal to obtain a phase error. The phase error is input into a digital phase-locked loop, and a proportional-integral controller is used to filter the phase error. The control voltage is calculated according to the weighted sum of the product of the phase error at the current moment and the proportional coefficient and the cumulative sum of the historical phase errors and the integral coefficient; The output phase of the numerically controlled oscillator is calculated according to the control voltage. The output phase is obtained by adding the output phase at the previous moment, the product of the center frequency and the sampling period, and the product of the control voltage and the gain coefficient. The output phase is used as the phase parameter of the local reference signal to achieve carrier synchronization; Establish a channel transmission model, where the channel transmission model includes a dispersion compensation model and a non - linear effect compensation model. The dispersion compensation model calculates the dispersion compensation coefficient according to the optical signal wavelength, transmission distance, and the speed of light. The non - linear effect compensation model calculates the non - linear compensation coefficient according to the input signal, non - linear coefficient, signal power, and effective length. Through deep learning technology, feature extraction and optimization of the loss function design are carried out. Combining with the adaptive equalization algorithm, the channel dispersion and the non - linear effect are compensated. Multiple delayed signals of the input signal are multiplied by the equalizer coefficients respectively and summed to obtain the output signal of the adaptive equalization algorithm; Calculate the adaptive step - size factor according to the link quality parameter, take the product of the adaptive step - size factor, the error signal, and the conjugate signal of the input signal delayed signal as the update amount to update the equalizer coefficients, and dynamically adjust the size of the adaptive step - size factor based on the link quality parameter to achieve the adaptive optimization of the equalizer coefficients and output the demodulated communication signal.

7. The method according to claim 6, wherein Through deep learning technology, feature extraction and optimization of the loss function design are carried out. Combining with the adaptive equalization algorithm to compensate the channel dispersion and the non - linear effect, and multiplying and summing multiple delayed signals of the input signal with the equalizer coefficients respectively to obtain the output signal of the adaptive equalization algorithm includes: Use a sliding time window with a preset length to perform time - domain feature extraction on the input signal to obtain signal amplitude, signal phase change, and signal instantaneous power. Perform a fast Fourier transform on the signal data within the sliding time window to obtain a frequency - domain signal, extract the spectral amplitude, channel dispersion characteristics, and non - linear effect characteristics of the frequency - domain signal, and form a feature vector with the signal amplitude, the signal phase change, the signal instantaneous power, the spectral amplitude, the channel dispersion characteristics, and the non - linear effect characteristics; Input the feature vector into the first neural network for feature dimensionality reduction to obtain the first feature. The first neural network includes a weight matrix and a bias vector. Input the first feature into the long - short - term memory network to obtain hidden - state features. The long - short - term memory network is used to model temporal features. Input the hidden - state features into the fully - connected neural network to obtain the equalizer coefficients; Compensate the channel dispersion and the non - linear effect, multiply and sum multiple delayed signals of the input signal with the equalizer coefficients respectively to obtain the output signal of the adaptive equalization algorithm. Subtract the output signal of the adaptive equalization algorithm from the desired output signal to obtain an error signal, and update the equalizer coefficients according to the error signal, the adaptive step - size factor, the conjugate signal of the input signal delayed signal, and the fusion factor; Construct a comprehensive loss function including a mean - square error loss term, a bit - error rate loss term, and a peak - to - average power ratio loss term, and use the gradient descent method to optimize the network parameters of the first neural network, the network parameters of the long - short - term memory network, and the network parameters of the fully - connected neural network according to the comprehensive loss function; Calculate the channel state index according to the signal-to-noise ratio, error vector magnitude, and bit error rate, dynamically adjust the adaptive step factor according to the channel state index and the error signal, and update the fusion factor according to the channel state index and the comprehensive loss function.

8. The method according to claim 1, characterized in that, Transmit the demodulated communication signal through the inter-satellite laser communication link, and monitor the link state parameters in real time. Updating the training data set of the deep learning neural network according to the link state parameters to realize the online optimization of the polarization state prediction model includes: Transmit the demodulated communication signal through the inter-satellite laser communication link, and use a polarization state analyzer to monitor the link state parameters of the inter-satellite laser communication link in real time. The link state parameters include signal-to-noise ratio, bit error rate, and polarization extinction ratio; Form a training sample pair with the link state parameter and the corresponding polarization state parameter, and update the training data set of the deep learning neural network according to the training sample pair. The deep learning neural network is used to establish a polarization state prediction model; Train the deep learning neural network with the training data set, adjust the network learning rate according to the signal-to-noise ratio in the link state parameter, adjust the weight coefficient of the network optimization objective function according to the bit error rate in the link state parameter, and dynamically adjust the network structure parameter according to the polarization extinction ratio in the link state parameter to realize the online optimization of the polarization state prediction model.

Images