High-altitude platform continuous variable quantum key distribution channel intelligent compensation method

Abstract

The invention provides an intelligent compensation method for a continuous variable quantum key distribution channel of a high-altitude platform, and relates to the technical field of quantum communication, and the method comprises the steps: obtaining a current light intensity distribution diagram of a to-be-compensated channel link, inputting the current light intensity distribution diagram into a constructed atmospheric turbulence phase prediction model, predicting phase distortion caused by turbulence in the to-be-compensated channel link to obtain a phase prediction result; and compensating the to-be-compensated channel link by using the phase prediction result. The network architecture of the atmospheric turbulence phase prediction model comprises an input layer, a feature extraction module, a small sample learning module, a dynamic optimization module and an output layer which are connected in sequence. According to the invention, the problem of poor correction precision in the traditional optical correction technology is effectively solved, and the efficiency and stability of the high-altitude platform CV-QKD channel are improved.

Description

Technical Field

[0001] The present invention relates to the field of quantum communication technology, and in particular to a high-altitude platform continuous variable quantum key distribution channel intelligent compensation method based on deep learning. Background Art

[0002] With the development of aviation technology, high-altitude platforms such as drones, hot air balloons, and airships have emerged. These platforms typically operate within the atmosphere at an altitude of 20 km. They are not only rapidly deployable and easy to maintain, but can also achieve long-duration fixed-point hovering or regional patrols. Space laser communication systems, as a new generation of space information transmission technology, are rapidly developing and gaining widespread application. Using high-altitude platforms such as drones as carriers and establishing laser communication links between them and the ground, they offer a new solution for high-speed, interference-resistant regional communications. Quantum key distribution (QKD), a core technology in quantum communication, has made significant progress in recent years. Because the signal modulation and measurement methods of continuous variable quantum key distribution (CV-QKD) are highly compatible with existing space laser communication systems, CV-QKD is often used to transmit quantum keys over laser communication links, thus providing "absolute" security for information transmission within laser communication systems, as defined by quantum theory.

[0003] In fiber-based CV-QKD systems, the channel transmission medium is a solid-state waveguide structure, whose attenuation coefficient and noise characteristics can be quantitatively modeled through laboratory calibration. In contrast, high-altitude platform CV-QKD faces the significant challenge of dynamic atmospheric channels. Specifically, atmospheric turbulence induces random fluctuations in the refractive index, leading to physical effects such as wavefront distortion, beam drift, and intensity fluctuations in the signal light. These effects cause spatial mode mismatch in the signal beam, reducing coupling efficiency and degrading the signal-to-noise ratio. Furthermore, the enhanced signal-to-noise ratio leads to increased parameter estimation errors, and intensity fluctuations also cause modulation variance drift. These effects can reduce the key rate and, especially when the channel transmittance falls below a security threshold, can even cause transmission interruption. Therefore, to compensate for noise in high-altitude platform CV-QKD channels, conventional optics, a well-established turbulence compensation solution in classical optical communications, is often used. This approach primarily relies on real-time detection of wavefront phase distortion using a beacon beam and the use of deformable mirrors to achieve wavefront correction accuracy on the order of λ / 10.

[0004] However, in the high-altitude platform CV-QKD scenario, the response bandwidth of the deformable mirror is usually <1kHz, making it difficult to track high-frequency turbulent disturbances; at the same time, the multi-stage correction devices in the deformable mirror will bring additional insertion loss, which is inconsistent with the sensitivity of high-altitude platform CV-QKD to channel efficiency; in addition, high-altitude platform CV-QKD has strict constraints on payload mass and power consumption.

[0005] Therefore, there is an urgent need for an intelligent compensation method for continuous variable quantum key distribution channels on high-altitude platforms to solve the problem of poor correction accuracy in traditional optical correction technology due to the difficulty in tracking high-frequency turbulence disturbances and insertion loss. Summary of the Invention

[0006] In response to the above-mentioned problems, the present invention provides an intelligent compensation method for high-altitude platform continuous variable quantum key distribution channels, which is used to solve the problem of poor correction accuracy in traditional optical correction technology, thereby improving the efficiency and stability of high-altitude platform CV-QKD channels.

[0007] The present invention provides a high-altitude platform continuous variable quantum key distribution channel intelligent compensation method, the method comprising: Obtaining a current light intensity distribution map of the channel link to be compensated, inputting the map into the constructed atmospheric turbulence phase prediction model, predicting the phase distortion caused by turbulence in the channel link to be compensated, and obtaining a phase prediction result; Compensating the channel link to be compensated by using the phase prediction result; The network architecture of the atmospheric turbulence phase prediction model includes the following components connected in sequence: The input layer is used to obtain the link’s sample data set or current light intensity distribution map; A feature extraction module, configured to extract turbulence dynamic change features from the sample data set or the current light intensity distribution map using a hierarchical and progressive feature learning network; A small sample learning module is used to train and learn the dynamic change characteristics of the turbulence; Dynamic optimization module, used to adjust and optimize the atmospheric turbulence phase prediction model; The output layer is used to output the phase prediction result.

[0008] Furthermore, the feature extraction module includes: The pixel perception layer is used to obtain the sample data set and adopt the local window attention mechanism of fixed pixel units to capture the pixel-level phase change characteristics of the phase screen from the sample data set to obtain the first-level turbulence dynamic change characteristics; The local correlation layer is used to obtain the first-order turbulence dynamic change characteristics. It constructs a dynamic receptive field through window merging, establishes a spatial correlation model between adjacent turbulence local characteristics, captures the structural changes of turbulence at the mesoscale, and obtains the second-order turbulence dynamic change characteristics. The global modeling layer is used to obtain the dynamic change characteristics of the second-order turbulence, analyze the structural changes of turbulence at the mesoscale and the statistical distribution of turbulence structure, and obtain the dynamic change characteristics of the third-order turbulence; The cross-scale feature interaction module is used to fuse the turbulence dynamic change characteristics of different levels to obtain the turbulence dynamic change characteristics.

[0009] Furthermore, the small sample learning module includes: The feature reuse channel is used to receive turbulence dynamic change features, build a turbulence feature prototype library including feature vectors corresponding to turbulence of different intensities, and construct an attention fusion gate that coordinates channels and space. Based on the turbulence feature prototype library and attention fusion gate, the key feature maps in the turbulence dynamic change features are identified and the attention heat map is generated. The knowledge distillation channel is used to obtain attention heat maps, use the teacher network trained with big data to supervise and guide the student network trained with small samples, and distill and extract key features.

[0010] Furthermore, the dynamic optimization module includes one or more optimization methods including turbulence perception enhancement strategy, dynamic network structure adjustment, and multi-task loss function.

[0011] Furthermore, the method for constructing the atmospheric turbulence phase prediction model includes: inputting the generated sample data set into the neural network model for pre-training to obtain a pre-trained phase result; using physical consistency conditions to train the pre-trained phase result to obtain the atmospheric turbulence phase prediction model.

[0012] Furthermore, the method for generating the sample data set includes: Build a dynamic communication simulation scenario between the high-altitude platform and the ground station to generate a link data set; The link dataset is characterized by using an atmospheric turbulence phase screen characterization model to obtain a superimposed phase screen caused by turbulence in the corresponding link, and a label dataset is generated; Inputting the label data set into an ideal free-space transmission model of a Gaussian beam to obtain an input light intensity distribution diagram of a Gaussian beam cross section after transmission via a corresponding link, thereby generating a training data set; The sample dataset is obtained according to the mapping relationship between the label dataset and the training dataset.

[0013] Furthermore, the method for building a dynamic communication simulation scenario between the high-altitude platform and the ground station is: real-time collection of parameter information of the high-altitude platform and the ground station, input into the STK simulation platform, and retrieve or load the change process of the refractive index structure constant distribution of the link between the high-altitude platform and the ground station during the communication process to realize the dynamic communication simulation scenario of the continuous variable quantum key distribution channel of the high-altitude platform.

[0014] Furthermore, using the physical consistency condition to constrain the pre-training phase result includes: Substituting the pre-trained phase result into the light intensity propagation physical model to obtain a theoretical light intensity distribution diagram; An input light intensity distribution map corresponding to the pre-training phase result is obtained, a difference between the input light intensity distribution map and the theoretical light intensity distribution map is calculated, a self-supervised loss function is constructed, and training constraints are performed based on the self-supervised loss function.

[0015] Furthermore, the self-supervised loss function is: ; in, Indicates the total number of samples; represents the input light intensity distribution diagram; Represents the theoretical light intensity distribution diagram.

[0016] Furthermore, the pre-training loss function of the pre-training is: ; in, Indicates the total number of samples; represents the superimposed phase screen in the label dataset; Represents the pre-training phase result; , represents the linear rectification function, which is a commonly used activation function in neural networks.

[0017] In general, the present invention provides a method for intelligent compensation of a continuous variable quantum key distribution channel on a high-altitude platform. The technical solution conceived by the present invention can achieve the following beneficial effects compared with the existing technology: (1) The present invention captures the dynamic change characteristics of turbulence in a large number of different links by introducing deep learning, learns and analyzes the complex mapping relationship between beam propagation and phase results in turbulence, and establishes an atmospheric turbulence phase prediction model based on a hierarchical and progressive feature learning network, small sample learning training and dynamic optimization module. It can accurately predict the beam wavefront phase distortion caused by turbulence in the communication link and perform real-time channel compensation in a short time, thereby meeting the real-time requirements of the channel compensation method for the rapid movement of satellites in the high-altitude platform CV-QKD channel scenario. Compared with the traditional adaptive optics method, the atmospheric turbulence phase prediction model established by the present invention can not only solve the problems of high-frequency turbulence disturbances being difficult to track and insertion loss in traditional optical correction technology, and improve the correction accuracy, but also give full play to the efficiency and adaptability of deep learning algorithms under the condition of limited resources of high-altitude platforms, effectively improve the efficiency and stability of high-altitude platform CV-QKD channels, provide strong technical support for the practical application of high-altitude platform CV-QKD, and provide a new idea and method for solving the noise compensation problem.

[0018] (2) The present invention uses a hierarchical and progressive method to capture the multi-scale dynamic characteristics of the complex dynamic changes of turbulence in the channel link; as the network level goes deeper, the subsequent network layer can extract more detailed and local features based on the basic features extracted by the previous network layer, and comprehensively capture the dynamic change characteristics of turbulence at different scales. Through the synergistic effect between various levels, the comprehensive extraction of the complex dynamic change characteristics of turbulence from macro to micro is achieved, avoiding the omission of important features, laying the foundation for subsequent small sample learning, and improving the accuracy of atmospheric turbulence phase prediction model training.

[0019] (3) The present invention adopts small sample learning and training. During the small sample learning and training, the student network uses the teacher network for supervision and guidance, which can better understand these key features. In addition, when the training data is limited, the small sample data can be fully utilized, so that the student network can mine more effective features on the limited data set and learn richer feature information, thereby improving its ability to identify the dynamic change characteristics of turbulence on the small sample data set and improving the accuracy of the atmospheric turbulence phase prediction model training.

[0020] (4) The present invention adopts a physics-driven self-supervised neural network architecture to reduce dependence on manually labeled data, use physical consistency conditions for training constraints, improve the rationality of network physics, enhance the adaptability to complex turbulence phase prediction models, and improve the generalization ability of the neural network model.

[0021] (5) The present invention uses the STK simulation platform to construct a dynamic communication simulation scenario. Compared with the static geometric model commonly used in the prior art, the present invention can describe the dynamic change process of the link between the high-altitude platform and the ground station during the communication process, providing a more diversified and accurate link data set, making the training of the atmospheric turbulence phase prediction model more accurate, improving the accuracy of the phase prediction results of the high-altitude platform CV-QKD channel, and making the channel compensation more accurate.

[0022] (6) The present invention not only has low hardware requirements and controllable power consumption, but also solves the problems of complex turbulence correction technology system and high hardware requirements of traditional optical technology; it can also be applied to any high-altitude platform with strict requirements on payload and power consumption, and can even be further extended to satellite platforms for application. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0024] Figure 1 This is a schematic diagram of the method steps of a high-altitude platform continuous variable quantum key distribution channel intelligent compensation method provided by the present invention; Figure 2 This is a schematic diagram of the network architecture of an atmospheric turbulence phase prediction model for an intelligent compensation method for a continuous variable quantum key distribution channel on a high-altitude platform provided by the present invention. DETAILED DESCRIPTION

[0025] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings and embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0026] It should be noted that, in the description of the embodiments of the present invention, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a method, step, or apparatus comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such method, step, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the method, step, or apparatus comprising the element.

[0027] In order to solve the problem of poor correction accuracy caused by the difficulty in tracking high-frequency turbulence disturbances and insertion loss in traditional optical correction technology, the present invention proposes a high-altitude platform continuous variable quantum key distribution channel intelligent compensation method based on deep learning. The atmospheric turbulence phase prediction model constructed by deep learning is used to predict the atmospheric turbulence phase change, and the phase prediction result is used to compensate for the light beam wavefront phase distortion caused by turbulence in the high-altitude platform channel, thereby realizing the correction of the high-altitude platform CV-QKD channel link and improving the communication performance.

[0028] Specifically, if Figure 1 As shown, the method includes: S101: Obtain a current light intensity distribution map of the channel link to be compensated, input the map into the constructed atmospheric turbulence phase prediction model, predict the phase distortion caused by turbulence in the channel link to be compensated, and obtain a phase prediction result; S102: Compensate the channel link to be compensated using the phase prediction result.

[0029] Current intelligent model architectures like SegNet and UNet, which rely on convolutional operations, are effective at capturing local features, but are limited in extracting global features. This makes these models only applicable to turbulence intensity within a narrow range of variation. In actual high-altitude platform communication scenarios, turbulence distributions change rapidly with the phase position between the satellite and the ground station. These models, with their limited generalization capabilities, struggle to capture the dependencies between features when dealing with complex turbulence variations. Furthermore, they incur high computational complexity and training time costs.

[0030] Therefore, the present invention constructs an atmospheric turbulence phase prediction model using a hierarchical and progressive feature learning network and a small-sample learning module. Based on the Transformer architecture, this atmospheric turbulence phase prediction model employs a "multi-scale feature decoupling → small-sample learning and knowledge distillation enhancement → dynamic optimization training" architecture. This effectively addresses key issues of existing models, such as high training time costs and weak generalization capabilities, and provides a new technical approach for turbulence compensation in quantum communications.

[0031] Specifically, such as Figure 2 As shown in FIG, the network architecture of the atmospheric turbulence phase prediction model includes an input layer, a feature extraction module, a small sample learning module, a dynamic optimization module and an output layer connected in sequence.

[0032] The input layer is used to obtain the sample data set or current light intensity distribution map of the link.

[0033] The feature extraction module is used to extract the turbulence dynamic change characteristics in the sample data set or the current light intensity distribution map using a hierarchical and progressive feature learning network.

[0034] During the feature extraction phase, a hierarchical and progressive feature learning network is used, leveraging the synergy between layers to comprehensively extract the complex dynamic characteristics of turbulence, from the macroscopic to the microscopic level. As an example, the feature extraction module includes a pixel perception layer, a local correlation layer, and a global modeling layer, all connected in sequence.

[0035] The pixel perception layer is used to obtain sample data sets. It adopts a local window attention mechanism with fixed pixel units to capture the pixel-level phase change characteristics of the phase screen from the sample data sets and obtain the first-order turbulence dynamic change characteristics, thus providing a basis for the detailed characterization of turbulence.

[0036] The local correlation layer is used to obtain the first-order turbulence dynamic change characteristics, construct a dynamic receptive field through window merging, establish a spatial correlation model between adjacent turbulence local features, capture the structural changes of turbulence at the mesoscale, and obtain the second-order turbulence dynamic change characteristics.

[0037] The global modeling layer is used to obtain the dynamic change characteristics of the second-order turbulence, deploy the global sparse attention mechanism, analyze the structural changes of turbulence at the mesoscale and the statistical distribution law of turbulence structure, and obtain the dynamic change characteristics of the third-order turbulence.

[0038] The cross-scale feature interaction module is used to fuse the turbulence dynamic change characteristics of different levels to obtain the turbulence dynamic change characteristics.

[0039] The dynamic change characteristics of turbulence are a multi-scale feature with rich layers and interrelated features. This hierarchical progressive, multi-scale feature decoupling and fusion strategy breaks through the limitations of traditional convolutional neural networks that extract single-scale feature maps through fixed-size convolution kernels in the convolution layer, and realizes the explicit expression of the multi-scale characteristics of the dynamic change characteristics of turbulence, providing a solid foundation for subsequent model training and prediction.

[0040] Small sample learning module, used to train and learn the dynamic change characteristics of turbulence.

[0041] To further improve the model's performance under small sample conditions, the present invention employs a knowledge transfer algorithm framework based on turbulence characteristics to mine more effective features, learn richer feature information, and enhance its ability to identify the dynamic changes in turbulence characteristics in small sample datasets. As an example, the small sample learning module includes a feature reuse channel and a knowledge distillation channel.

[0042] The feature reuse channel is used to receive the dynamic change characteristics of turbulence, build a turbulence feature prototype library including feature vectors corresponding to turbulence of different intensities, and build an attention fusion gate that coordinates channels and space; based on the turbulence feature prototype library and attention fusion gate, the key feature maps in the dynamic change characteristics of turbulence are identified to generate an attention heat map.

[0043] On the one hand, a prototype library of turbulence features is constructed through comparative learning; the essential feature vectors corresponding to turbulence of different intensities are extracted. These feature vectors serve as the basis for knowledge distillation and can help the model better understand the core characteristics of turbulence.

[0044] On the other hand, a channel-space collaborative attention gate is constructed; in the channel dimension, key feature maps are automatically identified through learnable weight vectors, and the weights of key features are adaptively improved; in the spatial dimension, an attention heat map is generated according to the structural distribution of the dynamic change characteristics of turbulence, thereby achieving targeted enhancement of important feature areas from two dimensions.

[0045] The feature multiplexing channel is like equipping the feature extraction module with a feature filter, which can effectively improve feature utilization under limited data conditions.

[0046] The knowledge distillation channel is used to obtain attention heat maps, use the teacher network trained with big data to supervise and guide the student network trained with small samples, and distill and extract key features.

[0047] In the knowledge distillation channel, the teacher network extracts key features through the attention mechanism and passes them to the student network, enabling the student network to learn richer feature information under small sample data conditions and improve its ability to identify the dynamic change characteristics of turbulence on small sample data sets.

[0048] Furthermore, the present invention introduces a feature importance assessment module to dynamically adjust the intensity and direction of knowledge transfer within the knowledge distillation pipeline, ensuring a more efficient and targeted knowledge transfer process. This knowledge distillation enhancement mechanism significantly improves the atmospheric turbulence phase prediction model's ability to discern the essential characteristics of turbulence, resulting in excellent learning results even with small sample sizes.

[0049] Dynamic optimization module, used to adjust and optimize the atmospheric turbulence phase prediction model; As an embodiment, the dynamic optimization module includes one or more optimization methods including turbulence perception enhancement strategy, dynamic network structure adjustment, and multi-task loss function.

[0050] The turbulence perception enhancement strategy is to increase the sample data corresponding to turbulence of different intensities; the dynamic network structure adjustment is to adjust and optimize the structure of the atmospheric turbulence phase prediction model, such as levels and modules; the multi-task loss function is to adjust and optimize the parameters in the atmospheric turbulence phase prediction model based on the loss function.

[0051] The output layer is used to output the phase prediction results.

[0052] Based on the network architecture of the atmospheric turbulence phase prediction model, the model is trained. Since supervised learning neural networks require a large dataset of atmospheric turbulence samples in practical applications to achieve high-precision predictions of the turbulence patterns present in the training dataset, turbulence models are complex and diverse, and the training dataset's ability to fully capture their variations is limited. Furthermore, increasing the size of the training dataset also increases the model size and training costs.

[0053] Therefore, the present invention introduces a "data + physics" driven self-supervised neural network module to use physical consistency conditions to constrain the training results. This not only improves the rationality of the atmospheric turbulence phase prediction model and enhances the adaptability of complex turbulence models, but also reduces the dependence on manually labeled data.

[0054] As an embodiment, a method for constructing an atmospheric turbulence phase prediction model includes: inputting the generated sample data set into a neural network model for pre-training to obtain a pre-trained phase result; using physical consistency conditions to train constraints on the pre-trained phase result to obtain an atmospheric turbulence phase prediction model.

[0055] That is to say, the sample data set is input into the neural network model for pre-training, the corresponding function value during the training process is calculated according to the selected loss function, and the weight parameters in the model are continuously updated according to the change of the function value until the model loss function value converges stably, and a trained atmospheric turbulence phase prediction model is obtained.

[0056] It should be noted that the sample data set is a light intensity distribution map-superimposed phase screen with a mapping relationship; that is, one light intensity distribution map corresponds to one superimposed phase screen.

[0057] Furthermore, the method for generating the sample data set includes: S201: Build a dynamic communication simulation scenario between the high-altitude platform and the ground station to generate a link data set.

[0058] Existing technologies often use static geometric models to simulate communication links. However, on high-altitude platforms, channels change dynamically with atmospheric turbulence, and this static simulation cannot simulate the dynamic changes caused by atmospheric turbulence. Therefore, the present invention uses the STK simulation platform (Satellite Tool Kit) to simulate the dynamic changes of different links between the high-altitude platform and the ground station under different turbulence intensities.

[0059] As an embodiment, the method for setting up a dynamic communication simulation scenario between a high-altitude platform and a ground station is as follows: real-time collection of parameter information of the high-altitude platform and the ground station, input into the STK simulation platform, and retrieve or load the change process of the refractive index structure constant distribution of the link between the high-altitude platform and the ground station during the communication process, that is, the link turbulence distribution information, to realize the dynamic communication simulation scenario of the continuous variable quantum key distribution channel of the high-altitude platform.

[0060] Among them, the parameter information includes high-altitude platform parameters such as flight altitude, motion trajectory, and flight speed, ground station parameters such as latitude and longitude coordinates and altitude, and atmospheric weather parameters such as temperature, humidity, and wind direction published by NASA.

[0061] Based on the changes in the high-altitude platform parameters, ground station parameters, atmospheric weather parameters, the real-time distance between the high-altitude platform and the ground station, and the relative pitch angle data between the transmitter and receiver, multiple different link data sets are obtained; among them, one link data corresponds to one type of turbulence distribution information, and multiple link data correspond to multiple types of turbulence distribution information.

[0062] S202: Characterize the link dataset using an atmospheric turbulence phase screen characterization model to obtain a superimposed phase screen caused by turbulence in the corresponding link, and generate a label dataset.

[0063] The phase distortion caused by turbulence is represented by the atmospheric phase screen. The atmospheric turbulence phase screen characterization model can be obtained by filtering the atmospheric turbulence phase power spectrum density function through a complex Gaussian random matrix. It is a function of the atmospheric refractive index structure constant. Taking the Kolmogorov power spectrum as an example, the expression of the atmospheric turbulence phase screen characterization model can be: ; in, is the wave number, and The wave numbers are Components on the axis; is the distance between the atmospheric equivalent phase screens on the beam propagation path; Represents the atmospheric refractive index structure constant, which is used to characterize the strength of atmospheric turbulence. When the turbulence is weak, When the turbulence is moderate, It is a strong turbulence.

[0064] Therefore, the atmospheric phase screen can be expressed as: ; in, The variance is 1 dimensional complex Gaussian random matrix, The number of grid divisions is equal to that of the atmospheric phase screen; represents the fast Fourier transform.

[0065] As an embodiment, generating a label dataset includes: The different links in the link dataset are layered, and the atmospheric turbulence phase screen characterization model is used to characterize each layer of turbulence to obtain multiple phase screens. All the phase screens are superimposed and visualized to obtain a superimposed phase screen. The superimposed phase screen is used as label data to represent the total phase distortion caused by turbulence in the corresponding channel link.

[0066] One type of turbulence distribution information corresponds to one label data; multiple types of turbulence distribution information correspond to multiple label data, thereby obtaining a label data set.

[0067] S203: Input the label data set into the ideal free-space transmission model of the Gaussian beam to obtain an input light intensity distribution diagram of the Gaussian beam cross section after transmission via the corresponding link, thereby generating a training data set.

[0068] That is to say, the label data is substituted as a phase term into the ideal free-space transmission model of a Gaussian beam to obtain a beam propagation model affected by turbulent phase noise. Based on the beam propagation model, the light intensity distribution diagram of the Gaussian beam cross-section after the corresponding link transmission can be calculated and used as training data.

[0069] One label data corresponds to one light intensity distribution map; multiple label data correspond to multiple light intensity distribution maps, thereby obtaining a training data set.

[0070] S204: Obtain a sample dataset based on the mapping relationship between the label dataset and the training dataset.

[0071] Furthermore, data augmentation can be performed on the sample dataset. That is, since the images in the label dataset and the training dataset are all grayscale images with pixel grayscale values ​​in the range [0, 255], all images are normalized so that the pixels in the sample dataset are scaled to a set range. At the same time, the normalized sample dataset is divided into a training set and a test set according to a certain ratio.

[0072] The generated sample data set is input into the neural network model for pre-training to obtain the pre-training phase result; then the physical consistency condition is used to train the pre-training phase result.

[0073] First, initialize the model, including the initialization of model parameters, optimizer, and learning rate, and then determine the model training loss function. For example, you can set the maximum number of training iterations to 200, the batch size to 64, the optimizer to Adam, and the initial learning rate to 0.001. The pre-training loss function can be: ; in, Indicates the total number of samples; represents the superimposed phase screen in the label dataset; Represents the pre-training phase result; , represents the linear rectification function, which is a commonly used activation function in neural networks.

[0074] As an embodiment, using physical consistency conditions to constrain the pre-training phase results includes: Substitute the pre-trained phase result into the light intensity propagation physical model to obtain a theoretical light intensity distribution diagram; wherein the light intensity propagation physical model can be angular spectrum method, Fresnel diffraction propagation, or Fraunhofer diffraction propagation.

[0075] Obtain the input light intensity distribution map corresponding to the pre-training phase result in the sample data set, calculate the difference between the input light intensity distribution map and the theoretical light intensity distribution map, construct a self-supervised loss function, and perform training constraints based on the self-supervised loss function.

[0076] As an embodiment, the loss function of the physically driven self-supervised neural network module adopts mean square error. Furthermore, the self-supervised loss function is: ; in, Indicates the total number of samples; represents the input light intensity distribution diagram; Represents the theoretical light intensity distribution diagram.

[0077] By adding a self-supervised neural network module and using physical consistency conditions to constrain the training of pre-trained phase results, the generalization ability of the atmospheric turbulence phase prediction model is improved, and the robustness of continuous variable quantum key distribution on high-altitude platforms in complex turbulent environments is enhanced.

[0078] The neural network model is iteratively trained, with stable convergence of the loss function value as the sign of training completion. After training is completed, the final atmospheric turbulence phase prediction model is obtained. The change in the loss function value can be used as a basis for adjustment. For example, if the loss function value does not decrease after more than 10 iterations, the learning rate is reduced by 0.1. If the loss function value does not decrease after more than 50 iterations, the training is terminated early.

[0079] A phase prediction result is obtained through an atmospheric turbulence phase prediction model; and the phase prediction result is used to compensate the channel link to be compensated.

[0080] That is: deploy the trained atmospheric turbulence phase prediction model to the hardware computing platform embedded in the CV-QKD receiver, use the pilot light intensity distribution map recorded by the image sensor as the model input, make real-time predictions on the turbulence phase changes in the channel, output the phase prediction results, and compensate the phase into the quantum signal light to achieve compensation for the channel turbulence phase distortion.

[0081] Specifically, at the transmitter, an intensity modulator modulates the Gaussian beam generated by a continuous-wave laser into a pulse sequence, and a polarization beam splitter divides the optical path into two parts: quantum signal light and pilot light. The modulator transforms the quantum signal light into a continuous variable quantum signal with a Gaussian or discrete distribution in the position and momentum directions. The pilot light is delayed by the optical fiber, achieving time-division multiplexing with the signal light.

[0082] The pilot light and quantum signal light are then transmitted through a spatial channel. Within this spatial channel link, a spatial light modulator (SLM) generates an atmospheric turbulence-equivalent phase screen to modulate the transmitted signal, simulating the wavefront phase distortion experienced during atmospheric transmission (or establishing an actual long-distance spatial link). The signal light and pilot light experience consistent phase distortion when traveling through the simulated turbulent channel. After the signal light pulse and pilot light separately reach a polarization beam splitter, the perpendicularly polarized pilot light is detected by an image sensor, obtaining its intensity distribution. This information is then sent as input to a hardware platform equipped with a neural network model, which generates a real-time prediction of the corresponding turbulence-equivalent phase screen image. The grayscale information from the phase screen is used to control the SLM to generate a compensated phase value, which is then applied to the next signal light pulse. The compensated signal light is then split by the polarization beam splitter and sent to a homodyne (or heterodyne) detector for response.

[0083] Finally, the security key is extracted through post-processing, completing the entire key distribution process.

[0084] In summary, the present invention introduces deep learning to capture the dynamic change characteristics of turbulence of a large number of different links, learn and analyze the complex mapping relationship between light beam propagation and phase results in turbulence, and establish an atmospheric turbulence phase prediction model based on a hierarchical and progressive feature learning network, small sample learning training and dynamic optimization module. It can accurately predict the light beam wavefront phase distortion caused by turbulence in the communication link in a short time and perform real-time channel compensation, thereby meeting the real-time requirements of the channel compensation method for the rapid movement of satellites in the high-altitude platform CV-QKD channel scenario, giving full play to the efficiency and adaptability of the deep learning algorithm, effectively improving the efficiency and stability of the high-altitude platform CV-QKD channel, and providing strong technical support for the practical application of high-altitude platform CV-QKD.

[0085] It should be noted that for the aforementioned embodiments, for simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that this application is not limited by the order of the actions described, because according to this application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily required by this application.

[0086] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0087] In the several embodiments provided in this application, it should be understood that the disclosed methods or systems can be implemented in other ways. For example, the embodiments described above are merely illustrative, and the division of the units described is merely a logical functional division. In actual implementation, other division methods may be used, such as combining or integrating multiple units or components into another system, or ignoring or not implementing certain features.

[0088] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0089] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0090] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable memory. Based on this understanding, the technical solution of this application, or the part that contributes to the existing technology, or all or part of the technical solution can be embodied in the form of a software product. This computer software product is stored in a memory and includes a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of this application.

[0091] Those skilled in the art will appreciate that all or part of the various circuits in the above embodiments may be implemented by instructing related hardware through a program. The program may be stored in a computer-readable memory, which may include a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0092] The above is only an exemplary embodiment of the present disclosure and cannot be used to limit the scope of the present disclosure. That is, any equivalent changes and modifications made according to the teachings of the present disclosure are still within the scope of the present disclosure. After considering the specification and practicing the disclosure herein, those skilled in the art will easily think of the implementation scheme of the present disclosure. This application is intended to cover any variation, use or adaptation of the present disclosure, which follows the general principles of the present disclosure and includes common knowledge or customary technical means in the art that are not recorded in the present disclosure. The description and examples are to be regarded as exemplary only, and the scope and spirit of the present disclosure are defined by the claims.

[0093] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0094] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A high-altitude platform continuous variable quantum key distribution channel intelligent compensation method, characterized in that: The method comprises: Obtaining a current light intensity distribution map of the channel link to be compensated, inputting the map into the constructed atmospheric turbulence phase prediction model, predicting the phase distortion caused by turbulence in the channel link to be compensated, and obtaining a phase prediction result; Compensating the channel link to be compensated by using the phase prediction result; The network architecture of the atmospheric turbulence phase prediction model includes the following components connected in sequence: The input layer is used to obtain the link’s sample data set or current light intensity distribution map; A feature extraction module, configured to extract turbulence dynamic change features from the sample data set or the current light intensity distribution map using a hierarchical and progressive feature learning network; A small sample learning module is used to train and learn the dynamic change characteristics of the turbulence; Dynamic optimization module, used to adjust and optimize the atmospheric turbulence phase prediction model; The output layer is used to output the phase prediction result.

2. The intelligent compensation method for continuous variable quantum key distribution channel on a high altitude platform according to claim 1, characterized in that: The feature extraction module includes: The pixel perception layer is used to obtain the sample data set and adopt the local window attention mechanism of fixed pixel units to capture the pixel-level phase change characteristics of the phase screen from the sample data set to obtain the first-level turbulence dynamic change characteristics; The local correlation layer is used to obtain the first-order turbulence dynamic change characteristics. It constructs a dynamic receptive field through window merging, establishes a spatial correlation model between adjacent turbulence local characteristics, captures the structural changes of turbulence at the mesoscale, and obtains the second-order turbulence dynamic change characteristics. The global modeling layer is used to obtain the dynamic change characteristics of the second-order turbulence, analyze the structural changes of turbulence at the mesoscale and the statistical distribution of turbulence structure, and obtain the dynamic change characteristics of the third-order turbulence; The cross-scale feature interaction module is used to fuse the turbulence dynamic change characteristics of different levels to obtain the turbulence dynamic change characteristics.

3. The intelligent compensation method for continuous variable quantum key distribution channel on a high altitude platform according to claim 1, characterized in that: The small sample learning module includes: The feature reuse channel is used to receive turbulence dynamic change features, build a turbulence feature prototype library including feature vectors corresponding to turbulence of different intensities, and construct an attention fusion gate that coordinates channels and space. Based on the turbulence feature prototype library and attention fusion gate, the key feature maps in the turbulence dynamic change features are identified and the attention heat map is generated. The knowledge distillation channel is used to obtain attention heat maps, use the teacher network trained with big data to supervise and guide the student network trained with small samples, and distill and extract key features.

4. The intelligent compensation method for continuous variable quantum key distribution channel on a high altitude platform according to claim 1, characterized in that: The dynamic optimization module includes one or more optimization methods including turbulence perception enhancement strategy, dynamic network structure adjustment, and multi-task loss function.

5. The intelligent compensation method for continuous variable quantum key distribution channel on a high altitude platform according to claim 1, characterized in that: The method for constructing the atmospheric turbulence phase prediction model includes: inputting the generated sample data set into a neural network model for pre-training to obtain a pre-trained phase result; and using physical consistency conditions to constrain the pre-trained phase result to obtain the atmospheric turbulence phase prediction model.

6. The intelligent compensation method for continuous variable quantum key distribution channel on a high altitude platform according to claim 5, characterized in that: The method for generating the sample data set includes: Build a dynamic communication simulation scenario between the high-altitude platform and the ground station to generate a link data set; The link dataset is characterized by using an atmospheric turbulence phase screen characterization model to obtain a superimposed phase screen caused by turbulence in the corresponding link, and a label dataset is generated; Inputting the label data set into an ideal free-space transmission model of a Gaussian beam to obtain an input light intensity distribution diagram of a Gaussian beam cross section after transmission via a corresponding link, thereby generating a training data set; The sample dataset is obtained according to the mapping relationship between the label dataset and the training dataset.

7. The intelligent compensation method for continuous variable quantum key distribution channel on a high altitude platform according to claim 6, characterized in that: The method for building a dynamic communication simulation scenario between the high-altitude platform and the ground station is: real-time collection of parameter information of the high-altitude platform and the ground station, input into the STK simulation platform, retrieve or load the change process of the refractive index structure constant distribution of the link between the high-altitude platform and the ground station during the communication process, and realize the dynamic communication simulation scenario of the continuous variable quantum key distribution channel of the high-altitude platform.

8. The intelligent compensation method for continuous variable quantum key distribution channel on a high altitude platform according to claim 6, characterized in that: Using physical consistency conditions to constrain the pre-training phase results includes: Substituting the pre-trained phase result into the light intensity propagation physical model to obtain a theoretical light intensity distribution diagram; An input light intensity distribution map corresponding to the pre-training phase result is obtained, a difference between the input light intensity distribution map and the theoretical light intensity distribution map is calculated, a self-supervised loss function is constructed, and training constraints are performed based on the self-supervised loss function.

9. The intelligent compensation method for continuous variable quantum key distribution channel on a high altitude platform according to claim 8, characterized in that: The self-supervised loss function is: ； in, Indicates the total number of samples; represents the input light intensity distribution diagram; It represents the theoretical light intensity distribution diagram.

10. The intelligent compensation method for continuous variable quantum key distribution channel on a high altitude platform according to claim 6, characterized in that: The pre-training loss function of the pre-training is: ； in, Indicates the total number of samples; represents the superimposed phase screen in the label dataset; Represents the pre-training phase result; , represents the linear rectification function, which is a commonly used activation function in neural networks.

Images