A method for predicting short-window gamma-gamma turbulence parameters in satellite-to-ground laser communication

Abstract

This invention discloses a short-window gamma-gamma turbulence parameter prediction method for space-to-ground laser communication, belonging to the technical field of space-to-ground laser communication and atmospheric turbulence channel parameter prediction. This method addresses the problems of traditional methods, such as strong dependence on long observation windows, significant degradation in prediction accuracy under short-window scenarios, and insufficient robustness under low signal-to-noise ratio conditions. It establishes a short-window observation model for the space-to-ground optical link and a gamma-gamma channel statistical model, constructs time-dependent short-window training data, and designs a short-window gamma-gamma network. Temporal features are extracted through a convolutional backbone, and a scintillation exponential physical regularization auxiliary head is used to achieve joint parameter prediction under physical constraints, outputting predicted gamma-gamma distributed parameters. This method can achieve high accuracy and good stability in parameter prediction under short observation windows and low signal-to-noise ratio conditions, and can be used for turbulence channel state characterization at the space-to-ground laser communication receiver.

Description

Technical Field

[0001] This invention belongs to the field of satellite-to-ground laser communication and atmospheric turbulence channel parameter prediction technology, and particularly relates to a short-window gamma-gamma turbulence parameter prediction method for satellite-to-ground laser communication. Background Technology

[0002] Space-to-ground laser communication can provide high-bandwidth, low-latency data transmission capabilities for integrated space-based and ground-based information networks. However, when the laser beam passes through the atmosphere, it is affected by refractive index fluctuations, resulting in a significant scintillation effect, which leads to random fluctuations in received light intensity and reduces link reliability. Currently, under moderate to strong turbulence conditions, gamma-gamma distributions are commonly used to characterize the statistical properties of received light intensity, among which large-scale fading parameters... and small-scale fading parameters It is a key parameter describing the statistical state of a turbulent channel.

[0003] Existing parameter prediction methods mainly include maximum likelihood methods, expectation-maximization algorithms, and generalized method of moments. These methods are generally based on the assumption of independent and identically distributed samples, and often require a long observation window to obtain stable prediction results. When the satellite-to-ground link is affected by changes in zenith angle, wind speed, solar heating, and local non-stationary turbulence, the statistical characteristics of atmospheric turbulence may drift within a short time scale, and the statistical results obtained through long windows will be difficult to reflect the current true channel state in a timely manner.

[0004] Some deep learning methods have attempted to predict turbulence parameters using histogram features or multi-aperture receiver structures. However, histogram construction destroys the original temporal information and typically still relies on long observation windows. Multi-aperture schemes require dedicated hardware, resulting in high engineering deployment costs. While Long Short-Term Memory (LSTM) networks can process temporal signals, their sequential computation limits inference speed, and their performance advantage in short-window gamma-gamma parameter prediction tasks is not stable. Therefore, there is an urgent need for a gamma-gamma parameter prediction method that can maintain high accuracy, low complexity, and strong real-time performance under conditions of short observation windows, low signal-to-noise ratios, and rapidly changing turbulence. Summary of the Invention

[0005] To address the problems of low prediction accuracy, poor convergence with low signal-to-noise ratio, and difficulty in fully utilizing time-related information in existing technologies under short-window conditions, this invention proposes a short-window gamma-gamma turbulence parameter prediction method for satellite-to-ground laser communication. The aim is to utilize the autocorrelation and fading duration characteristics in the original time-series light intensity sequence to achieve fast and high-precision prediction of atmospheric turbulence channel parameters for satellite-to-ground laser communication without relying on long-window statistics and dedicated hardware.

[0006] The above objectives are achieved through the following technical solutions:

[0007] A method for predicting short-window gamma-gamma turbulence parameters in satellite-to-ground laser communication, comprising the following steps:

[0008] Step S1: Based on the characteristics of the satellite-to-ground laser communication link, establish a short-window reception. Signal The correspondence between the sequence and the atmospheric turbulence fading characteristics, wherein the atmospheric turbulence fading characteristics are characterized by gamma-gamma atmospheric turbulence statistical laws.

[0009] Step S2: Based on the correspondence determined in Step S1, combine the statistical laws of gamma-gamma atmospheric turbulence to generate a time-dependent light intensity sequence and superimpose Gaussian white noise to construct the labeled dataset required for network training, validation and testing.

[0010] Step S3: Construct a short-window gamma-gamma network t, and input the labeled dataset generated in step S2 into the network. The short-window gamma-gamma network includes a convolutional feature extraction backbone, a global average pooling layer, and a layer for output... and The distribution of predicted values ​​predicts the master head and is used to output the flicker index. Physical regularization auxiliary head for predicted values;

[0011] Step S4: Based on the short-window gamma-gamma network constructed in step S3, the network is trained end-to-end by fusing the joint loss function of distributed prediction loss and physical constraint loss to obtain the trained short-window gamma-gamma parameter prediction model.

[0012] Step S5: In the online parameter prediction stage, the sequence of short-window received signals from the satellite-to-ground laser communication system to be detected is input into the short-window gamma-gamma parameter prediction model trained in step S4, and the output parameters are obtained after one forward propagation. Predicted value and parameters Predicted value and the and This serves as a prediction of the current state of the turbulent channel in the satellite-to-ground laser link.

[0013] Furthermore, the short window reception mentioned in step S1 Signal The correspondence between the sequence and the atmospheric turbulence fading characteristics is as follows:

[0014] ;

[0015] in, Discrete time The short-window received signal sequence, For photodetector responsivity, For deterministic path loss, For the normalized turbulent light intensity sequence, The noise is additive white Gaussian noise, and N is the time length of the short-window received signal sequence;

[0016] Furthermore, in step S1, the atmospheric turbulence fading characteristics are characterized using gamma-gamma statistical laws, and the received light intensity... It can be expressed as the product of two independent gamma random variables, i.e.:

[0017] ;

[0018] in, Represents large-scale turbulence components. Let represent the small-scale turbulent components, and:

[0019] ;

[0020] ;

[0021] The corresponding gamma-gamma probability density function is:

[0022] ;

[0023] in, For large-scale fading parameters, For small-scale fading parameters, This is a modified Bessel function of the second type.

[0024] Furthermore, in step S2, to ensure that the generated training samples meet the physical conditions, Rytov variance is introduced. With parameters , The resolution relationship between them. For a link height of... zenith angle is For a satellite-to-ground laser communication link, the Rytov variance satisfies:

[0025] ;

[0026] in, Indicates the wavenumber of light. λ is the wavelength, and π is pi. Indicates height The refractive index structure constant at that location; This represents the secant function.

[0027] Based on this, the parameters and They respectively satisfy:

[0028] ;

[0029] ;

[0030] flicker index With parameters and The following relationship exists between them:

[0031]

[0032] Further, in step S2, to generate the time-dependent light intensity sequence, two first-order autoregressive Gaussian processes are constructed, with autocorrelation coefficients satisfying:

[0033] ;

[0034] ;

[0035] in, Represents large-scale turbulent components The autocorrelation coefficient corresponding to a first-order autoregressive process, Represents large-scale turbulent components The coherence time, This represents the autocorrelation coefficient of the small-scale turbulent component Y corresponding to the first-order autoregressive process. This represents the coherence time of the small-scale turbulent component Y. Represents the natural constant.

[0036] Then, large-scale components with time correlation are obtained through Gaussian cumulative distribution function mapping and gamma inverse cumulative distribution function mapping. and small-scale components Thus forming a light intensity sequence :

[0037] ;

[0038] Based on this, additive white Gaussian noise is superimposed on the light intensity sequence to obtain short-window reception. Signal The sequences are divided into training, validation, and test sets.

[0039] Furthermore, the short-window gamma-gamma network described in step S3 includes a convolutional feature extraction backbone, a global average pooling layer, and a layer for output. and The distribution of predicted values ​​predicts the master head and is used to output the flicker index. Physical regularization auxiliary head for predicted values.

[0040] The convolutional feature extraction backbone comprises three cascaded one-dimensional convolutional feature extraction modules with kernel sizes of 7, 5, and 3, respectively, a stride of 2, and output channels of 32, 64, and 128, respectively. Each convolutional feature extraction module includes a one-dimensional convolutional layer, a batch normalization layer, and a modified linear unit activation layer. The first convolutional block focuses on extracting large-scale slow-changing patterns, the second convolutional block extracts medium-scale changing features, and the third convolutional block further enhances the representation of small-scale flickering features. The output of the convolutional feature extraction backbone is processed by a global average pooling layer to form a 128-dimensional bottleneck feature vector.

[0041] Set the output after the global average pooling layer. and The distribution of predicted values ​​predicts the master head and is used to output the flicker index. Physical regularization auxiliary head for predicted values.

[0042] The distributed prediction master head uses a 128-dimensional input, a 64-dimensional hidden layer, and a 2-dimensional output layer. Predicted value and Predicted value ;

[0043] The physical regularization auxiliary head employs a 128-dimensional input, a 64-dimensional hidden layer, and a 1-dimensional output layer, outputting a predicted flicker index. .

[0044] The physical regularization auxiliary head does not change the final independent output target of the invention, but by comparing the network output with... The analytical relationship coupling forces the shared backbone to learn feature representations that satisfy the gamma-gamma physical laws.

[0045] Furthermore, the joint loss function of the fusion distribution prediction loss and physical constraint loss mentioned in step S4 is:

[0046] ;

[0047] in, MSE(·) represents the total loss, and MSE(·) represents the mean squared error function. Indicates parameters The predicted value, Indicates parameters The predicted value, This represents the predicted value of the flicker index;

[0048] During training , and The loss is calculated after standardization. The Adam optimizer, cosine annealing learning rate scheduling, batch size of 256 and gradient pruning with a maximum norm of 1.0 are used, and the best training parameters are retained based on the principle of minimizing the validation set loss, to obtain the trained short-window gamma-gamma parameter prediction model.

[0049] The evaluation index uses normalized mean square error:

[0050] ;

[0051] in, These represent the actual and predicted values ​​of the corresponding parameters, respectively.

[0052] Furthermore, in the online parameter prediction stage described in step S5, the short-window received signal sequence to be detected is input into the short-window gamma-gamma parameter prediction model trained in step S4, and the parameters are simultaneously obtained in one forward propagation. , and .in and The final output is the predicted gamma-gamma parameters.

[0053] The advantages of this invention compared to the prior art are:

[0054] 1. Addressing the core pain points of existing methods such as maximum likelihood estimation, generalized method of moments, long short-term memory networks, and histogram convolutional networks, which rely on long observation windows, experience significant decreases in prediction accuracy under short window conditions, and suffer from poor convergence in low signal-to-noise ratio environments, this invention directly processes the original short-window light intensity time-series data, rather than histogram statistical features. It fully preserves the autocorrelation characteristics and fading duration information of the light intensity signal, without relying on long-window statistics under the assumption of independent and identically distributed signals. This fundamentally breaks through the strong dependence of traditional methods on long observation windows, while avoiding the loss of original time-series information during the histogram feature construction process.

[0055] 2. This invention introduces physical consistency constraints through a scintillation index auxiliary head, enabling the network to output not only large-scale fading parameters. Small-scale fading parameters The prediction results still explicitly utilize the scintillation index during the training phase. and , The analytical coupling relationship imposes physical constraints on the parameter prediction process, which greatly improves the consistency between the model output and the gamma-gamma channel model and the physical laws of turbulence.

[0056] 3. This invention employs a three-layer one-dimensional convolutional backbone structure to extract multi-scale temporal features, replacing traditional serial recursive structures and serial temporal models such as long short-term memory networks with a lightweight one-dimensional convolutional structure, significantly improving inference efficiency; the network contains only 50K trainable parameters, and the single forward inference time is at the millisecond level, ensuring the accuracy of short-window prediction while being more suitable for the requirements of real-time short-window processing and engineering real-time deployment.

[0057] 4. This invention demonstrates stable and accurate prediction performance under extreme conditions such as short observation windows, low signal-to-noise ratio, and rapidly changing turbulence. With 512 sampling points and a signal-to-noise ratio of 20dB, the average normalized mean square error can reach 0.065, which is 43% and 42% lower than the traditional maximum likelihood method and generalized method of moments, respectively. Under low signal-to-noise ratio conditions of 5dB, the error is reduced by about 70% compared to the maximum likelihood method, and the overall performance is significantly better than the existing mainstream methods.

[0058] 5. This invention does not require dedicated hardware such as multi-aperture receivers, making engineering implementation simpler; at the same time, the gamma-gamma parameters output by the model can be further mapped to Rytov variance and path integral turbulence intensity, which can provide a unified parameter interface for state perception and adaptive control of satellite-to-ground laser communication links, and has strong engineering application value and scenario scalability. Attached Figure Description

[0059] Figure 1 This is a flowchart of the short-window gamma-gamma parameter prediction model of the present invention.

[0060] Figure 2 This is a schematic diagram of the architecture of the short-window gamma-gamma network of the present invention.

[0061] Figure 3 This invention relates to different observation window lengths and different methods. Parameter prediction bar chart comparison.

[0062] Figure 4 This invention relates to different observation window lengths and different methods. Parameter prediction bar chart comparison.

[0063] Figure 5 This invention relates to different signal-to-noise ratio conditions and different methods. Parameter prediction bar chart comparison.

[0064] Figure 6 This invention relates to different signal-to-noise ratio conditions and different methods. Parameter prediction bar chart comparison. Detailed Implementation

[0065] The present invention will be further described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following embodiments.

[0066] This embodiment uses the downlink of satellite-to-ground laser communication as an example to illustrate a short-window gamma-gamma turbulence parameter prediction method for satellite-to-ground laser communication according to the present invention.

[0067] like Figure 1 The present invention provides a method for predicting short-window gamma-gamma turbulence parameters in satellite-to-ground laser communication, comprising the following steps:

[0068] Step S1: Based on the characteristics of the satellite-to-ground laser communication link, establish the correspondence between the short-window received signal sequence and the atmospheric turbulence fading characteristics, wherein the atmospheric turbulence fading characteristics are characterized by gamma-gamma atmospheric turbulence statistical law.

[0069] In this embodiment, the downlink of the satellite-to-ground laser communication operates at a wavelength of 1550 nm, with a satellite orbital altitude of 500 km and a zenith angle of 0°. A short-window received signal sequence of length N is used; in this embodiment, N is successively set to 64, 128, 256, 512, and 1024 to determine large-scale fading parameters. and small-scale fading parameters Comparison of mean squared error estimates.

[0070] The short window receives Signal The correspondence between the sequence and the atmospheric turbulence fading characteristics is as follows:

[0071] ;

[0072] in, Discrete time The short-window received signal sequence, For photodetector responsivity, For deterministic path loss, It is a time-dependent light intensity sequence. It is additive white Gaussian noise.

[0073] The atmospheric turbulence fading characteristics are characterized using gamma-gamma statistical laws, specifically, the received light intensity... It can be expressed as the product of two independent gamma random variables, i.e.:

[0074]

[0075] in, Represents large-scale turbulence components. Represents the small-scale turbulent components and satisfies:

[0076] ;

[0077] ;

[0078] The corresponding gamma-gamma probability density function for:

[0079] ;

[0080] in, For large-scale fading parameters, For small-scale fading parameters, This is a modified Bessel function of the second kind. This represents the gamma function.

[0081] Step S2, receiving based on the short window determined in step S1 Signal The correspondence between the sequence and atmospheric turbulence fading characteristics is as follows: a time-dependent light intensity sequence is generated by combining the gamma-gamma atmospheric turbulence statistical law and superimposing Gaussian white noise to construct a labeled dataset.

[0082] To ensure that sample generation meets the physical realizability conditions, Rytov variance and parameters are introduced. , The resolution relationship between them. For a link height of... zenith angle is downlink, Rytov variance satisfy:

[0083] ;

[0084] in, Indicates the wavenumber of light. , Let π be the mathematical constant pi and λ be the wavelength. Indicates height The refractive index structure constant at that location.

[0085] Based on this, the parameters and They respectively satisfy:

[0086] ;

[0087] ;

[0088] in, This represents an exponential function with the natural constant as its base.

[0089] flicker index With parameters and The following conditions must be met:

[0090] .

[0091] In this embodiment, in order to uniformly cover the gamma-gamma parameter space from strong turbulence to weak turbulence, to achieve approximately consistent relative resolution across different turbulence intensity ranges, and to avoid insufficient coverage of linear sampling in small parameter regions, the parameters are optimized within the range of [1.5, 20]. and Log-interval sampling was performed, and parameter combinations satisfying the above analytical relationship were retained within a 15% relative tolerance, resulting in 294 effective turbulence conditions. For each parameter combination, two first-order autoregressive Gaussian processes were constructed, with autocorrelation coefficients satisfying:

[0092] ;

[0093] ;

[0094] in, Represents large-scale turbulent components The autocorrelation coefficient corresponding to a first-order autoregressive process, Represents large-scale turbulent components The coherence time, This represents the autocorrelation coefficient of the small-scale turbulent component Y corresponding to the first-order autoregressive process. Let Y represent the coherence time of the small-scale turbulent component Y, and satisfy the following conditions: Subsequently, based on the inverse transform sampling method, and through Gaussian cumulative distribution function mapping and gamma distribution inverse cumulative distribution function mapping, the large-scale fading sequence X[n] and small-scale fading sequence Y[n] that maintain time correlation are obtained, and a time-correlated light intensity sequence is constructed. :

[0095] .

[0096] Based on this, the light intensity sequence with time correlation is... Gaussian white noise is superimposed to obtain a short-window received signal sequence; this sequence is then divided into a training set, a validation set, and a test set. In this embodiment, the number of samples in the training set, validation set, and test set are 20,000, 4,000, and 1,500, respectively.

[0097] Step S3: Construct a short-window gamma-gamma network and input the labeled dataset generated in step S2 into the short-window gamma-gamma network. The short-window gamma-gamma network includes a convolutional feature extraction backbone, a global average pooling layer, and a parameter for large-scale fading. and small-scale fading parameters The distribution of predicted values ​​predicts the master head and is used to output the flicker index. Physical regularization auxiliary head for predicted values.

[0098] The short-window gamma-gamma network constructed in this embodiment is as follows. Figure 2 As shown, the network includes a convolutional feature extraction backbone, a global average pooling layer, a distribution prediction head, and a physical regularization auxiliary head.

[0099] The convolutional feature extraction backbone is constructed using three cascaded one-dimensional convolutional feature extraction modules, designed with multi-scale receptive fields to address the large, medium, and small-scale undulation characteristics of space-to-ground turbulence. Specifically, the kernel sizes of the three one-dimensional convolutional feature extraction modules are 7, 5, and 3 respectively, with a stride of 2 for each module, and output channels of 32, 64, and 128 respectively. Each one-dimensional convolutional feature extraction module consists of a cascaded one-dimensional convolutional layer, a batch normalization layer, and a modified linear unit activation layer. Specifically, the first convolutional module uses a large kernel to extract slow-changing feature patterns related to large-scale fading in short-window sequences; the second convolutional module uses a medium kernel to extract medium-scale turbulent undulation features; and the third convolutional module uses a small kernel to enhance the representation of fine-grained small-scale flicker features, achieving complete extraction of turbulent features across all scales. The convolutional feature extraction backbone ultimately outputs a 128-channel one-dimensional depth feature map.

[0100] A global average pooling layer is connected to the output of the convolutional feature extraction backbone. It performs a global average pooling operation on the input 128-channel one-dimensional depth feature map, compressing the feature map into a 128-dimensional bottleneck feature vector. This reduces the number of network parameters and suppresses the risk of overfitting while preserving global turbulence feature information.

[0101] The 128-dimensional bottleneck feature vector output by the global average pooling layer is synchronously input into the parallel distributed prediction master head and the physical regularization auxiliary head. The two prediction heads are independent of each other and share the same set of deep features extracted by the backbone. The distributed prediction master head is a fully connected network structure, adopting a three-layer architecture of "128-dimensional input layer - 64-dimensional hidden layer - 2-dimensional output layer". The hidden layer uses the ReLU activation function, and the output layer finally outputs two scalar results, namely the large-scale fading parameters. Predicted value Small-scale fading parameters Predicted value The physical regularization auxiliary head serves as the final independent output target of this network. It is a fully connected network structure employing a three-layer architecture: a 128-dimensional input layer, a 64-dimensional hidden layer, and a 1-dimensional output layer. The hidden layers use the ReLU activation function, and the output layer ultimately outputs a single scalar result: the flicker index. Predicted value The physical regularization auxiliary head does not participate in the final output of channel parameters during the online prediction stage of this invention. Its core function is to output the scintillation index as described in step S2 through the network. With parameters and The strong coupling formed by the physical relationships between them imposes physical constraints on the shared convolutional feature extraction backbone, enabling the backbone network to learn feature representations that conform to the physical laws of gamma-gamma turbulence distribution. Therefore, the labeled dataset generated in step S2 is input into the network.

[0102] Step S4: Based on the short-window gamma-gamma network constructed in step S3, the joint loss function of distributed prediction loss and physical constraint loss is fused to train the short-window gamma-gamma network end-to-end, resulting in a trained short-window gamma-gamma parameter prediction model. For example... Figure 2 As shown, during training, the distribution is used to predict the output of the master head simultaneously. , and the output of the physical regularization auxiliary head Construct the joint loss function:

[0103] ;

[0104] in, , The large-scale fading parameters and small-scale fading parameters are respectively used, and the effective turbulence conditions obtained in step S2 are applied. and These are the predicted parameter values ​​output by the main head. For parameters and The calculated flicker index, This is the predicted flicker index value output by the auxiliary head. During training, firstly... , and The target value is standardized before being used in the loss calculation; then the Adam optimizer is used to update the parameters. In this embodiment, the training epochs are 50, the batch size is 256, and the initial learning rate is... The cosine annealing learning rate scheduling and gradient pruning strategy with a maximum norm of 1.0 are adopted. Finally, the optimal model parameters are retained according to the principle of minimizing the validation set loss, and the trained short-window gamma-gamma parameter prediction model is obtained.

[0105] To evaluate the prediction accuracy of the model, this embodiment uses normalized mean square error as the evaluation metric:

[0106] ;

[0107] in, These represent the actual and predicted values ​​of the corresponding parameters, respectively.

[0108] Step S5: In the online parameter prediction stage, the short-window received signal sequence of the satellite-to-ground laser communication to be detected is input into the short-window gamma-gamma parameter prediction model trained in step S4, and after one forward propagation, the large-scale fading parameters are output. Predicted value and small-scale fading parameters Predicted value and will and This serves as the prediction result for the turbulent channel state of the satellite-to-ground laser link at the current moment. To verify the effectiveness of the method described in this embodiment, after completing the training in step S4 and performing the online prediction in step S5, the prediction results under different observation window lengths and different signal-to-noise ratios are tested.

[0109] like Figure 3 As shown in the figure, the method of this invention and other methods are compared under different observation window lengths. Parameter prediction comparison results. The comparison objects include the method of this invention, the maximum likelihood method, the generalized method of moments, long short-term memory networks, and histogram convolutional networks. The method of this invention was compared at 64, 128, 256, 512, and 1024 sampling points. The parameter-normalized mean square error is generally lower than other methods, and it increases with the increase of the observation window length. The continuous decrease in parameter prediction error indicates that the method can effectively utilize relevant information in the short-window time series.

[0110] like Figure 4 As shown in the figure, the method of this invention and other methods are compared under different observation window lengths. Parameter prediction comparison results. Compared with methods relying on independent and identically distributed statistics or histogram features, the method of this invention maintains lower performance under most observation window conditions. Normalized mean square error of parameters, especially in and The advantages become even more pronounced under certain conditions.

[0111] like Figure 5 As shown in the figure, the method of this invention and other methods are compared under different signal-to-noise ratio conditions. Parameter prediction comparison results. Under the given conditions, when the signal-to-noise ratio increases from 5dB to 30dB, the performance of each method... The overall parameter prediction error decreased, but the method of this invention maintained optimal or near-optimal performance at all signal-to-noise ratios, with the most significant advantage under a low signal-to-noise ratio of 5dB.

[0112] like Figure 6 As shown in the figure, the method of this invention and other methods are compared under different signal-to-noise ratio conditions. Parameter prediction comparison results. It can be seen that the method of the present invention significantly outperforms the maximum likelihood method and long short-term memory network under low signal-to-noise ratio conditions, and maintains high stability under mixed signal-to-noise ratio training conditions.

[0113] For those skilled in the art, without departing from the principles of this invention, equivalent substitutions or combinations can be made to the sampling window length, training signal-to-noise ratio distribution, number of convolution channels, and parameter prediction process. All such equivalent substitutions should fall within the protection scope of this invention.

Claims

1. A method for predicting short-window gamma-gamma turbulence parameters in satellite-to-ground laser communication, characterized in that, Includes the following steps: Step S1: Based on the characteristics of the satellite-to-ground laser communication link, establish the correspondence between the short-window received signal sequence and the atmospheric turbulence fading characteristics, wherein the atmospheric turbulence fading characteristics are characterized by gamma-gamma atmospheric turbulence statistical law. Step S2: Based on the correspondence determined in Step S1, and combined with the statistical laws of gamma-gamma atmospheric turbulence, a time-dependent light intensity sequence is generated and superimposed with Gaussian white noise to construct a labeled dataset. The specific method is as follows: S2.

1. Introducing Rytov variance With large-scale fading parameters and small-scale fading parameters The resolution relationship between them, for a link height of zenith angle is Rytov variance of satellite-to-ground laser communication link satisfy: in, Indicates the wavenumber of light. , Pi is the mathematical constant, and λ is the wavelength. Indicates height The refractive index structure constant at that location; Represents the secant function; Based on this, large-scale fading parameters and small-scale fading parameters They respectively satisfy: in, Represents an exponential function with the natural constant as its base; flicker index and and The following relationship exists between them: S2.

2. Construct two first-order autoregressive Gaussian processes whose autocorrelation coefficients satisfy: In the formula, Represents large-scale turbulent components The autocorrelation coefficient corresponding to a first-order autoregressive process, Represents large-scale turbulent components The coherence time, This represents the autocorrelation coefficient of the small-scale turbulent component Y corresponding to the first-order autoregressive process. This represents the coherence time of the small-scale turbulent component Y. Represents the natural constant; Then, large-scale components with time correlation are obtained through Gaussian cumulative distribution function mapping and gamma inverse cumulative distribution function mapping. and small-scale components This forms a time-dependent light intensity sequence. : Based on this, additive white Gaussian noise is superimposed on the time-dependent light intensity sequence to obtain a short-window received signal sequence, which is then divided into a training set, a validation set, and a test set. Step S3: Construct a short-window gamma-gamma network and input the labeled dataset generated in step S2 into the short-window gamma-gamma network. The short-window gamma-gamma network includes a convolutional feature extraction backbone, a global average pooling layer, and a layer for outputting large-scale fading parameters. and small-scale fading parameters The distribution of predicted values ​​predicts the master head and is used to output the flicker index. Physical regularization auxiliary head for predicted values; Step S4: Based on the short-window gamma-gamma network constructed in step S3, the joint loss function of distributed prediction loss and physical constraint loss is fused to train the short-window gamma-gamma network end-to-end, resulting in a trained short-window gamma-gamma parameter prediction model. Step S5: In the online parameter prediction stage, the short-window received signal sequence of the satellite-to-ground laser communication to be detected is input into the short-window gamma-gamma parameter prediction model trained in step S4, and after one forward propagation, the large-scale fading parameters are output. Predicted value and small-scale fading parameters Predicted value and will and This serves as a prediction of the current state of the turbulent channel in the satellite-to-ground laser link.

2. The method for predicting short-window gamma-gamma turbulence parameters in satellite-to-ground laser communication according to claim 1, characterized in that, The correspondence between the short-window received signal sequence and the atmospheric turbulence fading characteristics in step S1 is as follows: in, Discrete time The short-window received signal sequence, For photodetector responsivity, For deterministic path loss, It is a time-dependent light intensity sequence. The noise is additive white Gaussian noise, and N is the time length of the short-window received signal sequence.

3. The method for predicting short-window gamma-gamma turbulence parameters in satellite-to-ground laser communication according to claim 2, characterized in that, In step S1, the atmospheric turbulence fading characteristics are characterized using gamma-gamma statistical laws. Specifically, the received light intensity... It can be expressed as the product of two independent gamma random variables, i.e.: in, Represents large-scale turbulence components. Let represent the small-scale turbulent components, and: The corresponding gamma-gamma probability density function for: in, For large-scale fading parameters, For small-scale fading parameters, This is a modified Bessel function of the second kind. This represents the gamma function.

4. A method for predicting short-window gamma-gamma turbulence parameters in satellite-to-ground laser communication according to claim 1, 2, or 3, characterized in that, The short-window gamma-gamma network described in step S3 includes a convolutional feature extraction backbone, a global average pooling layer, and a layer for output. and The distribution of predicted values ​​predicts the master head and is used to output the flicker index. The physical regularization auxiliary head for the predicted values, where: The convolutional feature extraction backbone consists of three cascaded one-dimensional convolutional feature extraction modules with kernel sizes of 7, 5, and 3, stride of 2, and output channels of 32, 64, and 128, respectively. Each convolutional feature extraction module includes a one-dimensional convolutional layer, a batch normalization layer, and a modified linear unit activation layer. The first convolutional block extracts large-scale slow-change patterns, the second convolutional block extracts medium-scale change features, and the third convolutional block further enhances the representation of small-scale flickering features. The local average pooling layer is used to extract the output of the convolutional features into a 128-dimensional bottleneck feature vector. The distribution prediction master head employs a 128-dimensional input, a 64-dimensional hidden layer, and a 2-dimensional output layer, outputting... Predicted value and Predicted value ; The physical regularization auxiliary head employs a 128-dimensional input, a 64-dimensional hidden layer, and a 1-dimensional output layer to output a predicted flicker index value. .

5. The method for predicting short-window gamma-gamma turbulence parameters in satellite-to-ground laser communication according to claim 4, characterized in that, The joint loss function of the fusion distribution prediction loss and physical constraint loss mentioned in step S4 is: in, MSE(·) represents the total loss, and MSE(·) represents the mean squared error function. Indicates parameters The predicted value, Indicates parameters The predicted value, This represents the predicted value of the flicker index.

6. The method for predicting short-window gamma-gamma turbulence parameters in satellite-to-ground laser communication according to claim 5, characterized in that, Step S4 describes end-to-end training of the short-window gamma-gamma network, during which... , and After standardization, the loss is calculated, and the Adam optimizer, cosine annealing learning rate scheduling, batch size of 256 and gradient pruning with maximum norm of 1.0 are used. The best training parameters are retained based on the principle of minimizing the loss on the validation set, and the trained short-window gamma-gamma parameter prediction model is obtained. The evaluation index uses normalized mean square error: in, These represent the actual and predicted values ​​of the corresponding parameters, respectively.

Images