Refractive index structure constant adaptive forecasting method and system based on atmospheric turbulence multi-scale characteristics

Abstract

The invention discloses a refractive index structure constant adaptive forecasting method and system based on atmospheric turbulence multi-scale characteristics. The method comprises the following steps: 1, measuring and preprocessing an atmospheric refractive index structure constant; 2, constructing a spatial-temporal feature extraction module of an atmospheric turbulence refractive index structure constant; 3, learning a dependency relationship representing a long-time sequence by using a gating mechanism, and constructing a long-time correlation feature extraction module; 4, capturing dependency relationships of different times through convolution kernels of different time steps, and constructing a short-time correlation feature extraction module; 5, designing a self-adaptive turbulence multi-scale feature fusion module which comprises three groups of bidirectional cross attention modules to realize alignment and fusion among different turbulence feature extraction modules; and introducing a gating mechanism, performing dynamic control and feature selection on output information streams of the three groups of bidirectional cross attention modules, and simulating the change of importance of different scale features of turbulence along with conditions to obtain a predicted value.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of atmospheric optical turbulence and artificial intelligence, specifically relating to a refractive index structure constant based on the multi-scale characteristics of atmospheric turbulence. An adaptive forecasting method and system. Background Technology

[0002] When light waves propagate through the troposphere, they are easily affected by atmospheric turbulence, resulting in phenomena such as field strength attenuation, phase fluctuations, and beam distortion, leading to signal fading or even interruption. Therefore, in-depth research on high-precision prediction methods for atmospheric turbulence parameters is of great significance for studying propagation effects and developing compensation schemes in advance to ensure the stability and reliability of communication and detection systems.

[0003] Atmospheric turbulent refractive index structure constant It is an important parameter characterizing the undulations of the medium in the atmospheric propagation path of electromagnetic waves. Currently, the atmospheric refractive index structure constant is predicted. The main methods include numerical weather prediction and machine learning. Numerical weather prediction combines atmospheric forecasting models and... A parameterized model, taking initial meteorological conditions and boundary conditions as input, forecasts weather patterns over a period of time. However, due to the insufficient regional adaptability of existing meteorological forecast results and parameterized models, the simulation error of cross-regional optical wave transmission links has increased significantly. Furthermore, numerical forecasting requires iteratively solving a large number of partial differential equations to describe complex physical processes, relying on powerful computing resources and long computation time, making it difficult to meet real-time forecasting needs. Existing machine learning forecasting methods, through training and analysis of large amounts of measured data, seek patterns from the training dataset, continuously iterate to improve their accuracy, and ultimately obtain the best possible forecasting effect. Machine learning forecasting methods are divided into two categories based on the different learning relationships: one is to adaptively construct... The mapping relationship between Lionis and relevant meteorological parameters, for example: Lionis uses random forest model, decision tree model, gradient boosting regression model and single-layer artificial neural network respectively, based on 11 months of meteorological parameters of the port of Piraeus and Actual measurement data for marine atmospheric environment Prediction. Su Changdong et al. from the Anhui Institute of Optics and Fine Mechanics, Chinese Academy of Sciences, proposed a prediction method based on an adaptive niche genetic algorithm and a backpropagation neural network, and utilized Antarctic measurements... The data and meteorological parameters were used for training to predict... The results generally agree well with the measured results, but this type of method relies on local real-time meteorological parameters and cannot achieve long-term forecasts. The second type of machine learning forecasting method is based on past atmospheric refractive index structure constants. Real data, the time-varying characteristics of turbulent flow are learned by machine learning method, so as to predict future atmospheric refractive index structure constant. Ma et al. proposed to use recurrent neural network and long short-term memory network to predict the next time point of Results. However, the space-time variation of turbulent flow is severe and complex, and the model only considers the time-varying characteristics of turbulent flow, resulting in low prediction accuracy. SUMMARY

[0004] In order to solve the problem of low prediction accuracy caused by insufficient consideration of turbulent flow characteristics in the existing machine learning prediction method, the present application provides a refractive index structure constant adaptive prediction method and system based on the multi-scale characteristics of atmospheric turbulence.

[0005] To achieve the above purpose, the technical scheme adopted by the present application is:

[0006] The refractive index structure constant adaptive prediction method based on the multi-scale characteristics of atmospheric turbulence comprises the following steps:

[0007] Step 1: Develop atmospheric refractive index structure constant Measurement, the collected atmospheric refractive index structure constant space-time data is preprocessed to represent the space-time distribution of atmospheric refractive index structure constant .

[0008] Step 2: Based on the turbulent motion process and the space-time distribution characteristics of turbulence, first use sliding window to extract local features, generate multiple input-output data pairs, then use different convolution kernels to learn the local complex mixed structure representing the turbulent motion process, and construct the space-time feature extraction module of atmospheric turbulence refractive index structure constant , which is used to extract the space-time features of atmospheric turbulence refractive index structure constant .

[0009] Step 3: Atmospheric turbulence is influenced by historical state (such as thermal stratification, wind speed profile evolution, etc.), and has large-scale long-time correlation. According to the long-time correlation of turbulence, the three-layer stacked gate mechanism is used to learn the dependence relationship representing the long-time sequence, and the long-time correlation feature extraction module of atmospheric turbulence refractive index structure constant is constructed, which is used to extract the long-time correlation features of atmospheric turbulence refractive index structure constant , and improve the long-time prediction accuracy of atmospheric refractive index structure constant.

[0010] Step 4: According to the short-time correlation of local turbulence, for the time dimension of each layer height, different time step convolution kernels are used to make the model capture the dependence relationship of different time, and the long-time correlation feature extraction module of atmospheric turbulence refractive index structure constant a short-time correlation feature extraction module for extracting the atmospheric turbulence refractive index structure constant a short-time correlation feature.

[0011] Step five: design an adaptive turbulence multi-scale feature fusion module, which includes three groups of bidirectional cross attention modules, the adaptive turbulence multi-scale feature fusion module is used to realize the alignment and fusion between the three different turbulence feature extraction modules of the space-time feature extraction module, the long-time correlation feature extraction module and the short-time correlation feature extraction module, so that the model can form more accurate and consistent profile prediction from the multi-scale characteristics of turbulence. In addition, a gating mechanism is introduced to dynamically control and select the output information flow of the above three groups of bidirectional cross attention modules, simulating the importance of turbulence features of different scales changing with conditions (such as the intermittency of turbulence), and finally obtaining the prediction value of the atmospheric turbulence refractive index structure constant .

[0012] As a preferred, in step one, the data collection scheme needs to consider the measurement equipment, measurement location and measurement time, etc. The preprocessing of the atmospheric refractive index structure constant mainly includes data extraction and data imaging.

[0013] As a preferred, in step one, the preprocessing of the atmospheric refractive index structure constant includes:

[0014] Data extraction: extract the measurement values of the atmospheric refractive index structure constant at different times and altitudes;

[0015] Data imaging: convert the extracted measurement values into two-dimensional images according to the order of height and time.

[0016] As a preferred, in step two, turbulence is a complex motion, which exhibits a variety of irregular patterns in space-time distribution. According to the image generated by the space-time data, local features are extracted using a sliding window, and the data set is enriched, and the data is divided to generate a training set, a validation set and a test set. Different convolution kernels are used to learn to represent the local complex mixed structure of turbulence. The shallow network captures small-scale, high-frequency (detail) vortex features, and the deep network captures large-scale, low-frequency (overall) vortex features.

[0017] As a preferred, in step two, local features are extracted using a sliding window to generate multiple groups of input-output data pairs.

[0018] As a preferred, in step two, the local features are extracted using a sliding window as follows:

[0019] 1) Set the sliding window width; the sliding window covers the time step history data used as input features, and the input features are X ∈ R m ×hwhere R represents the real number field, m represents the input sliding window width, h represents the number of height sampling points, and the output feature is set as Y ∈ R n×h where n represents the output feature window width, forming a set of input-output data pairs;

[0020] 2) Slide the window one step to the right, covering the measurement data of another time step, forming a new input-output data pair;

[0021] 3) Repeat step 2) until all measurement data is segmented; and randomly divide all input-output data pairs into training, validation, and test sets.

[0022] As a preferred, the long-term correlation feature extraction module of step three, each input feature X ∈ R 46×46 is regarded as a multivariate time series, where the behavior is in the height direction, and the column is in the time direction; for an input sequence of length n, the input at each time step is mapped to a 128-dimensional hidden state vector by the network, and the final representation h T ∈ R B ×128 where B represents the sample batch size, then the dimension is increased from 128 to 1656 by linear transformation, and then the vector R B×1656 output by the previous layer is restored to a two-dimensional tensor R B×46×36 , representing the prediction profile of 46 height points in the future 36 time steps in each sample.

[0023] As a preferred, in step four, the short-term correlation extraction module includes 4 layers of one-dimensional causal convolution layers stacked, each layer gradually expands the receptive field by inserting different numbers of holes in the convolution kernel elements, set to grow exponentially, in turn 1, 2, 4, 8; introduce a residual block connection structure, the number of residual blocks n = 4, and finally map it to R 32×36 through a linear layer; the same short-term correlation feature extraction module is used for all height points, and finally R 32×46×36 is obtained in the form.

[0024] As a preferred, in step five, the adaptive turbulent multi-scale feature fusion module unifies the output features of the spatio-temporal feature extraction module, the long-term correlation feature extraction module, and the short-term correlation feature extraction module into a tensor dimension X ∈ R 32×46×36 , where 32 is the batch size, 46 is the number of height sampling points, and 36 is the prediction time step length; three linear layers are used for linear projection, and the projection operation is to transform the time feature length of each height, and the transformations are Q = W q Q raw ∈ R 32×46×d , K = W k K​raw ∈R 32×46×d V = W v V raw ∈R 32×46×d , where Q raw K raw and V raw These are the original query, key, and value, respectively, W q W k and W v Let Q, K, and V be the linear projection weight matrices for the query, key, and value, respectively, and let d be 64.

[0025] The attention output Attn is obtained by using the output of the spatiotemporal feature extraction module as the main branch and the output of the long-term correlation feature extraction module as the auxiliary branch. ab The main branch uses the output of the long-term correlation extraction module, and the auxiliary branch uses the output of the spatiotemporal feature extraction module to obtain the attention output Attn. ba The main branch uses the output of the long-term correlation extraction module, and the auxiliary branch uses the output of the short-term correlation extraction module to obtain the attention output Attn. bc The main branch uses the output of the short-term correlation extraction module, and the auxiliary branch uses the output of the long-term correlation extraction module to obtain the attention output Attn. cb The attention output Attn is obtained by using the output of the short-term correlation extraction module as the main branch and the output of the spatiotemporal feature extraction module as the auxiliary branch. ca The attention output Attn is obtained by using the output of the spatiotemporal feature extraction module as the main branch and the output of the short-term correlation extraction module as the auxiliary branch. ac The formula for calculating attention output is:

[0026]

[0027] Where α is the attention score matrix; thus, the attention output Attn is obtained. ab Attn ba Attn bc Attn cb Attn ca and Attn ac ;

[0028] By mapping to the same dimension using fully connected layers and activation functions, and using three sets of learnable gating weights (Alpha1, Alpha2, and Alpha3) for bidirectional attention, the resulting images are obtained after pairwise fusion.

[0029] Fusion1 = Alpha1 × Attn ab +(1-Alpha1)×Attn ba

[0030] Fusion2 = Alpha2 x Attn bc + (1 - Alpha2) x Attn cb

[0031] Fusion3 = Alpha3 x Attn ca + (1 - Alpha3) x Attn ac

[0032] Two independent gating weights g are introduced to evaluate the importance of auxiliary branches, and the calculation formula of gating weights is as follows: 1, g2 e R 32×46×36

[0033] g1 = σ(W1(Fusion2 - Fusion1))

[0034] g2 = σ(W2(Fusion3 - Fusion1))

[0035] wherein the difference terms Fusion2 - Fusion1 and Fusion3 - Fusion1 respectively depict the feature residuals of the two auxiliary branches relative to the main branch, i.e. represent the complementary information between the main-auxiliary branches, and the difference features are projected within the dimensions via linear transformation weight matrices W1, W2 e R 46×36 ; the results are normalized to the interval [0, 1] through the sigmoid function σ to obtain interpretable attention weight ratios; a three-branch weighted fusion strategy dominated by the main branch is introduced for the feature residuals of the auxiliary branches relative to the main branch, and the final fused feature representation is defined as:

[0036] Fusion = (1 - g1 - g2) · Fusion1 + g1 · Fusion2 + g2 · Fusion3

[0037] wherein Fusion1 is taken as the main branch output, and its weight is expressed by 1 - g1 - g2; on the basis of preferentially retaining the main branch information, the auxiliary branches are dynamically injected only when they provide effective supplements, Fusion2 and Fusion3 are regarded as supplementary enhancement paths, and the gating network only allocates a higher fusion proportion when there is a significant difference between them and the main branch in the semantic space; the dimension of the final output tensor is: R 32×46×36 wherein 32 represents the batch size, 46 represents the number of sampling points in the vertical height profile, and 36 represents the future time steps; the tensor is used to represent the atmospheric refractive index structure constant at each height position within the future 36 time steps the prediction result in the logarithmic scale

[0038] ​As preferred, in step five, the predicted output constitutes a height-time two-dimensional profile, which can be directly used to reproduce the vertical distribution change of future atmospheric turbulence intensity.

[0039] The application also discloses a refractive index structure constant adaptive prediction system based on atmospheric turbulence multi-scale characteristics, which is used for executing the method and comprises the following modules:

[0040] A data acquisition module is used for measuring the atmospheric refractive index structure constant The atmospheric refractive index structure constant is preprocessed to represent the spatial and temporal distribution of the atmospheric refractive index structure constant .

[0041] A spatial and temporal feature extraction module is used for extracting local features by using a sliding window based on the turbulence motion process and the spatial and temporal distribution characteristics of turbulence, generating a plurality of input-output data pairs, and then learning the local mixed structure representing the turbulence motion process by using different convolution kernels.

[0042] A long-time correlation feature extraction module is used for learning the dependency relationship representing a long-time sequence by using a gating mechanism according to the long-time correlation of turbulence.

[0043] A short-time correlation feature extraction module is used for capturing the dependency relationship of different times by using different time step convolution kernels for the time dimension of each height according to the short-time correlation of the local turbulence.

[0044] An adaptive turbulence multi-scale feature fusion module comprises three groups of bidirectional cross-attention modules, the bidirectional cross-attention modules are used for realizing the alignment and fusion between the spatial and temporal feature extraction module, the long-time correlation feature extraction module and the short-time correlation feature extraction module, the gating mechanism is introduced to dynamically control and select the output information flow of the three groups of bidirectional cross-attention modules, the importance of different scale features of turbulence is simulated to change with conditions, and the predicted value of the atmospheric turbulence refractive index structure constant is obtained.

[0045] Compared with the prior art, the application has the beneficial effects that:

[0046] The application deeply studies the turbulence motion process, comprehensively considers the time scale and spatial scale characteristics, constructs different models to effectively learn a plurality of features, designs an adaptive feature fusion method, and finally proposes the adaptive prediction method of the refractive index structure constant based on the atmospheric turbulence multi-scale characteristics, so that the problem of low prediction accuracy caused by ignoring the turbulence characteristics in the prior art machine learning prediction method based on historical data is overcome. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 is the atmospheric refractive index structure constant The measurement data source file and the pseudo-color map after data conversion;

[0048] Figure 2 is a time-space feature extraction module structure diagram of atmospheric refractive index structure constant based on the turbulence characteristics in the preferred embodiment of the present application;

[0049] Figure 3 is a long-time correlation feature extraction module structure diagram of atmospheric refractive index structure constant based on the turbulence characteristics in the preferred embodiment of the present application;

[0050] Figure 4 is a short-time correlation feature extraction module structure diagram of atmospheric refractive index structure constant based on the turbulence characteristics in the preferred embodiment of the present application;

[0051] Figure 5 is a self-adaptive turbulence multi-scale feature fusion module structure diagram in the preferred embodiment of the present application;

[0052] Figure 6 is a predicted output atmospheric refractive index structure constant height-time two-dimensional profile diagram;

[0053] Figure 7 is an error evaluation and model performance verification related parameter diagram;

[0054] Figure 8 is a block diagram of a refractive index structure constant adaptive prediction system based on atmospheric turbulence multi-scale characteristics in the preferred embodiment of the present application. DETAILED DESCRIPTION

[0055] The technical solutions of the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0056] As shown in Figures 1-5 , the preferred embodiment of the present application relates to a refractive index structure constant adaptive prediction method based on atmospheric turbulence multi-scale characteristics, comprising the following steps:

[0057] Step one: carry out atmospheric refractive index structure constant measurement, pre-process the measured atmospheric refractive index structure constant time-space data to represent the time-space distribution of atmospheric refractive index structure constant .

[0058] In this step, the data collection process considers the measurement equipment, measurement location, and measurement time, etc. Atmospheric refractive index structure constant The measurement can be realized by differential image motion method, scintillometer, acoustic detection radar, temperature fluctuation instrument, meteorological sounding instrument, meteorological tower, wind profile radar, pulse Doppler radar, microwave radiation instrument, atmospheric detection laser radar, etc. Preferably, the atmospheric refractive index structure constant is measured for a long period of time and at multiple heights The profile is then extracted, and the measured data is converted into a two-dimensional image in order of height and time.

[0059] Specifically, in this embodiment, the atmospheric refractive index structure constant is measured by using a TWP3 boundary layer wind profile radar The wind profile radar accurately obtains the value of the atmospheric refractive index structure constant by processing the scattering echo of electromagnetic waves. The TWP3 boundary layer wind profile radar mainly consists of a mask array antenna, a radio acoustic detection system, a transmitting and receiving system, and a high-speed signal processing system. The measurement site is located in Mianzhu City, Sichuan Province (31°22′N, 104°07′E, altitude 724 m). The measurement time is from December 3, 2020 to January 5, 2021, a total of 34 days, and the measurement height is [0.1, 0.16, …, 2.8] km. In order to characterize the spatial and temporal distribution of the atmospheric refractive index structure constant

[0060] , the atmospheric refractive index structure constant data is extracted from the data measurement file, and the measured data is converted into a two-dimensional image in order of height and time. Figure 1 The measurement data source file and the pseudo-color map after data conversion. It can be seen that the distribution of the atmospheric refractive index structure constant presents a height-time evolution pattern, such as strips, clusters, and lines.

[0061] Step two: according to the turbulent motion process and the spatial and temporal distribution characteristics of turbulence, first use a sliding window to extract local features, generate multiple input-output data pairs, and randomly divide them into a training set, a validation set, and a test set for training and validating the following model. Then, different convolution kernels are used to learn the local mixed structure representing the turbulent motion process, and a spatial and temporal feature extraction module of the atmospheric turbulent refractive index structure constant is constructed, as shown in Figure 2 , which is used to extract the spatial and temporal features of the atmospheric turbulent refractive index structure constant .

[0062] ​​​Turbulence is a complex motion, which is a dynamic cascade process from large-scale energy input vortex to small-scale dissipation vortex. In this process, the generation of large-scale vortex, energy transfer and dissipation of small-scale vortex interact with each other, forming different scale turbulent vortex, turbulent filament, thermal bubble and other distribution, which shows strip, cluster and linear pattern in space-time distribution.

[0063] This step uses a sliding window to extract local features and enrich the data set to generate multiple sets of "input-output" data pairs for training and testing. The specific process is as follows: 1. Set the sliding window width. The sliding window covers the time step history data used as input features, and the input features are X∈R m×h (where m represents the input sliding window width, i.e. the time step, and h represents the number of high sampling points), and set the output features as Y∈R n×h (where n represents the output feature window width), forming a set of "input-output" data pairs. 2. Slide the window one step to the right to cover another time step of measurement data, forming a new "input-output" data pair. 3. Repeat step 2 until all measurement data is segmented. 4. Randomly divide all "input-output" data pairs into training set, validation set and test set, the training set is used to train and fit the model designed below, the validation set is used to debug the model and select hyperparameters, and the test set is used to evaluate the model performance.

[0064] Different convolution kernels are used to learn the local complex mixed structure of turbulence. The shallow network captures small-scale, high-frequency (detail) vortex features such as fine turbulent filaments and vortex core structures, which correspond to the scale at which turbulent kinetic energy is finally dissipated into heat energy. The deep network (marked in Figure 2 ) has a wider receptive field and can learn to represent large-scale, low-frequency vortex features in turbulence, such as large-scale turbulent vortexes and thermal bubbles, which are derived from energy input and contain the main energy. During feature extraction, the network uses pooling operations to gradually abstract spatial information, reflecting the process of information (energy) convergence from fine small-scale to coarse large-scale in the energy cascade process of turbulence. This hierarchical feature extraction mechanism enables the network to learn effective representations of complex structures such as turbulent vortexes, turbulent filaments and thermal bubbles covering different scales (large-scale generation / transmission, small-scale dissipation).

[0065] Specifically, in this embodiment, the measured data is segmented by sliding window to generate "input-output" data pairs for training and testing, and the specific steps are as follows: 1. Set the sliding window width so that the input features are X∈R 46×46 , and the data measured at the following time steps are used as output Y∈R 36×46, forming a set of "input-output" data pairs. 2. Slide the window one step to the right, covering the measurement data of another time step, forming a new set of "input-output" data pairs. 3. Repeat step 2 to segment all measurement data, a total of 9710 pairs of "input-output" data pairs are generated for the experiment. 4. Randomly divide all data pairs into training set, test set and validation set, get 6797 groups of training set, 1457 groups of test set and 1456 groups of validation set.

[0066] The spatiotemporal feature extraction module structure of the atmospheric refractive index structure constant based on the turbulence characteristics is shown in Figure 2 , using convolution kernel to learn the local complex mixed structure representing the turbulence motion process. The shallow layer uses 64 3x3 convolution kernels to learn the local, small-scale high-frequency variation of turbulence mixing structure in the time-height dimension. These layers capture the subtle and rapidly changing vortex characteristics between adjacent height layers. With the deepening of the network and the application of pooling operation, the receptive field increases. The middle layer uses 128 feature maps to integrate the local information extracted by the shallow layer and identify more complex spatial patterns. The deep layer uses 256 feature maps to learn large-scale, low-frequency turbulence vortex characteristics. The extracted feature vector R 32×256×1×1 is converted to the feature vector of R 32×256 through the global average pooling layer (where 32 is the sample batch size), and the 256-dimensional vector is mapped to a 1656-dimensional vector after the linear layer. Then the vector R 32×1656 is restored to the form of a two-dimensional tensor R 32×46×36 , so as to align with the target value and be used for loss calculation. This module can learn the typical spatiotemporal distribution patterns in the atmospheric refractive index profile, including turbulent vortices, turbulent filaments and thermal bubbles, etc.

[0067] Step three: Atmospheric turbulence is affected by historical states (such as thermal stratification, wind speed profile evolution), and has large-scale long-time correlation. According to the long-time correlation of turbulence, a three-layer stacked gating mechanism is used to learn the dependence relationship representing long time series, and the long-time prediction accuracy of the atmospheric refractive index structure constant is improved. A long-time correlation feature extraction module for the atmospheric refractive index structure constant based on the turbulence characteristics is constructed, which is used to extract the long-time correlation features of the atmospheric turbulence refractive index structure constant .

[0068] The turbulence is influenced by historical states (e.g. thermal stratification, wind speed profile evolution) and has significant long-time correlation. The gating mechanism remembers important historical turbulence state information through input gate, forget gate, output gate, etc., forgets irrelevant or weakly correlated information, and captures the hysteresis effect and periodicity of the atmospheric refractive index structure constant variation. The embodiment adopts a three-layer stacked gating mechanism to extract deeper time features. In addition, a random inactivation mechanism is added to prevent overfitting and improve generalization ability.

[0069] Specifically, in the embodiment, the long-time correlation extraction module structure of the atmospheric refractive index structure constant based on turbulence characteristics is shown in Figure 3 Each input feature X ∈ R 46×46 (behavioral height direction, column as time direction) is regarded as a multivariate time series. For an input sequence of length n (46 in this example), the input at each time step is mapped to a 128-dimensional hidden state vector through the network, and the final representation h T ∈ R B×128 of the entire sequence (where B represents the sample batch size) is obtained, and then the dimension is increased from 128 to 1656 through linear transformation. Subsequently, the vector R B×1656 output by the previous layer is restored to the form of a two-dimensional tensor R B×46×36 , representing the prediction profile of 46 height points in each sample at future 36 time steps.

[0070] In addition, this branch adopts a three-layer stacked gating structure, i.e. the output of the previous layer is taken as the input of the next layer, multiple layers can extract deeper time features, a random inactivation mechanism is added with Dropout = 0.1 to prevent overfitting and improve generalization ability. The goal of this branch is to learn the long-term evolution trend of these sequences through the gating structure, including: persistent enhancement / decay of turbulence intensity, periodic pattern, and long-term dependence.

[0071] Step four: according to the time correlation of turbulence, for the time dimension of each height, through different time step convolution kernels, the model captures the dependence of different times, and constructs a short-time correlation feature extraction module of the atmospheric turbulence refractive index structure constant , which is used to extract the short-time correlation features of the atmospheric turbulence refractive index structure constant .

[0072] ​In this step, for each layer's time dimension, multiple stacked convolutional layers are designed to process time-series data. Each layer gradually expands its receptive field by inserting holes into the convolutional kernel elements, thereby enabling parallel processing across multiple time steps, learning turbulence features at different time scales, and improving feature extraction efficiency. Furthermore, to avoid gradient vanishing and improve model training efficiency, n residual block connections are added. The output of each layer includes not only the result of the convolution calculation but also the weighted result of the input signal, making it easier to train deeper networks.

[0073] Specifically, in this embodiment, the atmospheric refractive index structure constant is constructed based on turbulence characteristics. Short-term correlation extraction module, such as Figure 4 As shown, it contains four stacked one-dimensional causal convolutional layers. Each layer progressively expands its receptive field by inserting a different number of holes into the kernel elements, with the number of holes increasing exponentially: 1, 2, 4, 8. A residual block connection structure is introduced, with n = 4 residual blocks. Finally, a linear layer maps it to R. 32×36 The same temporal correlation extraction module was used for all elevation points, ultimately yielding R0. 32×46×36 The module is designed to align with the target value and be used for loss calculation. It can learn the dependence of atmospheric refractive index over different time periods.

[0074] Step 5: Design an adaptive turbulence multi-scale feature fusion module. This module includes three sets of bidirectional cross-attention modules. The adaptive turbulence multi-scale feature fusion module aligns and fuses different turbulence features, enabling the model to obtain more accurate and consistent profile predictions based on the multi-scale characteristics of turbulence. Furthermore, the adaptive turbulence multi-scale feature fusion module introduces a gating mechanism to dynamically control and select features from the output information streams of the three sets of bidirectional cross-attention modules, simulating the variation in the importance of different scale features of turbulence under conditions (such as the intermittency of turbulence), to obtain the final predicted value.

[0075] In this step, three sets of bidirectional cross-attention modules are designed to align and fuse different turbulence feature extraction modules. The core of the bidirectional cross-attention module lies in the QKV attention mechanism, which quantifies the attention one module gives to another. Here, Q represents the query ("focus"), and K and V represent the key ("location index") and value ("actual content"), respectively. The outputs of the three modules are linearly projected to obtain a unified dimension vector. The attention weights of Q and all K are calculated to simulate the correlation between different turbulence features. The cross-attention module dynamically fuses multimodal information, enabling different feature extraction modules to enhance each other and effectively alleviating the problem of isolated modeling of single branches. By automatically determining "which information from another module the current module pays attention to," adaptive modeling across modules is achieved.

[0076] To improve the adaptive ability when fusing different attention outputs, reduce the offset or redundancy between different attention outputs, two gating mechanisms are introduced to more flexibly control the information flow, thereby finely controlling the strength and direction of multi-module information fusion, and avoiding excessive fusion or redundant interference. Specifically, first, two independent gating weights g1 and g2 are introduced to automatically evaluate the importance of the auxiliary module, and then the final output result is obtained according to the three-branch weighted fusion strategy dominated by the main module, which realizes the dynamic adaptive adjustment of the importance of different features. For example, under strong shear conditions, more attention may be paid to small-scale vortices; while under stable stratification, more attention may be paid to large-scale intermittent structures.

[0077] Specifically, in the embodiment, the adaptive turbulent multi-scale feature fusion module structure is as shown in Figure 5 The output features of the above three modules are unified to a tensor dimension X e R 32×46×36 , 32 is the batch size, 46 is the number of height sampling points, and 36 is the prediction time step. Three linear layers are used for linear projection, and the projection operation is to transform the time feature length of each height, and the transformations are Q = W q Q raw e R 32×46×d , K = W K K raw e R 32×46×d , and V = W V V raw e R 32×46×d , where Q raw , K raw , and V raw are the original query, key, and value, W q , W k , and W v are the linear projection weight matrices of the query, key, and value, Q, K, and V are the projected query, key, and value, and d is 64.

[0078] The output of the spatiotemporal feature extraction module is taken as the main branch, and the output of the long-time correlation extraction module is taken as the auxiliary branch to obtain the attention output Attn ab , which means that the weak perception ability of time dependence is compensated by means of long-time sequence trend information; the output of the long-time correlation extraction module is taken as the main branch, and the output of the spatiotemporal feature extraction module is taken as the auxiliary branch to obtain the attention output Attn ba , which means that the local spatial perception ability is fused to make up for the deficiency in spatial feature modeling, so that the time sequence output is more spatially recognizable, and the spatiotemporal coupling expression ability is improved; the output of the long-time correlation extraction module is taken as the main branch, and the output of the short-time correlation extraction module is taken as the auxiliary branch to obtain the attention output Attn bc; the output of the short-term correlation extraction module is the main branch, and the output of the long-term correlation extraction module is the auxiliary branch to obtain the attention output Attn cb , which means that the dependent information of different time scales is fused to better model complex dynamic changes; the output of the spatial-temporal feature extraction module is the main branch, and the output of the short-term correlation extraction module is the auxiliary branch to obtain the attention output Attn ca , which means that the fused image spatial features make its output more spatially different; the output of the spatial-temporal feature extraction module is the main branch, and the output of the short-term correlation extraction module is the auxiliary branch to obtain the attention output Attn ac , which means that the sequence structure perception ability is introduced to improve its understanding of the time evolution trend. The attention output calculation formula is:

[0079]

[0080] Thus, the attention output Attn ab , Attn ba , Attn bc , Attn cb , Attn ca , and Attn ac .

[0081] Through a fully connected layer and an activation function, it is uniformly mapped to the same dimension, and then through three groups of learnable gating weights Alpha1, Alpha2, and Alpha3 of bidirectional attention, the fused images are obtained:

[0082] Fusion1 = Alpha1 x Attn ab + (1-Alpha1) x Attn ba

[0083] Fusion2 = Alpha2 x Attn bc + (1-Alpha2) x Attn cb

[0084] Fusion3 = Alpha3 x Attn ca + (1-Alpha3) x Attn ac

[0085] Then two independent gating weights g 1, g2 e R 32×46×36 are introduced to automatically evaluate the importance of the auxiliary branch. The calculation formula of the gating weight is as follows:

[0086] g1 = sigma (W1 (Fusion2-Fusion1))

[0087] g2 = σ(W2(Fusion3 - Fusion1))

[0088] where the difference terms Fusion2 - Fusion1 and Fusion3 - Fusion1 depict the feature residuals of two auxiliary branches relative to the main branch, i.e., represent the complementary information between the main-auxiliary branches, and the difference features are transformed via linear transformation weight matrices W1, W2 ∈ R 46×36 Dimensional projection is performed to learn the fusion sensitivity of the corresponding position; then the result is normalized to the interval [0, 1] through a sigmoid function σ, so as to obtain an interpretable attention weight ratio. In order to realize effective integration of branch outputs, a three-branch weighted fusion strategy dominated by the main branch is introduced for the feature residuals of the auxiliary branches relative to the main branch, and the final fusion feature representation is defined as:

[0089] Fusion = (1 - g1 - g2) · Fusion1 + g1 · Fusion2 + g2 · Fusion3

[0090] where Fusion1 is taken as the main branch output, and its weight is expressed by 1 - g1 - g2; on the basis of preferentially retaining the main branch information, Fusion2 and Fusion3 are dynamically injected only when the auxiliary branches provide effective supplements, and the two are regarded as supplementary enhancement paths, and the gating network only allocates a higher fusion proportion when there is a significant difference between the two and the main branch in the semantic space. The fusion strategy effectively avoids the interference and feature cancellation problems of redundant information that may be introduced by "average enhancement", thereby improving the discrimination ability of the model to multi-source features. The dimension of the final model output tensor is: R 32×46×36 where 32 represents the batch size, 46 represents the number of sampling points in the vertical height profile, and 36 represents the future time steps. The tensor is used to represent the model's prediction of the atmospheric refractive index structure constant at each height position in the future 36 time steps Prediction results on a logarithmic scale The prediction output constitutes a height-time two-dimensional profile, which can be directly used to reproduce the vertical distribution change of the future atmospheric turbulence intensity, as shown in Figure 6 Error evaluation and model performance verification are performed based on relevant indicators (such as RMSE, MAPE, R 2 ), as shown in Figure 7 The prediction results can show the atmospheric turbulence evolution behavior in the time series dimension, providing prior reference for optical communication, remote sensing, laser weapon systems, etc.

[0091] As shown in Figure 8 The embodiment discloses a refractive index structure constant adaptive prediction system based on atmospheric turbulence multi-scale characteristics, which is used to execute the above method, and includes the following modules:

[0092] Data acquisition module: for measuring atmospheric refractive index structure constant The atmospheric refractive index structure constant is preprocessed to represent the spatial and temporal distribution of the atmospheric refractive index structure constant .

[0093] The spatial and temporal feature extraction module: based on the turbulent motion process and the spatial and temporal distribution characteristics of turbulence, the local features are extracted using a sliding window to generate multiple input-output data pairs, and then different convolution kernels are used to learn the local mixed structure representing the turbulent motion process.

[0094] The long-time correlation feature extraction module: according to the long-time correlation of turbulence, a gating mechanism is used to learn the dependence relationship representing the long-time sequence.

[0095] The short-time correlation feature extraction module: according to the short-time correlation of the local turbulence, for the time dimension of each layer height, different time step convolution kernels are used to capture the dependence relationship of different times.

[0096] The adaptive turbulence multi-scale feature fusion module: including three groups of bidirectional cross attention modules, the bidirectional cross attention modules are used to realize the alignment and fusion between the spatial and temporal feature extraction module, the long-time correlation feature extraction module and the short-time correlation feature extraction module; the gating mechanism is introduced to dynamically control and select the output information flow of the three groups of bidirectional cross attention modules, simulate the importance of different scale features of turbulence changing with conditions, and obtain the prediction value of the atmospheric turbulence refractive index structure constant. .

[0097] The other contents of the embodiment can refer to the above-mentioned method embodiment.

[0098] The above is only a preferred embodiment of the present application, and does not limit the present application in any form. Although the present application has been disclosed as above, it is not intended to limit the present application. Any skilled person in the art can make some changes or modifications to the above-mentioned technical content without departing from the scope of the technical solution of the present application, and any simple modification, equivalent change and modification of the above-mentioned embodiment according to the technical essence of the present application are still within the scope of the technical solution of the present application.

Claims

1. An adaptive prediction method for refractive index structure constant based on the multi-scale characteristics of atmospheric turbulence, characterized by: Includes the following steps: Step 1: Measure the atmospheric refractive index structure constant The atmospheric refractive index structure constant is preprocessed to characterize the atmospheric refractive index structure constant. The spatiotemporal distribution; Step 2: Based on the turbulent motion process and the spatiotemporal distribution characteristics of turbulence, a sliding window is first used to extract local features, generating multiple sets of input-output data pairs. Then, different convolutional kernels are used to learn the local mixing structure representing the turbulent motion process, constructing the atmospheric turbulent refractive index structure constant. The spatiotemporal feature extraction module is used to extract the atmospheric turbulent refractive index structure constant. The spatiotemporal characteristics; Step 3: Based on the long-term correlation of turbulence, a gating mechanism is used to learn the dependence of long-term series and construct the atmospheric turbulence refractive index structure constant. The module for extracting long-term correlation features is used to extract the refractive index structure constant of atmospheric turbulence. Long-term correlation characteristics; Step 4: Based on the short-term correlation of local turbulence, for each layer height, the dependence at different times is captured by convolution kernels at different time steps, and the atmospheric turbulent refractive index structure constant is constructed. The module for extracting short-time correlation features is used to extract the refractive index structure constant of atmospheric turbulence. The short-term correlation characteristics; Step 5: Design an adaptive turbulence multi-scale feature fusion module. This module includes three sets of bidirectional cross-attention modules. The bidirectional cross-attention modules are used to align and fuse the spatiotemporal feature extraction module, the long-term correlation feature extraction module, and the short-term correlation feature extraction module. The adaptive turbulence multi-scale feature fusion module introduces a gating mechanism to dynamically control and select features from the output information streams of the three sets of bidirectional cross-attention modules. It simulates the change in the importance of turbulence features at different scales with conditions to obtain the atmospheric turbulence refractive index structure constant. The predicted value.

2. The adaptive prediction method for refractive index structure constant based on the multi-scale characteristics of atmospheric turbulence as described in claim 1, characterized in that, Step one, the preprocessing of the atmospheric refractive index structure constant includes: Data extraction: Atmospheric refractive index structure constants at different times and altitudes. The measured values ​​were extracted; Data visualization: Convert the extracted measurements into a two-dimensional image in order of height and time.

3. The adaptive prediction method for refractive index structure constant based on the multi-scale characteristics of atmospheric turbulence as described in claim 1 or 2, characterized in that, In step two, a sliding window is used to extract local features and generate multiple sets of input-output data pairs.

4. The adaptive prediction method for refractive index structure constant based on the multi-scale characteristics of atmospheric turbulence as described in claim 3, characterized in that, In step two, the sliding window is used to extract local features as follows: 1) Set the sliding window width; the sliding window covers the historical data of the time step used as the input feature, where the input feature is X∈R. m×h Where R represents the real number field, m represents the input sliding window width, h represents the number of height sampling points, and the output feature is set to Y∈R n×h Where n represents the width of the output feature window, forming a set of input-output data pairs; 2) Slide the window one step to the right to cover the measurement data of another time step, forming a new input-output data pair; 3) Repeat step 2) until all measurement data are split; and randomly divide all input-output data pairs into training, validation and test sets.

5. The adaptive prediction method for refractive index structure constant based on the multi-scale characteristics of atmospheric turbulence as described in claim 4, characterized in that, Step three, the long-term correlation feature extraction module, extracts each input feature X∈R. 46×46 It is viewed as a multivariate time series, where rows represent the height direction and columns represent the time direction. For an input of length n, the input at each time step is mapped to a 128-dimensional hidden state vector through a Long Short-Term Memory (LSTM) network. The final representation h of the entire sequence is... T ∈R B ×128 Where B represents the batch size of the samples, and then the dimension is increased from 128 to 1656 through a linear transformation. Then, the vector R output from the previous layer is... B×1656 Restored to a two-dimensional tensor R B×46×36 The form represents the 46 elevation points in each sample over the next 36 time steps. Predicted profile.

6. The adaptive prediction method for refractive index structure constant based on the multi-scale characteristics of atmospheric turbulence as described in claim 5, characterized in that, In step four, the short-term correlation extraction module contains four stacked one-dimensional causal convolutional layers. Each layer gradually expands its receptive field by inserting different numbers of holes into the kernel elements, with the number of holes increasing exponentially: 1, 2, 4, and 8. A residual block connection structure is introduced, with n = 4 residual blocks. Finally, a linear layer maps the result to R. 32×36 The same short-term correlation feature extraction module is used for all elevation points to finally obtain R. 32×46×36 In the form of.

7. The adaptive prediction method for refractive index structure constant based on the multi-scale characteristics of atmospheric turbulence as described in claim 6, characterized in that, In step five, the adaptive turbulence multi-scale feature fusion module unifies the output features of the spatiotemporal feature extraction module, the long-term correlation feature extraction module, and the short-term correlation feature extraction module into a tensor dimension X∈R. 32×46×36 Where 32 is the batch size, 46 is the number of height sampling points, and 36 is the prediction time step; three linear layers are used for linear projection, and the projection operation transforms the time feature length of each height, with the transformations being Q = W. q Q raw ∈R 32×46×d K = W k K raw ∈R 32×46×d V = W v V raw ∈R 32×46×d , where Q raw K raw and V raw These are the original query, key, and value, respectively, W q W k and W v Let Q, K, and V be the linear projection weight matrices for the query, key, and value, respectively, and let d be 64. The attention output Attn is obtained by using the output of the spatiotemporal feature extraction module as the main branch and the output of the long-term correlation feature extraction module as the auxiliary branch. ab The main branch uses the output of the long-term correlation extraction module, and the auxiliary branch uses the output of the spatiotemporal feature extraction module to obtain the attention output Attn. ba The main branch uses the output of the long-term correlation extraction module, and the auxiliary branch uses the output of the short-term correlation extraction module to obtain the attention output Attn. bc The main branch uses the output of the short-term correlation extraction module, and the auxiliary branch uses the output of the long-term correlation extraction module to obtain the attention output Attn. cb The attention output Attn is obtained by using the output of the short-term correlation extraction module as the main branch and the output of the spatiotemporal feature extraction module as the auxiliary branch. ca The attention output Attn is obtained by using the output of the spatiotemporal feature extraction module as the main branch and the output of the short-term correlation extraction module as the auxiliary branch. ac The formula for calculating attention output is: Where α is the attention score matrix; Obtain attention output Attn ab Attn ba Attn bc Attn cb Attn ca and Attn ac ; By mapping to the same dimension using fully connected layers and activation functions, and using three sets of learnable gating weights (Alpha1, Alpha2, and Alpha3) for bidirectional attention, the resulting images are obtained after pairwise fusion. Fusion1=Alpha1×Attn ab +(1-Alpha1)×Attn ba Fusion2=Alpha2×Attn bc +(1-Alpha2)×Attn cb Fusion3=Alpha3×Attn ca +(1-Alpha3)×Attn ac Introduce two independent gating weights g 1, g2∈R 32×46×36 To assess the importance of auxiliary branches, the gating weight is calculated using the following formula: g1 = σ(W1(Fusion2-Fusion1)) g2=σ(W2(Fusion3-Fusion1)) Among them, the difference terms Fusion2-Fusion1 and Fusion3-Fusion1 respectively characterize the feature residuals of the two auxiliary branches relative to the main branch, that is, they represent the supplementary information between the main and auxiliary branches. The difference features are transformed by the linear transformation weight matrix W1,W2∈R 46×36 Intra-dimensional projection is performed; the result is normalized to the [0,1] interval using the sigmoid function σ to obtain an interpretable attention weight ratio; a three-branch weighted fusion strategy dominated by the main branch is introduced for the feature residuals of the auxiliary branch relative to the main branch, and the final fused feature representation is defined as: Fusion=(1-g1-g2)·Fusion1+g1·Fusion2+g2·Fusion3 In this architecture, Fusion1 is used as the main branch output, and its weight is expressed by 1-g1-g2. While prioritizing the preservation of information from the main branch, auxiliary branches are dynamically injected only when they provide effective supplementation. Fusion2 and Fusion3 are considered supplementary enhancement paths, and the gating network assigns them a higher fusion weight only when they differ significantly from the main branch in the semantic space. The final output tensor has the dimension R. 32×46×36 Where 32 represents the batch size, 46 represents the number of sampling points in the vertical height profile, and 36 represents the future time step; this tensor is used to represent the atmospheric refractive index structure constant at each height position within the next 36 time steps. Prediction results on a logarithmic scale 8. An adaptive prediction system for refractive index structure constant based on the multi-scale characteristics of atmospheric turbulence, used to perform the method as described in any one of claims 1-7, characterized in that, Includes the following modules: Data acquisition module: used to measure the atmospheric refractive index structure constant. The atmospheric refractive index structure constant is preprocessed to characterize the atmospheric refractive index structure constant. The spatiotemporal distribution; Spatiotemporal feature extraction module: Based on the turbulent motion process and the spatiotemporal distribution characteristics of turbulence, it first uses a sliding window to extract local features and generate multiple sets of input-output data pairs. Then, it uses different convolutional kernels to learn the local mixed structure that represents the turbulent motion process. Long-term correlation feature extraction module: Based on the long-term correlation of turbulence, the module uses a gating mechanism to learn and represent the dependencies of long-term series. Short-term correlation feature extraction module: Based on the short-term correlation of local turbulence, for the time dimension of each layer height, the module captures the dependencies at different times through convolution kernels at different time steps. The adaptive turbulence multi-scale feature fusion module includes three sets of bidirectional cross-attention modules. These modules are used to align and fuse spatiotemporal feature extraction, long-term correlation feature extraction, and short-term correlation feature extraction modules. A gating mechanism is introduced to dynamically control and select features from the output information streams of the three sets of bidirectional cross-attention modules, simulating the variation of the importance of turbulence features at different scales under conditions, and obtaining the atmospheric turbulence refractive index structure constant. The predicted value.

Images