CN114417742B - A laser atmospheric scintillation index prediction method and system - Google Patents

Abstract

The invention discloses a laser atmospheric scintillation index prediction method and system, comprising: constructing an atmospheric turbulence simulation model; recording a light intensity pattern after a reference beam passes through the atmospheric turbulence simulation model, and calculating the light intensity pattern after the reference beam passes through the atmospheric turbulence simulation model Flicker index; take the light intensity pattern and flicker index as training data, send them to the neural network for training, and obtain a trained neural network model; send the light intensity pattern of the unknown beam into the neural network model, and perform a calculation on the flicker index of the unknown beam. predict. The invention overcomes the defects in the prior art such as complicated calculation, long time consumption, difficulty in eliminating the influence of external noise factors, and poor robustness. The prediction is accurate, the calculation is simple, and the influence of external noise factors is avoided.

Description

A laser atmospheric scintillation index prediction method and system

technical field

The invention relates to the fields of laser engineering and optical communication, in particular to a laser atmospheric scintillation index prediction method and system.

Background technique

To study the influence of atmospheric turbulence on laser propagation, it is important to calculate the laser atmospheric scintillation index. At present, the laser atmospheric scintillation index is mainly calculated by the following methods:

(1) A random phase screen is used to simulate atmospheric turbulence. The effect of turbulence in continuous space is equivalent to the superposition of multiple equally spaced turbulent phase screens, and then the actual transmission of light beams in the atmosphere is simulated by the turbulent random phase screen model and the angular spectrum transmission formula. Calculate the scintillation index of the beam according to the following formula:

In the formula, the angle brackets represent the ensemble average, and I represents the light intensity at the detection end.

According to this method, the light intensity and scintillation index of the laser beam after passing through atmospheric turbulence can be simulated numerically.

(2) The atmospheric scintillation index is estimated by the method of projection optics. Based on the multiplicative modulation assumption of light source intensity fluctuation and atmospheric scintillation, the method measures the light intensity scintillation on two different receiving apertures at the same time through projection optics, and solves the measurement model by combining the aperture smoothing factor under weak fluctuation conditions, thereby estimating separately. Atmospheric Scintillation Index. This method needs to eliminate the effects of background light intensity and detector random noise.

The shortcomings of the above methods for calculating the laser atmospheric scintillation index are: 1. The simulation calculation is complex and time-consuming; 2. The experimental measurement is difficult to completely eliminate the influence of external factors (noise, etc.); 3. The robustness of these methods is poor, and the calculation results are only suitable for The current turbulence case, to be recalculated for other turbulence cases.

SUMMARY OF THE INVENTION

The technical problem to be solved by the present invention is that the calculation of the atmospheric scintillation index in the prior art is complicated, time-consuming, difficult to eliminate the influence of external noise factors, poor robustness, etc. The purpose is to provide a laser atmospheric scintillation index prediction method and system, Through the neural network model, the problem of quickly and accurately calculating the atmospheric scintillation index and predicting the scintillation index is solved.

The present invention is achieved through the following technical solutions:

A method for predicting a laser atmospheric scintillation index, comprising the following steps: Step S1: building an atmospheric turbulence simulation model; recording a light intensity pattern after a reference beam passes through the atmospheric turbulence simulation model, and calculating the reference beam passing through the atmospheric turbulence simulation model The flicker index after the model; Step S2: take the light intensity pattern and the flicker index as training data, send them into a neural network for training, and obtain a trained neural network model; The scintillation index of the unknown beam is predicted.

The present invention uses the existing data generated by the reference beam (including the light intensity pattern and the scintillation index corresponding to the light intensity pattern) as the training data to train the neural network to form a mature atmospheric scintillation index neural network model, and then uses the unknown light beam to train the neural network. The light intensity pattern of the beam is sent into the atmospheric scintillation index neural network model, and the scintillation index of the unknown light beam can be quickly obtained, thereby realizing the rapid prediction of the scintillation index. The invention adopts the neural network for learning and prediction, realizes the fast and simple calculation of the atmospheric scintillation index, avoids the influence of external noise factors, and has good robustness.

The core idea of the present invention is that the neural network model is adopted to realize the intelligent prediction of the flicker index. The atmospheric turbulence simulation model can be constructed according to the existing technical data.

Further, the neural network includes a convolutional neural network, and the neural network model includes a convolutional neural network model; the step S2 specifically includes: taking the light intensity pattern of the reference beam as the input of the convolutional neural network , take the flicker index of the reference beam as the output of the convolutional neural network, and train the convolutional neural network; send the light intensity pattern of the unknown beam into the convolutional neural network model to obtain the unknown beam flicker index.

The convolutional neural network is trained with the known training data, and the mapping relationship between the light intensity pattern and the flicker index is represented by the convolutional neural network model, so as to realize the prediction of the flicker index of the unknown beam.

Further, the neural network further includes a time-series neural network, and the neural network model further includes a time-series neural network model; the step S1 specifically includes: recording the reference beam after the reference beam passes through the atmospheric turbulence simulation model in a continuous time. flicker index, forming a flicker index sequence in a continuous time; the step S2 specifically includes: sending the flicker index sequence in the continuous time into the time-series neural network, and training the time-series neural network; A time-series neural network model to predict the sequence of flicker indices for future periods of light beams.

Using the known training data to train the time series neural network, the time series neural network model is used to characterize the mapping relationship between the flicker index sequences in different time periods before and after, so that the flicker index sequence in the previous period can predict the flicker in the next period. The technical effect of exponential series.

Further, the flicker index sequence in the continuous time is divided into the flicker index sequence of the previous period and the flicker index sequence of the next period according to time; the flicker index sequence of the previous period is used as the training set, and the sliding window is used. The strategy trains and revises the time-series neural network to obtain a time-series neural network model; uses the flicker index sequence of the latter period as a test set to verify the time-series neural network model.

Further, the length ratio of the flicker index sequence of the previous period and the flicker index sequence of the next period is: 8.5:1.5.

Further, in the step S1, building an atmospheric turbulence simulation model includes: using a random phase screen to simulate atmospheric turbulence, and constructing the position distribution of the random phase screen according to the atmospheric model.

表示，范围为10^-18-10^-14；将范围为10^-18-10^-14的湍流强度按数量级划分为五种湍流强度。采用神经网络模型，可以训练出参考光束在每一种湍流强度下的大气闪烁指数。大气模型模拟了五种不同数量级的湍流强度。Further, the turbulence intensity of the atmospheric turbulence is determined by

represents, the range is 10 ^-18 -10 ^-14 ; the turbulence intensity in the range of 10 ^-18 -10 ^-14 is divided into five turbulence intensities according to the order of magnitude. Using a neural network model, the atmospheric scintillation index of the reference beam at each turbulence intensity can be trained. The atmospheric model simulates turbulent intensities of five different orders of magnitude.

Further, in the step S1, record the light intensity pattern after the reference beam passes through the atmospheric turbulence simulation model, and calculate the scintillation index after the reference beam passes through the atmospheric turbulence simulation model, specifically: recording each type of turbulence The light intensity pattern after the reference beam passes through the atmospheric turbulence simulation model under the intensity, and the scintillation index after the reference beam passes through the atmospheric turbulence simulation model under each turbulence intensity is calculated.

Further, the reference beam is a simulated Gaussian beam. The simulated Gaussian beam is used, which simplifies the calculation process of the flicker index in the training data and avoids the influence of external noise factors.

The second implementation manner of the present invention is a laser atmospheric scintillation index prediction system, the laser atmospheric scintillation index prediction system includes a processor and a machine-readable storage medium, where machine-executable instructions are stored in the machine-readable storage medium , the machine-executable instructions are loaded and executed by the processor to implement the above-mentioned method for predicting the laser atmospheric scintillation index.

Compared with the prior art, the present invention has the following advantages and beneficial effects:

It overcomes the defects in the prior art such as complicated calculation, long time consumption, difficulty in eliminating the influence of external noise factors, and poor robustness. The convolutional neural network is used to realize the separate prediction of atmospheric scintillation indices in atmospheric models with different turbulence intensities. The prediction is accurate, the calculation is simple, and the influence of external noise factors is avoided; at the same time, the prediction of the atmospheric scintillation index in the future period is realized through the time series neural network. The invention provides a new method for predicting the atmospheric scintillation index applied to laser.

Description of drawings

In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings required in the embodiments will be briefly introduced below. It should be understood that the following drawings only illustrate some embodiments of the present invention, Therefore, it should not be regarded as a limitation of the scope. For those of ordinary skill in the art, other related drawings can also be obtained from these drawings without any creative effort. In the attached image:

Fig. 1 is a flowchart of a method for predicting laser atmospheric scintillation index by using convolutional neural network;

Fig. 2 is a flow chart of a method for predicting the laser atmospheric scintillation index by using a time-series neural network;

Figure 3 is a schematic diagram of the spot intensity pattern of the simulated laser after passing through five intensities of turbulence (where the numerical value represents the turbulence intensity);

Figure 4 is a schematic diagram of the results of laser atmospheric scintillation index obtained by using convolutional neural network for slowly changing atmospheric turbulence (turbulent intensity is within the same order of magnitude), including the predicted value of the scintillation index, the actual value, and the maximum and minimum value intervals;

Figure 5 is a schematic diagram of the prediction results of the laser atmospheric scintillation index obtained by using the convolutional neural network for atmospheric turbulence (turbulent intensity in different orders of magnitude, 10 ^-18 -10 ^-16 ), including the predicted value of the scintillation index, the actual value and the maximum value of the scintillation index. , the minimum value interval;

Figure 6 is a graph of the prediction results of the laser atmospheric scintillation index obtained by using the time series neural network for the slowly changing atmospheric turbulence (the turbulence intensity is within the same order of magnitude), including the predicted value and the actual value of the scintillation index;

Figure 7 is a graph of the prediction results of the laser atmospheric scintillation index sequence obtained by using the time-series neural network for atmospheric turbulence (turbulent intensity is in different orders of magnitude, 10 ^-18 -10 ^-16 ), including the predicted value of the scintillation index and the actual value;

FIG. 8 is a schematic diagram of a sequence of flicker indices trained by a time-series neural network.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and the accompanying drawings. as a limitation of the present invention.

Example 1

The present embodiment 1 is a method for predicting a laser atmospheric scintillation index, comprising the following steps:

Step S1: constructing an atmospheric turbulence simulation model, recording the light intensity pattern after the reference beam passes through the atmospheric turbulence simulation model, and calculating the scintillation index after the reference beam passes through the atmospheric turbulence simulation model;

Step S2: The light intensity pattern and the flicker index are used as training data, and sent to the neural network for training to obtain a trained neural network model; the trained neural network model is used to predict the flicker index of the unknown light beam.

In Example 1, the existing data generated by the reference beam (including the light intensity pattern and the scintillation index corresponding to the light intensity pattern) are used as training data to train the neural network to form a mature atmospheric scintillation index neural network model, and then the The light intensity pattern of the unknown beam is sent into the atmospheric scintillation index neural network model, and the scintillation index of the unknown beam can be quickly obtained, thereby realizing the rapid prediction of the scintillation index. This embodiment 1 adopts neural network for learning and prediction, realizes fast and simple calculation of atmospheric scintillation index, avoids the influence of external noise factors, and has good robustness. The core idea of this embodiment 1 is that the neural network model is adopted to realize the intelligent prediction of the flicker index. The atmospheric turbulence simulation model can be constructed according to the existing technical data.

In a possible embodiment, the neural network includes a convolutional neural network, and the neural network model includes a convolutional neural network model; step S2 specifically includes: taking the light intensity pattern of the reference beam as an input of the convolutional neural network, and using the reference beam The scintillation index is used as the output of the convolutional neural network to train the convolutional neural network; the light intensity pattern of the unknown beam is sent into the convolutional neural network model to obtain the scintillation index of the unknown beam. The convolutional neural network is trained by the known training data, and the mapping relationship between the light intensity pattern and the flicker index is represented by the convolutional neural network model, so as to realize the prediction of the flicker index of the unknown beam.

In a possible embodiment, the neural network further includes a time-series neural network, and the neural network model further includes a time-series neural network model; step S1 specifically includes: recording the scintillation index after the reference beam passes through the atmospheric turbulence simulation model in continuous time, forming The flicker index sequence in the continuous time; step S2 specifically includes: sending the flicker index sequence in the continuous time into the time-series neural network to train the time-series neural network; predict.

The flicker index sequence in the continuous time is divided into the flicker index sequence of the previous period and the flicker index sequence of the next period according to time; the flicker index sequence of the previous period is used as the training set, and the time series neural network is analyzed according to the sliding window strategy. Perform training and correction to obtain a time-series neural network model; and then use the flicker index sequence in the later period as a test set to verify the time-series neural network model.

Each sliding window is a sequence of flicker index windows of equal length. The last flicker index of the sliding window sequence is used as the output of the time-series neural network, and the other flicker index sequences except the last flicker index in the window sequence are used as the time-series neural network. The input of the temporal neural network is trained and revised. That is to say, during the training of the time series neural network, only the last flicker index can be used as the output in a sliding window, and other flicker index sequences in the sliding window can be used as input.

In a possible embodiment, the temporal length ratio of the flicker index sequence in the previous period (ie the training set) and the flicker index sequence in the subsequent period (ie the test set) is 8.5:1.5.

For example: a total of 1000 flicker indices are divided into training sequence and test sequence according to the ratio of 8.5:1.5 before training, that is, the first 850 flicker indices are used for training, and the last 150 flicker indices are used for testing. The length ratio of the flicker index sequence of the previous period and the flicker index sequence of the following period. During training, use these 850 flicker index data, as shown in Figure 8, take 10 lengths each time (the first solid line box in the figure), take the first 9 as input, and the 10th (gray dot) as the input. output. Then move the solid line box one step back (it becomes a short horizontal line box in the figure), select the next 10 lengths, and use the first 9 as input and the 10th as output. Then push the dash box one step back, and so on. After the training, the first 9 of the 150 flicker index test data are used as the input of the time series neural network. The mapping relationship is learned in the exponential data, which can predict the value of a continuous period of time in the future. The reliability of the temporal neural network model is verified with 150 flicker index test data.

In this example, the training set sliding window contains 10 flicker indices, the first 9 are used as input and the last 1 is output. In different application scenarios, the sliding window contains N flicker indices, the Nth flicker index is used as the output of neural network training, and the sequence formed by the first to N-1th flicker indices is used as the input of time series neural network training.

So, even if each secondary sequence has 100 flicker indices, there are 99 as input and the 100th as output, and then the next secondary sequence is the 2nd to 101st flicker indices, "2~100" input, "101 " output. One flicker index value is predicted at a time.

In this embodiment, the length ratio of the flicker index sequence in the previous period and the flicker index sequence in the next period is preferably 8.5:1.5, and generally does not exceed 8:2.

Using the trained and corrected time series neural network model, the flicker index sequence of the beam in the future period is predicted. The time-series convolutional neural network is trained with the known training data, and the time-series convolutional neural network model is used to characterize the mapping relationship between the flicker indices in different consecutive time periods before and after, so as to realize the prediction of the flicker index sequence in the previous period. The technical effect of a sequence of flickering indices over a period of time.

In a possible embodiment, in step S1, building an atmospheric turbulence simulation model includes: using a random phase screen to simulate atmospheric turbulence, and constructing a position distribution of the random phase screen according to the atmospheric model. The atmospheric model here adopts the existing atmospheric model in the prior art.

表示，范围为10^-18-10^-14；将范围为10^-18-10^-14的湍流强度按数量级划分为五种湍流强度。采用神经网络模型，可以训练出参考光束在每一种湍流强度下的大气闪烁指数。大气模型模拟了五种不同数量级的湍流强度。The turbulence intensity of atmospheric turbulence is

represents, the range is 10 ^-18 -10 ^-14 ; the turbulence intensity in the range of 10 ^-18 -10 ^-14 is divided into five turbulence intensities according to the order of magnitude. Using a neural network model, the atmospheric scintillation index of the reference beam at each turbulence intensity can be trained. The atmospheric model simulates turbulent intensities of five different orders of magnitude.

In a possible embodiment, step S1 includes: recording the light intensity pattern of the reference beam after passing through the atmospheric turbulence simulation model under each turbulence intensity, and calculating after the reference beam passing through the atmospheric turbulence simulation model under each turbulence intensity flicker index. Realize the prediction of atmospheric scintillation index under different turbulence intensities.

的计算式如下：In this embodiment 1, the reference beam is a simulated Gaussian beam. The simulated Gaussian beam is used, which simplifies the calculation process of the flicker index in the training data and avoids the influence of external noise factors. In addition, the flicker index

The calculation formula is as follows:

Among them, the angle brackets represent the ensemble average, and I represents the light intensity received by the detection end.

Example 2

The present embodiment 2 is a laser atmospheric scintillation index prediction method based on the embodiment 1, comprising the following steps:

Step A: In Example 2, a random phase screen is used to simulate atmospheric turbulence, and the transmission of light beams in the atmosphere is equivalent to "free space transmission + phase screen", and the distribution of the phase screen is constructed according to the atmospheric model.

1. Simulation process of random phase screen:

中；(1) First, put the random complex numbers that obey the standard normal distribution into the N×N matrix

middle;

(2) It is then filtered with a random phase spectrum.

。其中：

表示随机相位谱，

表示N×N矩阵

，which is:

. in:

represents the random phase spectrum,

Represents an N×N matrix

,

为相位屏间距；

is the phase screen spacing;

，

,

是折射率结构常数；

，

一般取2mm，k是空间频率大小，

，L ₀一般取10m；

。

is the refractive index structure constant;

,

Generally take 2mm, k is the size of the spatial frequency,

, L ₀ is generally taken as 10m;

.

(3) Perform inverse Fourier transform on the result obtained above, and take the real part as the random phase loaded by the random phase screen.

2. Simulate the transmission process of the beam in the atmosphere:

According to the above simulation process of the phase screen, the change of the light field caused by the p-th phase screen can be described as:

。其中，

和

分别为相位屏前后的场。

为厚度可忽略不计的相位屏所引起的随机相位。

. in,

and

are the fields before and after the phase screen, respectively.

Random phase caused by a phase screen of negligible thickness.

在x-y平面上的二维傅里叶变换得到

；(1) Calculation

The 2D Fourier transform in the xy plane gives

;

得到

；(2) Multiply the result obtained by

get

;

；(3) After inverse Fourier transform, the light field is obtained as

;

(4) The light intensity can be obtained by using the obtained light field, that is, the light intensity pattern of the light beam after passing through the turbulent flow can be obtained.

Step B: record the light intensity pattern of the reference beam after passing through the turbulence and calculate the scintillation index; simulate the light intensity pattern of the Gaussian beam after passing through the turbulent flow and record it; calculate the scintillation index of the beam under each turbulence intensity. The schematic diagram of the spot intensity pattern after the simulated Gaussian light passes through the turbulent flow with five intensities is shown in Figure 3, where the value range of 10 ^-18 -10 ^-14 represents the turbulence intensity.

为10^-18-10^-14范围，分为五个数量级湍流强度。对于每种湍流强度，按照步骤A仿真高斯光束经过随机相位屏后的光强图样并记录；根据光强图样计算每一种湍流强度下光束的闪烁指数。根据以下公式计算闪烁指数：Turbulence intensity

For the 10 ^-18 -10 ^-14 range, the turbulence intensity is divided into five orders of magnitude. For each turbulence intensity, simulate and record the light intensity pattern of the Gaussian beam after passing through the random phase screen according to step A; calculate the scintillation index of the beam under each turbulence intensity according to the light intensity pattern. Calculate the flicker index according to the following formula:

（1）

(1)

where the angle brackets represent the ensemble mean, I the total intensity of the light intensity pattern, and N the number of light intensity patterns used to calculate the ensemble average.

Step C: Neural network model establishment; Convolutional neural network is used in this embodiment 2 to predict the scintillation index of light beams passing through a certain range of turbulence intensity (any sub-range or all of ^10-18-10-14 ); time- ^series neural network is also used. The network learns the scintillation index of beams passing through turbulent flow in the past and predicts the scintillation index of beams passing through turbulence in the future. Specifically, according to the previous flicker index sequence (length 1684), the next flicker index sequence of length 286 is predicted. The ratio of the sequence length of this flicker index to the predicted sequence length is 8.5:1.5, but it is better not to exceed 8 :2.

In this embodiment, two methods are used to predict the flicker index. One is to use a convolutional neural network to predict the scintillation index of beams passing through turbulent flow (in any sub-range of ^{intensity 10-18-10-14} ). Using the light intensity pattern and flicker index data pair obtained in step B to form a data set, the light intensity pattern is the network input, and the flicker index is the training label. The training convolutional network can predict the flicker index according to the light intensity pattern. The flow chart of the laser atmospheric scintillation index prediction method using convolutional neural network is shown in Figure 1. The second is to use the time series neural network to learn the trend of the flicker index sequence collected and calculated in the continuous time of the light beam, and predict the light beam flicker index sequence for a period of time in the future (the ratio of the learned sequence length to the predicted sequence length is 8.5:1.5, the ratio can be adjusted , but preferably no more than 8:2). The flow chart of the laser atmospheric scintillation index prediction method using time series neural network is shown in Figure 2.

Step D: network training; use the light intensity pattern and flicker index recorded by simulation to train the convolutional neural network, use the Adam optimizer, set the learning rate to 10 ^-4 , calculate the first-order moment estimation and the exponential decay factor setting of the second-order moment estimation are 0.5 and 0.999, and the loss function is set to mae; the flicker index recorded by the simulation is used to train the sequential neural network, and the Adam optimizer is used.

The convolutional neural network and the time-series neural network are trained using the light intensity pattern and flicker index obtained from the simulation experiments. When training a convolutional neural network, its objective function is:

表示卷积神经网络，

为网络初始化参数，

为训练得到的最优网络参数；

代表光强图，

为真实闪烁指数。卷积网络采用Adam梯度下降优化算法，旨在降低预测闪烁指数与真实闪烁指数之间的误差。训练期间，学习率设置为10^-4，损失函数设置为平均绝对误差：

。in,

represents a convolutional neural network,

Initialize parameters for the network,

are the optimal network parameters obtained by training;

represents the light intensity map,

is the real flicker index. The convolutional network adopts the Adam gradient descent optimization algorithm, which aims to reduce the error between the predicted flicker index and the real flicker index. During training, the learning rate is set to ^10-4 and the loss function is set to mean absolute error:

.

When training the time-series neural network, the flicker index sequence recorded by the simulation is used for training. The flicker index sequence was divided into training set and test set according to the ratio of 8.5:1.5. From the training sequence, each time a sequence window of length 10 is taken out, the first 9 values of the sequence are used as the network input, and the 10th value is used as the training label. The window is moved by 1 step each time, and the input data is used to train the network. When training the time series network, the Adam gradient descent optimization algorithm is also used, and the loss function is set to the mean absolute error.

Step E: Network test; use the light intensity pattern not seen by the network during training as input to test the convolutional neural network; use the test data of a training window length to input the time series neural network, and the network predicts and outputs the current turbulence intensity trend. The number of beam flicker indices.

Feed the trained network with the test set data to get the predicted flicker index. For the convolutional neural network, the light intensity pattern of the test set is used as the input to predict its flicker index; for the time series neural network, the test sequence is input to the time series neural network, and the network predicts and outputs the beam flicker index in the future under the current turbulent intensity trend.

This embodiment 2 overcomes the defects in the prior art, such as complicated calculation, long time consumption, difficulty in eliminating the influence of external noise factors, and poor robustness, and provides a new method for predicting the atmospheric scintillation index applied to laser.

Example 3

The third embodiment is based on the second embodiment,

1. Use MATLAB to numerically simulate the light intensity and scintillation index of the reference beam after passing through atmospheric turbulence. Simulate the effect of atmospheric turbulence on the beam using a random phase screen. Specifically: the light beam is first transmitted through a certain distance of free space, and then a phase plate is added, and the distribution of the phase plate is constructed according to the atmospheric model; then it is transmitted through a certain distance of free space, and a phase plate is added; a certain number of cycles are carried out. Similar structure transmission.

表示，范围为10^-18-10^-14。其中，每个数量级内的湍流强度均按照2000等分，一个数量级内模拟光束经过2000种强度湍流。为了更好地计算闪烁指数，相同条件下（湍流强度不变），记录光束传输500次的光强平均值作为系综平均值，即为公式（1）中的<I>，代入公式，计算闪烁指数。最终每仿真一个数量级湍流强度光束传输，计算得到2000个闪烁指数。同时，选取500次传输中的一张光强图放入数据集，最终每仿真一个数量级湍流强度光束传输，得到2000张光强图样。这样采集的光强图数据集和闪烁指数数据将用于训练卷积神经网络和时序神经网络。2. The light intensity map of the simulated Gaussian beam after passing through different intensities of atmospheric turbulence, the turbulence intensity is calculated by

means, the range is 10 ^-18 -10 ^-14 . Among them, the turbulence intensity in each order of magnitude is divided by 2000, and the simulated beam passes through 2000 kinds of intensity turbulence in one order of magnitude. In order to better calculate the scintillation index, under the same conditions (the turbulent intensity remains unchanged), record the average value of the light intensity of the beam transmitted 500 times as the ensemble average value, which is <I> in the formula (1), and substitute it into the formula to calculate flicker index. Finally, 2000 scintillation indices are calculated for each order of magnitude turbulent intensity beam transmission is simulated. At the same time, a light intensity map in the 500 transmissions was selected and put into the data set, and finally 2000 light intensity patterns were obtained for each order of magnitude turbulent intensity beam transmission simulated. The light intensity map dataset and flicker index data thus collected will be used to train convolutional neural networks and time-series neural networks.

3. Preparations before training the convolutional neural network: The collected light intensity map and flicker index form a data pair. Calculate the mean, maximum and minimum values of the flicker indices of the 15 light intensity maps before and after each light intensity map as the flicker index of the current light intensity map, as well as its maximum and minimum values. The maximum value can be used to evaluate the network prediction results. The convolutional neural network is trained, and the light intensity map and its corresponding flicker index are extracted from the data set at equal intervals (such as every 150) to form the test set, and the rest of the data form the training set. For the temporal neural network, the flicker index dataset (that is, the flicker index sequence) is divided into training set and test set according to the ratio of 8.5:1.5.

4. Establish convolutional neural network and sequential neural network respectively. For the convolutional neural network, the input is a 64×64 light intensity map, the optimization algorithm adopts Adam, the number of training iterations is 100, the trained model parameters are saved, and then the light intensity map in the test set is input into the trained model, Obtain the predicted flicker index; for the time series neural network, the input is a flicker index sequence with a specified length (the length is determined by the amount of data, such as 2000 flicker indices, the input window length can be 10), the optimization algorithm uses Adam, and the number of training iterations is 2 Second, save the trained model parameters, and then input the test sequence into the trained model to predict the trend and value of the flicker index sequence compared to 15% of the length of the training sequence.

5. Use the test set to test the trained model. As shown in Figure 4, the result is the scintillation index predicted by the convolutional neural network in the range of ^10-17 turbulence intensity; its network test result curve includes the predicted value of the scintillation index, the actual value, and the maximum and minimum value intervals; its corresponding The error rate table of the network test results is as follows:

Actual flicker index Predicted flicker index Relative error rate 1.72×10^-5 1.54×10^-5 -10.47% 2.64×10^-5 2.93×10^-5 10.98% 3.51×10^-5 3.38×10^-5 -3.70% 4.40×10^-5 4.88×10^-5 10.91% 5.44×10^-5 6.66×10^-5 22.43% 6.16×10^-5 6.59×10^-5 6.98% 7.38×10^-5 8.28×10^-5 12.20% 8.31×10^-5 7.32×10^-5 -11.91% 8.99×10^-5 7.96×10^-5 -11.46% 9.86×10^-5 9.22×10^-5 -6.49%

It can be seen that using the trained convolutional network to deal with the slowly changing atmospheric turbulence, a good scintillation index prediction result can be obtained. Among them, the error rate of 90% is less than 30%.

As shown in Figure 5, the convolutional network has higher prediction accuracy for larger flicker indices, while lower prediction accuracy for smaller flicker indices, that is, bias occurs. A schematic diagram of the prediction results of the laser atmospheric scintillation index obtained by using the convolutional neural network for atmospheric turbulence with large changes (turbulence intensity is in different orders of magnitude, 10 ^-18 -10 ^-16 ), as shown in Figure 5, which corresponds to the error of the network test results The rate table is as follows:

Actual flicker index Predicted flicker index Relative error rate 1.72×10^-5 1.74×10^-5 1.41% 2.64×10^-5 3.64×10-5 38.08% 3.51×10^-5 3.63×10-5 3.45% 4.40×10^-5 4.98×10-5 13.09% 5.44×10^-5 5.56×10^-5 2.15% 6.16×10^-5 5.19×10^-5 -15.77% 7.38×10^-5 6.12×10^-5 -17.03% 8.31×10^-5 7.52×10^-5 -9.51% 8.99×10^-5 8.01×10^-5 -10.90% 9.86×10^-5 8.32×10^-5 -15.64% 1.70×10^-4 1.48×10^-4 -12.94% 2.57×10^-4 2.18×10^-4 -15.18% 3.59×10^-4 2.40×10^-4 -33.15% 4.54×10^-4 3.04×10^-4 -33.04% 5.37×10^-4 4.61×10^-4 -14.15% 6.38×10^-4 6.22×10^-4 -2.51% 7.39×10^-4 6.45×10^-4 -12.72% 8.35×10^-4 7.48×10^-4 -10.42% 8.96×10^-4 7.61×10^-4 -15.07% 1.01×10^-3 8.93×10^-4 -11.58% 1.70×10^-3 2.10×10^-3 23.53% 2.56×10^-3 3.22×10^-3 25.78% 3.59×10^-3 3.19×10^-3 -11.14% 4.54×10^-3 4.20×10^-3 -7.49% 5.37×10^-3 5.74×10^-3 6.89% 6.38×10^-3 7.37×10^-3 15.52% 7.36×10^-3 7.25×10^-3 -1.49% 8.32×10^-3 8.10×10^-3 -2.64% 8.92×10^-3 8.15×10^-3 -8.63% 1.01×10^-2 1.04×10^-2 2.97%

As shown in Figure 6-7, the time series network makes up for the bias defect of the convolutional neural network, and more accurately predicts the change trend of the flicker index with the change of turbulence. Figure 6 is a graph of the prediction results of the laser atmospheric scintillation index obtained by using the time-series neural network for the slowly changing atmospheric turbulence (turbulent intensity is within the same order of magnitude), including the predicted value of the scintillation index and the actual value; the error rate table of some network test results as follows:

true flicker index Predicted flicker index Relative error rate 1.10×10^-2 1.28×10^-2 16.36% 1.10×10^-2 1.21×10^-2 10.00% 1.96×10^-2 1.70×10^-2 -13.27% 1.32×10^-2 1.35×10^-2 2.27% 1.46×10^-2 9.96×10^-3 -31.78% 1.74×10^-2 1.41×10^-2 -18.97% 2.70×10^-2 2.11×10^-2 -21.85% 3.00×10^-2 2.51×10^-2 -16.33% 3.08×10^-2 2.91×10^-2 -5.52% 3.75×10^-2 3.42×10^-2 -8.80% 4.01×10^-2 3.88×10^-2 -3.24% 3.93×10^-2 3.78×10^-2 -3.82% 3.96×10^-2 4.10×10^-2 3.54% 3.44×10^-2 3.62×10^-2 5.23% 3.15×10^-2 3.05×10^-2 -3.17%

Figure 7 is a graph of the prediction results of the laser atmospheric scintillation index sequence obtained by using the time-series neural network for atmospheric turbulence (turbulent intensity in different orders of magnitude, 10 ^-18 -10 ^-16 ), including the predicted value of the scintillation index and the actual value; The error rate table of some network test results is as follows:

true flicker index Predicted flicker index Relative Error Rate Table 1.39×10-2 1.58×10-2 13.67% 6.65×10-3 7.98×10-3 20.00% 1.30×10-2 1.55×10-2 19.38% 7.43×10-3 5.25×10-3 -29.34% 1.03×10-2 7.20×10-3 -30.10% 1.05×10-2 1.29×10-2 22.86% 7.09×10-3 6.87×10-3 -3.10% 1.22×10-2 1.50×10-2 23.28% 1.51×10-2 1.29×10-2 -14.57% 1.12×10-2 1.29×10-2 15.18% 1.48×10-2 1.42×10-2 -4.05% 1.60×10-2 1.87×10-2 16.63% 1.21×10-2 1.40×10-2 15.70% 8.45×10-3 1.04×10-2 23.08% 9.23×10-3 1.20×10-2 30.01%

6. The software and hardware equipment used for training are: 4 GeForce2080 graphics cards, Ubuntu18.04.3 operating system, Python3.6 programming language, TensorFlow1.8.0, keras2.1.6 deep learning framework, and Vscode compilation environment.

The specific embodiments described above further describe the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above descriptions are only specific embodiments of the present invention, and are not intended to limit the scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims (8)

1. a laser atmospheric scintillation index prediction method, is characterized in that, comprises the following steps: Step S1: build an atmospheric turbulence simulation model; record the light intensity pattern after the reference beam passes through the atmospheric turbulence simulation model, and calculate the scintillation index after the reference beam passes through the atmospheric turbulence simulation model; Step S2: take the light intensity pattern and the flicker index as training data, send them into a neural network for training, and obtain a trained neural network model; Predict the flicker index of the unknown light beam through the trained neural network model; Wherein, the neural network includes a convolutional neural network, and the neural network model includes a convolutional neural network model; The step S2 specifically includes: using the light intensity pattern of the reference beam as the input of the convolutional neural network, using the flicker index of the reference beam as the output of the convolutional neural network, and using the convolutional neural network as the output. network for training; sending the light intensity pattern of the unknown beam into the convolutional neural network model to obtain the flicker index of the unknown beam; Alternatively, the neural network further includes a time-series neural network, and the neural network model further includes a time-series neural network model; The step S1 specifically includes: recording the scintillation index after the reference beam passes through the atmospheric turbulence simulation model in a continuous time, to form a scintillation index sequence in a continuous time; The step S2 specifically includes: sending the sequence of flicker indices in the continuous time into the time-series neural network to train the time-series neural network; sequence for prediction. 2. The laser atmospheric scintillation index prediction method according to claim 1, wherein the scintillation index sequence in the continuous time is divided into the scintillation index sequence of the previous period and the scintillation index sequence of the latter period by time; Using the flicker index sequence of the previous period as a training set, the sequential neural network is trained and revised according to the sliding window strategy to obtain a sequential neural network model; The sequential neural network model is verified by taking the flicker index sequence of the latter period as a test set. 3 . The laser atmospheric scintillation index prediction method according to claim 2 , wherein the length ratio of the scintillation index sequence of the previous period to the scintillation index sequence of the latter period is: 8.5:1.5. 4 . 4. The laser atmospheric scintillation index prediction method according to claim 1, wherein in the step S1, building an atmospheric turbulence simulation model, comprising: using a random phase screen to simulate atmospheric turbulence, and constructing the position of the random phase screen according to the atmospheric model distributed. 5. The laser atmospheric scintillation index prediction method according to claim 4, wherein the turbulence intensity of the atmospheric turbulence is determined by

represents, the range is 10 ^-18 -10 ^-14 ; the turbulence intensity in the range of 10 ^-18 -10 ^-14 is divided into five turbulence intensities according to the order of magnitude. 6. The laser atmospheric scintillation index prediction method according to claim 5, wherein the step S1 comprises: recording the light intensity pattern after the reference beam passes through the atmospheric turbulence simulation model under each turbulence intensity, and calculating each The scintillation index of a reference beam after passing through the atmospheric turbulence simulation model at a turbulent intensity. 7 . The laser atmospheric scintillation index prediction method according to claim 1 , wherein the reference beam is a simulated Gaussian beam. 8 . 8. A laser atmospheric scintillation index prediction system, characterized in that the laser atmospheric scintillation index prediction system comprises a processor and a machine-readable storage medium, wherein machine-executable instructions are stored in the machine-readable storage medium, and the Machine executable instructions are loaded and executed by the processor to implement the laser atmospheric scintillation index prediction method of any of claims 1-7.

Images