An Adaptive Demodulation Method for Free-Space Optical Communication Based on Channel Prediction

Abstract

The present invention relates to an adaptive demodulation method for free-space optical communication based on channel prediction, belonging to the field of free-space optical communication. The present invention includes an atmospheric attenuation prediction mechanism, a turbulence classification mechanism, and an adaptive demodulation mechanism. The atmospheric attenuation prediction mechanism can obtain the degree of atmospheric attenuation in advance according to weather characteristic parameters, and then perform power compensation at the transmitting end. The turbulence classification mechanism determines the current channel turbulence range based on the amplitude jitter of the received optical signal and selects a suitable adaptive demodulator. The adaptive demodulation mechanism uses a neural network model, and its demodulation performance is better than that of traditional demodulators. The present invention can greatly improve the demodulation ability of the FSO communication system under unknown channel conditions, thereby improving the reliability of the system.

Description

Technical Field

[0001] The present invention belongs to the field of free space optical communication, and relates to an adaptive demodulation method for free space optical communication based on channel prediction. Background Art

[0002] With the increasing richness of broadband services and the continuous expansion of the customer scale, the traditional radio frequency (RF) communication has the problem of insufficient bandwidth and can no longer meet the people's demand for high-capacity communication. FSO is a new type of communication method that uses laser as the signal carrier and free space as the transmission medium. It has the advantages of no need for spectrum license background, low cost, high bandwidth, strong security and confidentiality, high transmission rate, small system size, light weight, low construction and maintenance costs, etc., and can well make up for the deficiencies of traditional RF communication. In addition, wireless optical communication can adapt to communication under complex geographical conditions and can meet the needs of post-disaster emergency communication, satellite communication, and military communication. Therefore, it is of great significance to the communication field.

[0003] Due to the atmospheric turbulence effect caused by the random movement of atmospheric molecules, the optical refractive index fluctuates randomly, causing light intensity scintillation, arrival angle fluctuation, phase fluctuation, spot drift, beam expansion and other turbulence effects during the laser transmission process. Turbulence reduces the coherence of the transmitted beam, affects the beam convergence and collimation effects, resulting in imperfect laser transmission and seriously affecting the power of the laser communication system. Light intensity scintillation is the most important factor affecting communication performance. It will cause the received signal intensity at the receiving end to drop below the decision threshold, increasing the bit error rate of the communication system. When the light intensity scintillation effect is strong, it will even interrupt the current communication. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide an adaptive demodulation method based on neural network for atmospheric turbulence and atmospheric attenuation in FSO communication systems.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] An adaptive demodulation method for free space optical communication based on channel prediction, comprising the following steps:

[0007] S1: Establish an FSO communication system equipped with weather feature sensors, and obtain the original data set of atmospheric attenuation. The data includes transmitted optical power, received optical power, and weather feature parameters;

[0008] S2: Construct an LSTM neural network model based on an adaptive loss function and an attention mechanism;

[0009] S3: Processing the atmospheric attenuation original data set obtained in step S1 into an atmospheric attenuation prediction data set, and training the LSTM neural network model;

[0010] S4: Establish an FSO communication system with stable received optical power and obtain the original data set for neural network adaptive demodulator and turbulence classification, including turbulence intensity, transmission sequence, and received signal;

[0011] S5: Construct a neural network demodulator model for FSO communication system based on Adam and Dropout optimization algorithms;

[0012] S6: Process the original data set obtained in step S4 into a data set for the neural network adaptive demodulator; set simulation parameters to train the model described in step S5 to obtain the neural network demodulator model under different turbulence conditions: DL a1 Demodulator, DL a2 Demodulator, DL a3 Demodulator;

[0013] S7: Applying DL to datasets with different turbulence intensities a1 Demodulator, DL a2 Demodulator, DL a3 demodulator and calculate the bit error rate; the neural network adaptive demodulator with the lowest bit error rate under a certain turbulence intensity is called the optimal demodulator under the turbulence. a1 The turbulence set corresponding to the demodulator is called DL a1 Class, DL a2 The turbulence set corresponding to the demodulator is called DL a2 Class, DL a3 The turbulence set corresponding to the demodulator is called DL a3 kind;

[0014] S8: Build a classification model based on Bi-LSTM;

[0015] S9: combining the data obtained in step S4 and the turbulence classification obtained in step S7 into a data set; setting simulation parameters to train the Bi-LSTM-based classification model to obtain a Bi-LSTM model that classifies turbulence according to the received signal;

[0016] S10: The received signal is classified using a Bi-LSTM neural network model. After obtaining the classification, the corresponding neural network demodulator is used to demodulate the received signal to obtain the actual demodulation result.

[0017] Furthermore, the FSO communication system equipped with the weather feature sensor in step S1 includes a transmitter and a receiver, with the modulation method being binary on-off keying (OOK). Communication is carried out under different weather features, and the transmitted optical power, received optical power, and weather feature parameters are recorded. The weather feature parameters include temperature, pressure, visibility, and humidity.

[0018] Furthermore, the LSTM neural network model based on the adaptive loss function and attention mechanism includes 1 attention layer, 2 LSTM layers, and 1 fully connected layer.

[0019] The input for training the LSTM neural network model is the meteorological parameters of the previous 5 time periods, and the output is the received optical power of the next 5 time periods. Among them, the training data set is 80% of the atmospheric attenuation prediction data set, and the test data set is 20% of the atmospheric attenuation prediction data set. The input is expressed as:

[0020]

[0021] y = [RSSI1 RSSI2 RSSI3 RSSI4 RSSI5] (2)

[0022] The prediction model takes x as the input and the received optical power y as the output. Among them, P represents pressure [Pa], T represents air temperature [°C], V represents visibility [m], and H represents relative humidity [%]. The received optical power is predicted based on these parameters, and y is the target value. The transmitted optical power and the received optical power are used in the following formula to calculate the atmospheric attenuation:

[0023]

[0024] where Lbs represents the atmospheric attenuation, G output power represents the output optical power, G received power represents the received optical power, and d represents the transmission distance.

[0025] Furthermore, in step S4, the FSO communication system is made to communicate under different channel conditions. According to the atmospheric attenuation prediction model obtained in step S3, the degree of atmospheric attenuation is obtained in advance and compensated by the optical amplifier at the transmitting end, so that the received optical power at the receiving end remains stable. The transmitted sequence, received light intensity, and turbulence intensity are recorded, and the signal is transmitted according to the following formula:

[0026] x(t) = A(t)cos(2πf c t), m = 1,..., M, 1 ≤ t ≤ T (4)

[0027] where t is time, A(t) is the OOK modulation signal sequence, and T is the signal period;

[0028] At the receiver, the received signal y(t) is calculated according to the following formula:

[0029] y(t) = g(t)x(t) + n(t) (5)

[0030] where t is time in seconds, g(t) is the multiplicative noise caused by atmospheric turbulence and the attenuation caused by the atmosphere, x(t) is the signal, and n(t) is the noise caused by the photodetector.

[0031] Furthermore, the neural network demodulator model of the FSO communication system based on the Adam and Dropout optimization algorithms includes 1 convolutional layer, 1 max pooling layer, 1 smoothing layer, and 2 fully connected layers;

[0032] The received optical signal is converted into an electrical signal by a photodetector. After passing through a low-pass filter and a digital-to-analog converter, the received optical signal y(t) is converted from an analog signal to a digital signal, where y = [y1, y2,... y NL T is the total sampled digital signal sequence, is the nth sampling point, n ranges from 1 to N, N is the number of sampling points in one period, and L is the number of training signal periods.

[0033] Furthermore, in step S6, the sampling points at the highest signal-to-noise ratio for each bit are selected to form a new sequence Y = [y1, y2,..., y N , and it is normalized to the [0, 1] interval according to the following formula:

[0034]

[0035] Furthermore, in step S6, a dataset for communication under different turbulence intensities is established as the original dataset; different turbulence intensities are defined as c = [c1, c2,... c n ,..., c k T , where k represents the total number of turbulences, and c n represents a certain turbulence intensity;

[0036] is the labeled training dataset, represents the dataset under the c n turbulence intensity, where L is the number of periods of the training signal; the dataset is divided into a training dataset and a test dataset, where the training dataset is 80% of the dataset, and the test dataset is 20% of the dataset.

[0037] ​​Furthermore, in step S7, the data sets under different turbulence conditions are applied respectively, and the bit error rates of the demodulators under different turbulence conditions are compared to obtain DL a1 The best applicable turbulence set c for the demodulator a =[c a1 ,c a2 ,…,c an ], called DL a1 Class, DL a2 The best applicable turbulence set c for the demodulator b =[c b1 ,c b2 ,…,c bn ], called DL a2 Class, DL a3 The best applicable turbulence set c for the demodulator c =[c c1 ,c c2 ,…,c cn ], called DL a3 Class, where n is the number of turbulence with the lowest bit error rate under this demodulator;

[0038] Determine which turbulence class the i-th point in the Y sequence belongs to and construct a classification data set;

[0039] It is a marked DL a1 Class training data set, similarly construct DL a2 Class training dataset, DL a3 Class training dataset.

[0040] Furthermore, the Bi-LSTM neural network includes two LSTM layers in opposite directions and one fully connected layer.

[0041] The beneficial effects of the present invention are as follows: Based on neural network theory, it mainly solves the problem of increased bit error rate caused by traditional demodulators not considering light intensity attenuation and light intensity flicker caused by atmospheric channels. The present invention includes an atmospheric attenuation prediction mechanism, a turbulence classification mechanism, and an adaptive demodulation mechanism. The atmospheric attenuation prediction mechanism can obtain the degree of atmospheric attenuation in advance based on weather characteristic parameters, and then perform power compensation at the transmitting end. The turbulence classification mechanism determines the current channel turbulence range based on the amplitude jitter of the received optical signal and selects an appropriate adaptive demodulator. The adaptive demodulation mechanism uses a neural network model and has better demodulation performance than traditional demodulators. The present invention can significantly improve the demodulation capability of FSO communication systems under unknown channel conditions, thereby improving system reliability.

[0042] Other advantages, objects, and features of the present invention will be set forth in part in the following description, and in part will be obvious to those skilled in the art upon examination of the following, or may be learned by practice of the present invention. The objects and other advantages of the present invention may be realized and obtained by the following description of the specification. Description of the Drawings

[0043] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be described in detail below with reference to the accompanying drawings, where:

[0044] Figure 1 is a wireless optical communication channel model;

[0045] Figure 2 is an LSTM model structure diagram based on an attention mechanism and an adaptive loss function;

[0046] Figure 3 is a graph of the fitting situation of the atmospheric attenuation prediction model;

[0047] Figure 4 is the bit error rate of a fixed optical amplifier and an adaptive optical amplifier under different atmospheric attenuations and different channel lengths;

[0048] Figure 5 is an optisystem simulation diagram of a free-space optical OOK communication system;

[0049] Figure 6 is a curve of the probability density function of the light intensity fluctuation distribution of the Gamma-Gamma turbulence model;

[0050] Figure 7 is a simulation model diagram of the bit error rate of different neural network demodulators of the present invention at different turbulences;

[0051] Figure 8 is a Bi-LSTM neural network model diagram of the present invention;

[0052] Fig. 9 is a diagram of the model training steps;

[0053] Fig.10 is a diagram of the receiving and demodulating process after the neural network demodulator model training is completed. Detailed Embodiments

[0054] The following describes the implementation manners of the present invention through specific examples. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific implementation manners. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the drawings provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner. Without conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0055] Among them, the drawings are only for illustrative purposes, showing only schematic diagrams rather than physical diagrams, and should not be construed as a limitation on the present invention; in order to better illustrate the embodiments of the present invention, some components in the drawings will be omitted, enlarged or reduced, which do not represent the sizes of actual products; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0056] In the drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components; in the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "left", "right", "front", "rear", etc. indicating the orientation or positional relationship, they are based on the orientation or positional relationship shown in the drawings. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, the terms describing the positional relationship in the drawings are only for illustrative purposes and should not be construed as a limitation on the present invention. For those of ordinary skill in the art, the specific meanings of the above terms can be understood according to specific circumstances.

[0057] An adaptive demodulation method for free space optical communication based on channel prediction includes the following steps:

[0058] First, build an FSO communication system with the modulation format of OOK. Figure 1 This is the FSO communication system model. After completing the construction of the FSO communication system, run the FSO communication system. Obtain weather parameter, transmitted optical power, and received optical power data. Construct an LSTM neural network including 1 attention layer, 2 LSTM layers, and 1 fully connected layer. The LSTM neural network is as Figure 2 shown. The input of the model is the meteorological parameters of the previous 5 time periods, and the output is the received optical power of the next 5 time periods, where the training data set is 80% of the data set, and the test data set is 20% of the data set.

[0059] Due to the complexity and difficulty of conducting free-space optical communications in the field, an indoor channel was used to simulate the actual environment. The meteorological factor research area in this embodiment is Tongnan District, Chongqing. Tongnan District is adjacent to Hechuan District and Tongliang District to the east, Dazu District to the south, Anyue County, Ziyang City, Sichuan Province, Anju District, Suining City, Sichuan Province, and Chuanshan District, Suining City, Sichuan Province to the west, Pengxi County, Suining City, Sichuan Province, and Wusheng County, Guang'an City, Sichuan Province to the north, and faces Jialing District, Nanchong City, Sichuan Province. Weather data was input into the experimental chamber to obtain received optical power data.

[0060] The kernel density estimation method is one of the most commonly used techniques in non-parametric test methods. It is widely used to estimate an unknown probability density function. It is a natural extension of the histogram, improving the discontinuity problem of the histogram and providing higher analysis accuracy. The calculation formula is as follows:

[0061]

[0062] in represents the kernel density estimate of (x, y), is the total amount of sample data in the study area, h represents bandwidth, K represents sum function, (X i ,Y i ) represents the coordinates of the i-th sample.

[0063] The core of the kernel density estimation function lies in the kernel function, which is divided into triangular kernel, quadratic kernel, etc. This embodiment uses the triangular kernel, and the formula is as follows:

[0064]

[0065] It is easy to see that weather conditions are unevenly distributed, with less severe weather conditions. This is reflected in the fact that less data is collected during severe weather conditions. However, this data is of high value, so the degree of fit in this area needs to be improved as much as possible. Using a conventional loss function will make fitting difficult in other intervals. This embodiment improves the degree of fit in this area by modifying the loss function. The loss function is modified to:

[0066]

[0067] in:

[0068]

[0069] The simulation system parameters are shown in Table 1:

[0070] Table 1

[0071]

[0072]

[0073] The prediction results are as follows Figure 3As shown. By Figure 3 It can be seen that the model has a good prediction effect. After completing the atmospheric attenuation prediction model, at the transmitting end, an optical amplifier is used to compensate the received optical power to keep the received optical power stable. During the test process, the absolute error is used to evaluate the quality of the model. The evaluation formula is the absolute error:

[0074] μ = A - A' (11)

[0075] where A is the true value and A' is the predicted value. The average absolute error is calculated to be 0.1188 dBm.

[0076] Figure 4 and Figure 5 are the BER graphs of the traditional demodulator for the fixed-gain optical amplifier and the adaptive optical amplifier under the conditions of a 2-km channel and a 3-km channel. It can be seen from the figure that using an adaptive optical amplifier at the transmitting end to compensate the optical power can significantly reduce the bit error rate.

[0077] The present invention uses Optisystem for simulation. The specific situation of the simulation model is as Figure 6 , and the specific settings are as follows: The initial signal sequence is A(t). Using a Mach-Zehnder Modulator (MZM) and a continuous-wave laser for electro-optical phase modulation, the obtained signal is

[0078] x(t) = A(t)cos(2πf c t), m = 1,..., M, 1 ≤ t ≤ T (12)

[0079] where t is the time in seconds, A(t) is the initial signal sequence, and T is the signal period. After the optical signal modulation is completed, it passes through an adaptive optical amplifier. The adaptive optical amplifier adjusts the optical power amplification factor at the transmitting end according to the attenuation degree predicted by the atmospheric attenuation prediction model to keep the received optical power at the receiving end stable. Optical amplifiers can be classified into semiconductor optical amplifiers (SOA) and optical fiber amplifiers (OFA) according to their types. Generally, semiconductor optical amplifiers are used. They have a wide operating bandwidth and a small gain amplitude. They are devices with small volume, high efficiency, and low power consumption. The noise generated by the light can be ignored after passing through a low-pass filter. The simulation gain is set to 15 dB.

[0080] After that, it passes through the channel, and the channel mainly considers the influence brought by atmospheric turbulence. The currently most commonly used atmospheric turbulence model is the Gamma-Gamma turbulence model, Figure 7This is the probability density function of the turbulence model. This model is a two-parameter model, and its probability distribution can better reflect the distribution characteristics of light intensity. Assume that the received irradiance is the product of two statistically independent random processes I x and I y which is expressed as:

[0081] I = I x I y (13)

[0082] where I x represents large-scale vortices, and I y represents small-scale vortices. Its probability density function is:

[0083]

[0084] According to the total probability formula, we get:

[0085]

[0086] where α represents the effective number of large-scale vortices in the scattering process, β represents the effective number of small-scale vortices, K n (·) is the nth-order modified Bessel function of the second kind, and I is the normalized received irradiance. Among them, α and β are related to the beam model.

[0087] The channel length is set to 3 km, and the channel attenuation is set to 10 dB / km.

[0088] After passing through the optical amplifier, the simulation gain is set to 15 dB.

[0089] After completing the construction of the FSO communication system, the received optical signal y(t) is converted from an analog signal to a digital signal: where y = [y1, y2,... y NL T is the total sampled digital signal sequence. Among them is the nth sampling point, n ranges from 1 to N, N is the number of sampling points in a period, and L is the number of training signal periods. Each bit selects the sampling point with the highest signal-to-noise ratio to form a new sequence Y = [y1, y2,…, y N , and then it is normalized to obtain the normalized sequence Y = [y1, y2,…, y N . Different turbulence intensities are defined as c = [c1, c2,... c n ,..., c k T , where k represents the total number of turbulences, and c n represents a certain turbulence intensity. is the labeled training data set, represents at c n ​​A data set under turbulent intensity, where L is the number of periods of the training signal. The data set includes a training data set and a test data set, where the training data set is 5000 bits and the test data set is all data. The optisystem system simulation parameters are shown in Table 2

[0090] Table 2

[0091]

[0092]

[0093] The parameter settings of the neural network demodulator model are shown in Table 3 as follows:

[0094] Table 3

[0095] parameter Value / Form Number of neural network layers 5 Convolutional layer 1 Max pooling layer 1 Smooth layer 1 Fully connected layer 2 Number of training samples 5000 Activation Function Tanh Loss Function MSE Optimizer Adam Evaluation parameters Accuracy

[0096] DL a1 The demodulator is trained under the atmospheric structure constant DL a2 The demodulator is trained under the atmospheric structure constant DL a3 The demodulator is trained under the atmospheric structure constant Table 4 shows the bit error rates under the same conditions using the neural network model and the traditional decision.

[0097] Table 4

[0098] Turbulence intensity Neural Network Demodulator Bit Error Rate Traditional demodulator bit error rate 0.5e-012(m^2 / 3) 0.007320 0.032714 0.5e-015(m^2 / 3) 0.000411 0.001205 0.5e-018(m^2 / 3) 0.000305 0.0005416

[0099] Then determine the bit error rates of the three demodulators at all turbulences.

[0100] The bit error rates are shown in the figure. It can be seen from the figure that in some regions, even if the selected turbulence is not the optimal neural network demodulator, its bit error rate is still lower than that of the traditional demodulator.

[0101] The simulation is for illustration only. In this embodiment, only the atmospheric structure constants of 0.5e-012 (m^2 / 3), 0.5e-013 (m^2 / 3), 0.5e-014 (m^2 / 3), 0.5e-015 (m^2 / 3), 0.5e-016 (m^2 / 3), 0.5e-017 (m^2 / 3), and 0.5e-018 (m^2 / 3) in the simulation are tested.

[0102] From Figure 7 it can be known from Table 5 that Table 5 shows the selection of the neural network demodulator for each turbulence intensity.

[0103] Table 5

[0104] Turbulence intensity Neural Network Demodulator 0.5e-012(m^2 / 3) <![CDATA[DL a1 Demodulator]]> 0.5e-013(m^2 / 3) <![CDATA[DL a1 Demodulator]]> 0.5e-014(m^2 / 3) <![CDATA[DL a2 Demodulator]]> 0.5e-015(m^2 / 3) <![CDATA[DL a2 Demodulator]]> 0.5e-016(m^2 / 3) <![CDATA[DL a2 Demodulator]]> 0.5e-017(m^2 / 3) <![CDATA[DL a3 Demodulator]]> 0.5e-018(m^2 / 3) <![CDATA[DL a3 Demodulator]]>

[0105] Take DLa1 The received optical intensity obtained from the turbulence intensity corresponding to the demodulator is sampled and normalized to obtain a sequence Y = [y1, y2, …, y N . The sequence is randomly grouped into groups of 20 as the input, and the output is the DL a1 class. And so on to obtain the dataset of the DL a2 class. Take 80% as the training set and 20% as the test set, and the total number of the test set is 10,000. a2

[0106] Construct a Bi-LSTM neural network model for classification. The Bi-LSTM structure is as Figure 8 shown.

[0107] The parameter settings of the Bi-LSTM neural network model are as follows:

[0108] Table 6

[0109]

[0110]

[0111] The confusion matrix of the classification accuracy after training the model is shown in Table 7.

[0112] Table 7

[0113] <![CDATA[DL a1 class]]> <![CDATA[DL a2 class]]> <![CDATA[DL a3 class]]> <![CDATA[DL a1 class]]> 9250 431 319 <![CDATA[DL a2 class]]> 392 9152 456 <![CDATA[DL a3 Class]]> 106 381 9513

[0114] It can be seen from the table that the classification accuracy of strong turbulence is 92.50%, the classification accuracy of medium turbulence is 91.52%, and the classification accuracy of weak turbulence is 95.13%.

[0115] The operation process of model training is as Fig. 9 shown.

[0116] The specific operation process after model training is as Fig.10 shown.

[0117] The present invention designs an adaptive demodulator based on neural network in the FSO communication system. The operation process includes obtaining the transmitted optical power, received optical power, and weather parameters under different weather conditions. Training to obtain an atmospheric attenuation prediction model to ensure the stability of the received optical power. Obtaining datasets under different turbulence intensities, training the demodulator model of the FSO communication system, and obtaining a neural network demodulator. Classifying the received signal by using a Bi-LSTM neural network model, demodulating by using the corresponding neural network demodulator after classification, and obtaining the actual demodulation result.

[0118] ​The present invention takes into account the effects of attenuation and turbulence on the system more than traditional demodulators. Compared with the optimal decision-making demodulator, it does not require complex calculations and detailed channel conditions. In terms of actual applications, neural network demodulators are generally not used as demodulators in communication systems because the channels are complex and variable, and the demodulation performance will drop significantly without channel conditions. The present invention can well solve this problem by judging the channel conditions based on the fluctuation of the sampled signals at the receiving end.

[0119] The turbulence intensity and classification method adopted by the present invention are for example. The more classifications of turbulence, the better the performance of the neural network demodulator. However, the accuracy of the Bi-LSTM model for classifying turbulence will decrease, and vice versa. It can be changed according to the actual situation.

[0120] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the purpose and scope of the present technical solution, and they should all be covered by the scope of the claims of the present invention.

Claims

1. An adaptive demodulation method for free space optical communication based on channel prediction, characterized in that: Including the following steps: S1: Establish an FSO communication system equipped with weather feature sensors, and obtain the original dataset of atmospheric attenuation. The data includes transmitted optical power, received optical power, and weather feature parameters; S2: Construct an LSTM neural network model based on an adaptive loss function and an attention mechanism; S3: Process the original dataset of atmospheric attenuation obtained in step S1 into a prediction dataset of atmospheric attenuation, and train the LSTM neural network model; S4: Establish an FSO communication system with stable received optical power, and obtain the original dataset of neural network adaptive demodulation and turbulence classification. The data includes turbulence intensity, transmitted sequence, and received signal; S5: Construct an FSO communication system neural network demodulator model based on the Adam and Dropout optimization algorithms; S6: Process the original data set obtained in step S4 into a data set for the neural network adaptive demodulator; set simulation parameters to train the model described in step S5 to obtain a neural network demodulator model under different turbulences: DL a1 Demodulator, DL a2 Demodulator, DL a3 Demodulator; S7: Apply the data sets under different turbulence intensities to DL a1 Demodulator, DL a2 Demodulator, DL a3 Demodulator and calculate the bit error rate; The neural network adaptive demodulator with the lowest bit error rate under a certain turbulence intensity is called the optimal demodulator under that turbulence, and the turbulence set corresponding to the DL a1 Demodulator is called the DL a1 class, DL a2 Demodulator is called the DL a2 class, DL a3 Demodulator is called the DL a3 class; S8: Construct a classification model based on Bi-LSTM; S9: Combine the data obtained in step S4 and the turbulence classification obtained in step S7 into a dataset; set simulation parameters to train the Bi-LSTM-based classification model to obtain a Bi-LSTM model for classifying turbulence according to the received signal; S10: Use the Bi-LSTM neural network model to classify the received signal, and after obtaining the classification, demodulate it using the corresponding neural network demodulator to obtain the actual demodulation result.

2. The adaptive demodulation method for free space optical communication based on channel prediction according to claim 1, wherein: In step S1, the FSO communication system equipped with weather feature sensors includes a transmitter and a receiver. The modulation method is binary on-off keying (OOK). Communication is carried out under different weather features, and the transmitted optical power, received optical power, and weather feature parameters are recorded; the weather feature parameters include temperature, pressure, visibility, and humidity.

3. The adaptive demodulation method for free space optical communication based on channel prediction according to claim 1, wherein: The LSTM neural network model based on the adaptive loss function and the attention mechanism includes 1 attention layer, 2 LSTM layers, and 1 fully connected layer; The input for training the LSTM neural network model is the meteorological parameters of the previous 5 time periods, and the output is the received optical power of the next 5 time periods. Among them, the training dataset is 80% of the atmospheric attenuation prediction dataset, and the test dataset is 20% of the atmospheric attenuation prediction dataset; the input is expressed as: y = [RSSI1 RSSI2 RSSI3 RSSI4 RSSI5] (2) The prediction model takes x as the input and the received optical power y as the output. Among them, P represents pressure [Pa], T represents air temperature [°C], V represents visibility [m], and H represents relative humidity [%]; the received optical power is predicted based on these parameters, and y is the target value; the transmitted optical power and the received optical power are used to calculate the atmospheric attenuation using the following formula: where Lbs represents atmospheric attenuation, and G output power represents the output optical power, and G received power represents the received optical power, and d represents the transmission distance.

4. The adaptive demodulation method for free space optical communication based on channel prediction according to claim 1, wherein: In step S4, make the FSO communication system communicate under different channel conditions. According to the atmospheric attenuation prediction model obtained in step S3, obtain the degree of atmospheric attenuation in advance and compensate it at the optical amplifier at the transmitting end to keep the received optical power at the receiving end stable; record the transmitted sequence, received light intensity, and turbulence intensity, and transmit the signal according to the following formula: x(t) = A(t)cos(2πf c t), m = 1, …, M, 1 ≤ t ≤ T (4) where t is time, A(t) is the OOK modulation signal sequence, and T is the signal period; At the receiver, calculate the received signal y(t) according to the following formula: y(t) = g(t)x(t) + n(t) (5) where t is time in seconds, g(t) is the multiplicative noise caused by atmospheric turbulence and the attenuation caused by the atmosphere, x(t) is the signal, and n(t) is the noise caused by the photodetector.

5. The adaptive demodulation method for free space optical communication based on channel prediction according to claim 1, wherein: The neural network demodulator model of the FSO communication system based on the Adam and Dropout optimization algorithms includes 1 convolutional layer, 1 max pooling layer, 1 smoothing layer, and 2 fully connected layers; The received optical signal is converted into an electrical signal by a photodetector. After passing through a low-pass filter and a digital-to-analog converter, the received optical signal y(t) is converted from an analog signal to a digital signal, where y = [y1, y2,... y NL T is the total sampled digital signal sequence, is the nth sampling point, where n ranges from 1 to N, N is the number of sampling points in one period, and L is the number of training signal periods.​ 6. The adaptive demodulation method for free space optical communication based on channel prediction according to claim 1, wherein: In step S6, sample points at the highest signal-to-noise ratio are selected for each bit to form a new sequence Y = [y1, y2,..., y N , and it is normalized to the interval [0, 1] according to the following formula:

7. The adaptive demodulation method for free space optical communication based on channel prediction according to claim 1, wherein: In step S6, a data set for communication under different turbulence intensities is established as the original data set; different turbulence intensities are defined as c = [c1, c2,... c n ,..., c k T , where k represents the total number of turbulences, and c n represents a certain turbulence intensity; is the marked training data set, represents the data set under the turbulence intensity of c n , where L is the number of periods of the training signal;​ The dataset is divided into a training dataset and a test dataset, where the training dataset is 80% of the dataset and the test dataset is 20% of the dataset.

8. The adaptive demodulation method for free space optical communication based on channel prediction according to claim 6, characterized in that: In step S7, for data sets under different turbulences, the bit error rates of each demodulator under different turbulences are compared to obtain DL a1 The optimal applicable turbulence set c of the demodulator a =[c a1 , c a2 , …, c an , which is called DL a1 class, and the optimal applicable turbulence set c of DL a2 demodulator b =[c b1 , c b2 , …, c bn , which is called DL a2 class, and the optimal applicable turbulence set c of DL a3 demodulator c =[c c1 , c c2 , …, c cn , which is called DL a3 class, where n is the number of turbulences with the lowest bit error rate under this demodulator; Determine which type of turbulence the $i$-th point in the $Y$ sequence belongs to and construct a classification data set; Is the labeled DL a1 Class training data set. Similarly, construct the DL a2 Class training data set, DL a3 Class training data set.

9. The adaptive demodulation method for free space optical communication based on channel prediction according to claim 1, characterized in that: The Bi-LSTM neural network includes 2 LSTM layers in opposite directions and 1 fully connected layer.

Images