Anti-turbulence free space optical communication method and system based on receiving and transmitting joint optimization

Abstract

The invention discloses an anti-turbulence free space optical communication method based on receiving and transmitting joint optimization, which balances the code length and the code rate through a Polar coding mode of an atmospheric turbulence channel, and ensures the coding and decoding overhead; meanwhile, a Transform model based on convolution enhancement is designed at a receiving end, learning of local features is enhanced through convolution, the relation between phase screen pixels is established, and a conjugate phase screen is given by the model for turbulence compensation. In order to solve the problem that in a free space optical transmission system, the atmospheric turbulence effect interferes with an OAM light beam, the influence of atmospheric turbulence interference on free space optical communication is relieved in a mode of joint optimization of the receiving end and the transmitting end. Error codes caused by atmospheric turbulence are corrected through Polar coding of the transmitting end, and the coupling efficiency of receiving light spots is improved after phase compensation of the receiving end, so that the mode purity obtained by corrected light spot images is improved, and the error rate of the system is reduced.

Description

Technical Field

[0001] The present invention relates to the field of optical communication technology, and more specifically, to a method and system for turbulence-resistant free-space optical communication based on joint optimization of transmission and reception. Background Art

[0002] Free Space Optical Communication (FSO) is a wireless optical communication technology that uses atmospheric channels as a signal transmission medium. Compared with traditional fiber-optic communications, free space optical communication systems do not require digging the ground or laying cables. They have the advantages of high bandwidth, anti-electromagnetic interference, and flexible deployment. In some places where wiring is difficult, such as historical buildings, protected areas, or areas with complex terrain, FSO provides an effective means of communication. Therefore, FSO is widely used in many scenarios such as metropolitan area network expansion, fiber backup, wireless cellular network backhaul, and post-disaster emergency recovery. It has the important significance of "last mile" communication and is considered to be one of the next generation of wireless connection technologies with great potential. The research on free space optical communication systems is of great significance to the construction of high-speed optical communication networks.

[0003] However, free-space optical communication links are severely affected by complex atmospheric channels. Atmospheric molecules and other molecules will absorb, scatter and cause turbulence on signal light, causing a series of interferences such as power attenuation, pulse broadening, beam expansion and arrival angle fluctuations, which ultimately reduce the average signal-to-noise ratio of the communication system and cause random fluctuations in the detection signal. Among them, the most serious impact is atmospheric turbulence. Atmospheric turbulence and bad weather will cause random fluctuations in the amplitude or phase of the signal, thereby degrading the quality of the optical signal in atmospheric transmission, increasing the communication bit error rate and reducing the transmission quality. Therefore, it is crucial to design a free-space optical transmission solution that can alleviate the degradation of optical signal transmission caused by atmospheric channels and achieve resistance to atmospheric turbulence interference. Summary of the invention

[0004] The object of the present invention is to provide an anti-turbulence free-space optical communication solution based on joint optimization of transmission and reception, so as to alleviate the interference of atmospheric turbulence on signals in free space.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is as follows: a turbulence-resistant free-space optical communication method based on joint optimization of transmission and reception, comprising:

[0006] S1. At the transmitting end, the output of the neural network model determines the construction of the code in the Polar code, analyzes the reliability of each sub-channel, and determines whether the sub-channel is a frozen bit or an information bit under the current channel according to the turbulent channel;

[0007] S2, modulate the encoded information onto an optical carrier, transmit it after collimation and beam expansion, and control the size of the Gaussian beam by an aperture. Place a spiral phase plate at the rear end to convert the Gaussian beam into an OAM beam with better anti-turbulence performance;

[0008] S3, using a rotating turbulence phase plate to simulate the atmospheric turbulence channel, the OAM beam is disturbed by the atmospheric turbulence;

[0009] S4. At the receiving end, a beam splitter is used to split the received light into two beams. One beam of light is irradiated into an infrared camera, and the infrared camera is used to collect images and input into the neural network. The neural network analyzes the phase of the turbulent channel from the light intensity image and gives the conjugate phase. The SLM displays the compensated phase image.

[0010] S5, the other light split by the beam splitter, after the conjugate phase compensation on the SLM, uses the translation stage to couple the spatial light into the optical fiber and then enters the photodetector to complete the photoelectric conversion;

[0011] S6. Finally, the oscilloscope collects the data input by the PD, saves it, and performs offline processing by MATLAB to analyze the bit error rate.

[0012] Preferably, the code construction process of the Polar coding is to sort the reliabilities of the N sub-channels generated after the channel polarization operation from large to small, and then select the first K sub-channels as information bits, and the remaining NK sub-channels as frozen bits.

[0013] Preferably, the Polar encoding process includes:

[0014] Use simulation to collect subchannel status information, including the bit error rate of the received signal corresponding to different turbulence intensities as the channel transmission characteristics, and select the Bhattacharyya parameters and subchannel index of each subchannel after channel polarization as the characteristics of the network model for training;

[0015] Construct a data set, use the known Polar coding construction method as a benchmark, mark each subchannel as an information bit or a frozen bit, and use this as a training label;

[0016] A network model is constructed to accept turbulence intensity and coding parameters as input and output the reliability prediction of each subchannel, representing the polarization weight of each subchannel as a probability that can be learned by a neural network to represent the subchannel as an information bit.

[0017] Define the loss function to measure the error between the feedback result and the true value.

[0018] Preferably, the network model includes an input layer for receiving feature vectors of sub-channels, such as bit error rate and coding parameters; a plurality of fully connected layers for extracting features and learning nonlinear mappings; an LSTM layer for processing channel state changes; and an output layer for outputting the probability or reliability score of each sub-channel as an information bit.

[0019] Preferably, when defining the loss function, a binary cross entropy loss function is used as a basic loss to determine whether the current subchannel is a frozen bit or an information bit:

[0020]

[0021] Where N is the number of samples, y i is the actual reliability label of the subchannel, is the probability predicted by the model. In order to consider the bitrate constraint and penalize the deviation between the model prediction and the target bitrate, the bitrate loss shown in the following formula is added:

[0022]

[0023] In the formula, R target is the target bitrate, usually set to 0.5, λ R The weight used to balance the bit rate loss. In order to constrain the code length, we add the following formula to the loss function, and the code length is set to 2048:

[0024] L Length =λ L N

[0025] In order to quantify the change of channel conditions, the signal-to-noise ratio is used as a quantitative indicator of channel quality, and the bit error rate is used as a measure of the performance difference before and after the channel condition changes. The loss term under the turbulent channel is shown as follows:

[0026] L channel =λ ch f(ΔBER)

[0027] Where f is a non-negative penalty factor that is proportional to the magnitude of the change in channel conditions, ΔBER is the change in bit error rate after the channel conditions change, and λ is ch is a hyperparameter for balancing loss weights;

[0028] Combining the above, we can form the final loss function as shown below:

[0029] L total =L BCE +λ R L RATE +λ L L Length +λ ch Lchannel .

[0030] Preferably, the neural network at the receiving end determines the turbulence intensity of the current channel based on the light intensity image and feeds back to the network model at the transmitting end. After identifying the turbulence intensity of the received light intensity image, the trained model is continued to be used to evaluate the new sub-channel set to obtain the reliability prediction of each sub-channel. The output of the model is used as the input of the Polar coding construction to determine the final information bits and frozen bits.

[0031] Preferably, the receiving end uses a convolution-enhanced Transformer model to learn the mapping relationship between the received light intensity image and the turbulence phase screen.

[0032] Preferably, the convolution-enhanced Transformer model uses convolution layers in the encoder to gradually reduce the spatial dimensions of the feature map, then uses upsampling and convolution layers in the decoder to gradually restore the spatial dimensions of the feature map, and uses Transformer modules in the encoder and decoder to capture the dependencies between pixels with longer distances in the feature map. The model combines the encoder, the decoder and the final convolution layer to generate a prediction result;

[0033] At the same time, in the final output layer, the turbulence intensity of the current channel is fed back to the neural network at the transmitter. The output of the network includes the turbulence intensity fed back to the transmitter and the current turbulence phase screen. Based on the model architecture of the convolution-enhanced Transformer, the convolution module is combined in the architecture.

[0034] The anti-turbulence free-space optical communication system based on the joint optimization of transmission and reception includes a receiving end neural network model, a transmitting end network model, a polar encoder, a waveform generator, an intensity modulator, a laser, a collimating beam expander, a spiral phase plate, a turbulent phase plate, a beam splitter, a photodetector, an oscilloscope, an infrared camera and a spatial light modulator;

[0035] The transmitting end designs a neural network model for a turbulent channel, outputs the encoded signal, first sends a training sequence, and allows the receiving end neural network model to receive the light intensity image and then judge the turbulence intensity of the current atmospheric channel and feed it back to the transmitting end network model. The Polar encoder gives the distribution of the Polar coded information bits under the current channel conditions. The encoded information is input into an arbitrary waveform generator according to the output of the model. The RF signal output by the waveform generator is loaded onto the intensity modulator. The laser outputs a 1550nm wavelength laser and inputs it into the intensity modulator. The bias controller is used to make the intensity modulator work in the linear region. The encoded information is modulated onto the optical carrier. The output of the intensity modulator passes through a collimating beam expander and then is converted into an OAM beam with stronger turbulence resistance through a spiral phase plate.

[0036] In the channel part, the turbulent phase plate is rotated to simulate the interference of the OAM light beam in the atmospheric turbulence channel. At the receiving end, the light beam enters the 10:90 beam splitter, of which 10% of the light beam is displayed by the infrared camera and input into the network; the other part is phase compensated by the spatial light modulator. The phase diagram displayed on the spatial light modulator is given by the output of the neural network model at the receiving end. Finally, we use the translation stage to couple the compensated spatial light into the optical fiber and connect it to the photodetector and then the oscilloscope. After recording the data, we perform offline processing to analyze the bit error rate.

[0037] Preferably, the receiving end neural network model is a CNN_Transformer model.

[0038] Compared with the prior art, the advantages of the present invention are:

[0039] The present invention aims at the problem of atmospheric turbulence effect interfering with OAM light beams in free-space optical transmission systems. The present invention uses a joint optimization method at both ends of the transmitter and receiver to alleviate the impact of atmospheric turbulence interference on free-space optical communications. Polar coding at the transmitter corrects the bit errors caused by atmospheric turbulence, and phase compensation at the receiver improves the coupling efficiency of the received light spot, so that the mode purity of the corrected light spot image is improved and the system bit error rate is reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0041] Figure 1 It is a flow chart of the anti-turbulence free-space optical communication method based on the joint optimization of transmission and reception of the present invention;

[0042] Figure 2 It is the Polar encoding flow chart of the present invention;

[0043] Figure 3 This is the architecture diagram of the convolution-enhanced Transformer model of the present invention;

[0044] Figure 4 It is the architecture diagram of the MBConv module of the present invention;

[0045] Figure 5 It is an architecture diagram of the turbulence-resistant free-space optical communication system based on the joint optimization of transmission and reception of the present invention. DETAILED DESCRIPTION

[0046] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more definite definition of the protection scope of the present invention.

[0047] The present invention provides an anti-turbulence free-space optical communication method based on joint optimization of transmission and reception, and adopts coding and compensation methods at the transmitting end and the receiving end respectively to alleviate the interference of atmospheric turbulence on the signal. The patent of the present invention designs a Polar coding method for the atmospheric turbulence channel at the transmitting end, balances the code length and code rate, and ensures the encoding and decoding overhead; at the same time, a Transformer model based on convolution enhancement is designed at the receiving end, which strengthens the learning of local features through convolution and establishes the connection between phase screen pixels, and the conjugate phase screen is given by the model for turbulence compensation.

[0048] like Figure 1 As shown, the specific steps include:

[0049] S1. At the transmitting end, the output of the neural network model determines the construction of the code in the Polar code, analyzes the reliability of each sub-channel, and determines whether the sub-channel is a frozen bit or an information bit under the current channel according to the turbulent channel;

[0050] S2, modulate the encoded information onto an optical carrier, transmit it after collimation and beam expansion, and control the size of the Gaussian beam by an aperture. Place a spiral phase plate at the rear end to convert the Gaussian beam into an OAM beam with better anti-turbulence performance;

[0051] S3, using a rotating turbulence phase plate to simulate the atmospheric turbulence channel, the OAM beam is disturbed by the atmospheric turbulence;

[0052] S4. At the receiving end, a beam splitter is used to split the received light into two beams. One beam of light is irradiated into an infrared camera, and the infrared camera is used to collect images and input into the neural network. The neural network analyzes the phase of the turbulent channel from the light intensity image and gives the conjugate phase. The SLM displays the compensated phase image.

[0053] S5, the other light split by the beam splitter, after the conjugate phase compensation on the SLM, uses the translation stage to couple the spatial light into the optical fiber and then enters the photodetector to complete the photoelectric conversion;

[0054] S6. Finally, the oscilloscope collects the data input by the PD, saves it, and performs offline processing by MATLAB to analyze the bit error rate.

[0055] Preferably, the code construction process of the Polar coding is to sort the reliabilities of the N sub-channels generated after the channel polarization operation from large to small, and then select the first K sub-channels as information bits, and the remaining NK sub-channels as frozen bits.

[0056] The following is a detailed description of the transmitter process based on Polar coding:

[0057] As an emerging channel coding scheme, Polar code not only has a structured construction method and superior asymptotic performance, but also has lower complexity. However, it is relatively rarely used in practical scenarios such as FSO systems.

[0058] The construction process of Polar code is to sort the reliability of N sub-channels generated after the channel polarization operation from large to small, and then select the first K sub-channels as information bits, and the remaining NK sub-channels as frozen bits. The core of Polar code lies in channel polarization. When designing Polar coding under atmospheric turbulence channels, evaluating the reliability of sub-channels is a key step, which not only affects the selection of information bits and frozen bits, but also affects the control of computational complexity and delay.

[0059] At the same time, the atmospheric turbulence channel is a time-varying channel model, and it is necessary to predict the channel state information to adjust the parameters of Polar coding, such as code length, code rate, and frozen bit distribution. The information at the transmitting end is sent after coding and modulation. After transmission through the channel, the receiving end uses the Successive Cancellation (SC) decoding algorithm to calculate the Bhattacharyya parameters of each subchannel;

[0060] Through continuous large-scale simulation, the Bhattacharyya parameters obtained by each sub-channel are accumulated each time. The smaller the value of the Bhattacharyya parameter is, the higher the reliability of the sub-channel is. Then, the accumulated sums of the Bhattacharyya parameters of each sub-channel are sorted, and the K sub-channels with the smallest accumulated sums are selected as information bits.

[0061] Therefore, in order to balance the error correction capability and coding efficiency, the patent of the present invention adopts a method combining polarization weight with deep learning. The polarization weight can be used as the basis for selecting information bits and frozen bits. Sub-channels with higher weights are considered to be more reliable and therefore more suitable for transmitting information bits. The patent of the present invention uses a data-driven approach to train the model, learn to evaluate the reliability of each sub-channel after polarization, and predict whether the sub-channel is an information bit or a frozen bit.

[0062] like Figure 2 As shown, in this embodiment, the Polar coding process includes:

[0063] Use simulation to collect subchannel status information, including the bit error rate of the received signal corresponding to different turbulence intensities as the channel transmission characteristics, and select the Bhattacharyya parameters and subchannel index of each subchannel after channel polarization as the characteristics of the network model for training;

[0064] Construct a data set, use the known Polar coding construction method as a benchmark, mark each subchannel as an information bit or a frozen bit, and use this as a training label;

[0065] A network model is constructed to accept turbulence intensity and coding parameters (code length, code rate) as input, and output the reliability prediction of each subchannel, representing the polarization weight of each subchannel as a probability that can be learned through a neural network to represent the subchannel as an information bit.

[0066] Then, define the loss function to feedback the error between the result and the true value.

[0067] Preferably, the network model includes an input layer for receiving feature vectors of sub-channels, such as bit error rate and coding parameters; a plurality of fully connected layers for extracting features and learning nonlinear mappings; an LSTM layer for processing channel state changes; and an output layer for outputting the probability or reliability score of each sub-channel as an information bit.

[0068] Preferably, when defining the loss function, a binary cross entropy loss function is used as a basic loss to determine whether the current subchannel is a frozen bit or an information bit:

[0069]

[0070] Where N is the number of samples, y i is the actual reliability label of the subchannel, is the probability predicted by the model. In order to consider the bitrate constraint and penalize the deviation between the model prediction and the target bitrate, the bitrate loss shown in the following formula is added:

[0071]

[0072] In the formula, R target is the target bitrate, usually set to 0.5, λ R The weight used to balance the bit rate loss. In order to constrain the code length, we add the following formula to the loss function, and the code length is set to 2048:

[0073] L Length =λ L N

[0074] In order to quantify the change of channel conditions, the signal-to-noise ratio is used as a quantitative indicator of channel quality, and the bit error rate is used as a measure of the performance difference before and after the channel condition changes. The loss term under the turbulent channel is shown as follows:

[0075] L channel =λ ch f(ΔBER)

[0076] Where f is a non-negative penalty factor that is proportional to the magnitude of the change in channel conditions, ΔBER is the change in bit error rate after the channel conditions change, and λ is ch is a hyperparameter for balancing loss weights;

[0077] Combining the above, we can form the final loss function as shown below:

[0078] L total =L BCE +λ R L RATE +λ L L Length +λ ch L channel .

[0079] Furthermore, in actual use, since the network at the transmitting end is unaware of the channel conditions, it needs to be used in conjunction with the neural network at the receiving end. The neural network at the receiving end determines the turbulence intensity of the current channel based on the light intensity image and feeds back to the network model at the transmitting end. After identifying the turbulence intensity of the received light intensity image, the trained model is used to evaluate the new sub-channel set to obtain the reliability prediction of each sub-channel. The output of the model is used as the input of the Polar coding construction to determine the final information bits and frozen bits.

[0080] like Figure 3 As shown, in this embodiment, at the receiving end, a convolution-enhanced Transformer model is designed, and the receiving end uses the convolution-enhanced Transformer model to learn the mapping relationship between the received light intensity image and the turbulence phase screen.

[0081] Preferably, the convolution-enhanced Transformer model uses convolution layers in the encoder to gradually reduce the spatial dimensions of the feature map, then uses upsampling and convolution layers in the decoder to gradually restore the spatial dimensions of the feature map, and uses Transformer modules in the encoder and decoder to capture the dependencies between pixels with longer distances in the feature map. The model combines the encoder, the decoder and the final convolution layer to generate a prediction result;

[0082] At the same time, the turbulence intensity of the current channel is also fed back to the neural network at the transmitter in the final output layer. The output of the network includes the turbulence intensity fed back to the transmitter and the current turbulence phase screen. In the model architecture based on the convolution-enhanced Transformer, in order to address the problem that the traditional Transformer has poor ability to extract high-order information of feature maps, the convolution module is combined in the architecture to increase the network's learning of local image features.

[0083] like Figure 4 As shown in the figure, the Encoder part contains multiple MBConv modules, each of which is followed by a SE (Squeeze-and-Excitation) module. In detail, the SE module learns the relationship between channels in the feature map through global average pooling and fully connected layers, and uses the sigmoid function to output weights to adjust the channel response of the feature map. The network also contains a MaxPool2d layer for downsampling. Similarly, the Decoder part contains multiple convolutional transpose layers ConvTranspose2d for upsampling. Each upsampling layer is followed by a Transformer module, which contains a multi-head self-attention mechanism and a feedforward network for interaction between different regions of the feature map and enhanced feature representation. The Decoder part gradually restores the spatial resolution of the features, while maintaining rich feature representation through Skip Connection and Transformer modules.

[0084] Furthermore, the MBConv module is a module that is used more frequently in the encoder part. Its essence is an improved convolution module. The specific structure is as follows: Figure 5 As shown in the figure, it mainly includes input convolution layer, normalization, depth convolution layer, SE module, point convolution layer and downsampling convolution, and captures more complex features by using depth separable convolution and SE module.

[0085] like Figure 5 As shown, the anti-turbulence free-space optical communication system based on the joint optimization of transmission and reception includes a receiving end neural network model CNN_Transformer, a transmitting end network model, a Polar encoder, a waveform generator AWG, an intensity modulator, a laser, a collimator Col, a spiral phase plate SPP, a turbulent phase plate, a beam splitter BS, a photodetector PD, an oscilloscope, an infrared camera and a spatial light modulator SLM;

[0086] The transmitting end designs a neural network model for the turbulent channel and outputs the encoded data. A training sequence is first sent to allow the receiving end neural network model CNN_Transformer to receive the light intensity image and then judge the turbulence intensity of the current atmospheric channel and feed it back to the transmitting end network model. The Polar encoder gives the distribution of the Polar coded information bits under the current channel conditions. The encoded information is input into the arbitrary waveform generator AWG according to the output of the model. The RF signal output by the waveform generator AWG is loaded onto the intensity modulator. The laser outputs a laser with a wavelength of 1550nm and inputs it into the intensity modulator. The bias controller is used to make the intensity modulator work in the linear region. The encoded information is modulated onto the optical carrier. The output of the intensity modulator passes through the collimating beam expander Col and then through the spiral phase plate SPP to be converted into an OAM beam with stronger turbulence resistance.

[0087] In the channel part, the turbulent phase plate is rotated to simulate the interference of the OAM light beam in the atmospheric turbulence channel. At the receiving end, the light beam enters the 10:90 beam splitter, of which 10% of the light beam is displayed by the infrared camera and input into the network; the other part is phase compensated by the spatial light modulator. The phase diagram displayed on the spatial light modulator SLM is given by the output of the neural network model at the receiving end. Finally, we use the translation stage to couple the compensated spatial light into the optical fiber and connect it to the photodetector PD and then the oscilloscope. After recording the data, we perform offline processing to analyze the bit error rate.

[0088] Although the embodiments of the present invention are described in conjunction with the accompanying drawings, the patent owner may make various variations or modifications within the scope of the appended claims. As long as they do not exceed the protection scope described in the claims of the present invention, they should be within the protection scope of the present invention.

Claims

1. A turbulence-resistant free-space optical communication method based on joint optimization of transmission and reception, characterized in that: include: At the transmitting end, the output of the neural network model determines the construction of the code in the Polar code, analyzes the reliability of each sub-channel, and determines whether the sub-channel is a frozen bit or an information bit under the current channel according to the turbulent channel; The encoded information is modulated onto an optical carrier, which is then emitted after being collimated and expanded. The size of the Gaussian beam is controlled by an aperture, and a spiral phase plate is placed at the rear end to convert the Gaussian beam into an OAM beam with better anti-turbulence performance. The OAM beam is disturbed by atmospheric turbulence by using a rotating turbulence phase plate to simulate the atmospheric turbulence channel. At the receiving end, a beam splitter is used to split the received light into two beams. One beam of light is irradiated into an infrared camera, which is used to collect images and input into the neural network. The neural network analyzes the phase of the turbulent channel from the light intensity image and gives the conjugate phase. The SLM displays the compensated phase image. The other light split by the beam splitter is compensated for the conjugate phase on the SLM, and then the translation stage is used to couple the spatial light into the optical fiber and then enter the photodetector to complete the photoelectric conversion. Finally, the oscilloscope collects the data input by the PD, saves it, and uses MATLAB to perform offline processing and bit error rate analysis.

2. The anti-turbulence free-space optical communication method based on joint optimization of transmission and reception according to claim 1, characterized in that: The construction process of the code in the Polar coding is to sort the reliabilities of the N sub-channels generated after the channel polarization operation from large to small, and then select the first K sub-channels as information bits, and the remaining NK sub-channels as frozen bits.

3. The anti-turbulence free-space optical communication method based on joint optimization of transmission and reception according to claim 2, characterized in that: The Polar encoding process includes: Use simulation to collect subchannel status information, including the bit error rate of the received signal corresponding to different turbulence intensities as the channel transmission characteristics, and select the Bhattacharyya parameters and subchannel index of each subchannel after channel polarization as the characteristics of the network model for training; Construct a data set, use the known Polar coding construction method as a benchmark, mark each subchannel as an information bit or a frozen bit, and use this as a training label; A network model is constructed to accept turbulence intensity and coding parameters as input and output the reliability prediction of each subchannel, representing the polarization weight of each subchannel as a probability that can be learned by a neural network to represent the subchannel as an information bit. Define the loss function to measure the error between the feedback result and the true value.

4. The anti-turbulence free-space optical communication method and system based on joint optimization of transmission and reception according to claim 2, characterized in that: The network model includes an input layer for receiving feature vectors of sub-channels, such as bit error rate and coding parameters; multiple fully connected layers for extracting features and learning nonlinear mapping; an LSTM layer for processing channel state changes; and an output layer for outputting the probability or reliability score of each sub-channel as an information bit.

5. The anti-turbulence free-space optical communication method based on joint optimization of transmission and reception according to claim 4 is characterized in that: When defining the loss function, the binary cross entropy loss function is used as the basic loss to determine whether the current subchannel is a frozen bit or an information bit: Where N is the number of samples, y i is the actual reliability label of the subchannel, is the probability predicted by the model. In order to consider the bitrate constraint and penalize the deviation between the model prediction and the target bitrate, the bitrate loss shown in the following formula is added: In the formula, R target is the target bitrate, usually set to 0.5, λ R The weight used to balance the bit rate loss. In order to constrain the code length, we add the following formula to the loss function, and the code length is set to 2048: L Length =λ L N In order to quantify the change of channel conditions, the signal-to-noise ratio is used as a quantitative indicator of channel quality, and the bit error rate is used as a measure of the performance difference before and after the channel condition changes. The loss term under the turbulent channel is shown as follows: L channel =λ ch ·f(ΔBER) Where f is a non-negative penalty factor that is proportional to the magnitude of the change in channel conditions, ΔBER is the change in bit error rate after the channel conditions change, and λ is ch is a hyperparameter for balancing loss weights; Combining the above, we can form the final loss function as shown below: L total =L BCE +λ R L RATE +λ L L Length +λ ch L channel 。 6. The anti-turbulence free-space optical communication method and system based on joint optimization of transmission and reception according to claim 1, characterized in that: The neural network at the receiving end determines the turbulence intensity of the current channel based on the light intensity image and feeds it back to the network model at the transmitting end. After identifying the turbulence intensity of the received light intensity image, the trained model is used to evaluate the new sub-channel set to obtain the reliability prediction of each sub-channel. The output of the model is used as the input of the Polar coding structure to determine the final information bits and frozen bits.

7. The anti-turbulence free-space optical communication method based on joint optimization of transmission and reception according to claim 1, characterized in that: The receiving end uses a convolution-enhanced Transformer model to learn the mapping relationship between the received light intensity image and the turbulence phase screen.

8. The anti-turbulence free-space optical communication method based on joint optimization of transmission and reception according to claim 7, characterized in that: The convolution-enhanced Transformer model uses convolution layers in the encoder to gradually reduce the spatial dimensions of the feature map, and then uses upsampling and convolution layers in the decoder to gradually restore the spatial dimensions of the feature map. Transformer modules are used in the encoder and decoder to capture the dependencies between pixels with longer distances in the feature map. The model combines the encoder, decoder and the final convolution layer to generate a prediction result. At the same time, in the final output layer, the turbulence intensity of the current channel is fed back to the neural network at the transmitter. The output of the network includes the turbulence intensity fed back to the transmitter and the current turbulence phase screen. Based on the model architecture of the convolution-enhanced Transformer, the convolution module is combined in the architecture. 9.Anti-turbulence free-space optical communication system based on joint optimization of transmission and reception, characterized by: Including the receiving end neural network model, the transmitting end network model, the polar encoder, the waveform generator, the intensity modulator, the laser, the collimating beam expander, the spiral phase plate, the turbulent phase plate, the beam splitter, the photodetector, the oscilloscope, the infrared camera and the spatial light modulator; The transmitting end designs a neural network model for a turbulent channel, outputs the encoded signal, first sends a training sequence, and allows the receiving end neural network model to receive the light intensity image and then judge the turbulence intensity of the current atmospheric channel and feed it back to the transmitting end network model. The Polar encoder gives the distribution of the Polar coded information bits under the current channel conditions. The encoded information is input into an arbitrary waveform generator according to the output of the model. The RF signal output by the waveform generator is loaded onto the intensity modulator. The laser outputs a 1550nm wavelength laser and inputs it into the intensity modulator. The bias controller is used to make the intensity modulator work in the linear region. The encoded information is modulated onto the optical carrier. The output of the intensity modulator passes through a collimating beam expander and then is converted into an OAM beam with stronger turbulence resistance through a spiral phase plate. In the channel part, the turbulent phase plate is rotated to simulate the interference of the OAM light beam in the atmospheric turbulence channel. At the receiving end, the light beam enters the 10:90 beam splitter, of which 10% of the light beam is displayed by the infrared camera and input into the network; the other part is phase compensated by the spatial light modulator. The phase diagram displayed on the spatial light modulator is given by the output of the neural network model at the receiving end. Finally, we use the translation stage to couple the compensated spatial light into the optical fiber and connect it to the photodetector and then the oscilloscope. After recording the data, we perform offline processing to analyze the bit error rate.

10. The anti-turbulence free-space optical communication system based on joint optimization of transmission and reception according to claim 9, characterized in that: The receiving end neural network model is a CNN_Transformer model.

Images