OAM superposition mode identification and turbulence intensity detection method based on domain self-adaption

Abstract

The invention discloses an OAM superposition mode identification and turbulence intensity detection method based on domain self-adaption, and the method comprises the steps: S1, receiving a light intensity picture, and inputting the light intensity picture into a pre-trained multi-task neural network model; and S2, identifying an OAM light beam superposition mode based on the multi-task neural network model and detecting the current turbulence intensity. According to the method, a network originally suitable for a single-topological charge OAM light beam is realized, a data set does not need to be re-established for training, the superposed OAM mode can be identified and the intensity of a turbulent channel can be analyzed only by a small amount of data, the training cost is greatly reduced, and the method has the advantages of being simple and convenient to operate, high in cost effectiveness, high in real-time performance and the like, and has a wide application prospect. The method is expected to be widely applied to multiple fields such as channel monitoring, emergency communication and data return.

Description

Technical Field

[0001] The present invention relates to the technical field of turbulence intensity detection, and more specifically, to a method for OAM superposition mode recognition and turbulence intensity detection based on domain adaptation. Background Art

[0002] FSO is a wireless optical communication technology that uses the atmosphere as a transmission medium. Due to its advantages of high bandwidth, high confidentiality and low latency, it shows great application potential in urban communications, broadband access, data center interconnection and other fields. In 1992, Allen et al. proved that a vortex beam with a spiral wavefront exp(ilφ) carries orbital angular momentum (OAM), where l is the OAM state or topological charge, which can be any integer, and φ is the angular coordinate. Theoretically, OAM can provide an infinite number of orthogonal bases with different modes, and each order is orthogonal to each other. Based on this characteristic, OAM beams are usually used as optical carriers in FSO to transmit data, which can be used to improve the communication capacity and bandwidth utilization of FSO. Due to the orthogonality of OAM, a laser beam can contain multiple different OAM components at the same time, and the superposition state of OAM will present different spot images. However, atmospheric turbulence is one of the key factors affecting the performance of FSO systems. Turbulence is a complex fluid motion phenomenon that is ubiquitous in nature. The intensity and characteristics of turbulence in the atmosphere directly affect the change in the refractive index of the air, causing the light beam to shake, diverge and flicker, leading to signal attenuation, increased bit error rate, and even communication interruption. Therefore, for the dynamic channel of atmospheric turbulence, accurate identification of OAM beams and real-time detection of turbulence intensity are of vital importance for optimizing FSO system design and improving communication quality and reliability.

[0003] The current more mature OAM mode perception method mainly infers the mode components and distribution of OAM through the related characteristics of interference and diffraction. However, due to the limitations of parameters such as optical system performance and optical device resolution, the mode recognition and OAM spectrum measurement of multi-mode hybrid beams with a wide OAM mode range are difficult, and the recognition rate is low. Especially when transmitting in inhomogeneous media or scattering media, the distorted phase and intensity distribution will significantly reduce the accuracy of mode recognition. In application scenarios such as free-space optical communication and radar detection, research needs are more focused on accurately measuring the original light field information emitted from the transmitter at the receiving end. Traditional methods often reduce the impact of disturbances by introducing adaptive optics, and then combine subsequent analysis and processing to restore the original light field information. It is often for OAM beams with a single turbulence condition, a single transmission distance, and a single mode. In the face of dynamic turbulence, it is often difficult to adapt to superposition OAM beams with different link lengths. Therefore, how to quickly and accurately perceive the OAM mode field distribution of a multimode hybrid beam over a wide OAM mode range, and how to ensure the accuracy of OAM mode perception in the presence of inhomogeneous media or other disturbances, is one of the key challenges in this field.

[0004] With the significant improvement in computing power and the development of artificial intelligence technology, it has become possible to identify turbulence intensity using machine learning and image processing techniques. These technologies are able to process large amounts of remote sensing data, automatically identify turbulence features, and provide more accurate and comprehensive turbulence analysis. In 2018, Li Jin and others from Beijing University of Posts and Telecommunications proposed a scheme for atmospheric turbulence detection and adaptive demodulation using a convolutional neural network (CNN), which achieved a detection accuracy of 99.8% in simulated strong turbulence links. In 2022, Zheng Chonghui and others from Changchun University of Science and Technology proposed a method for identifying orbital angular momentum beams using a deep transfer learning method, achieving a recognition rate of more than 90% in both weak and medium turbulence environments. By comparing with traditional deep learning methods, it is proved that training time can be reduced while maintaining a high recognition rate in weak and medium turbulence environments.

[0005] In the existing Free Space Optical Communication (FSO) system, although the channel estimation and turbulence intensity identification technologies have made significant progress, there are still some challenges and limitations in the dynamic and complex atmospheric turbulence environment. Traditional beam identification methods may not be able to accurately capture the rapidly changing channel conditions in real time, especially in the environment of high turbulence intensity. In addition, the existing beam identification technology is often targeted at a single mode, and may lack sufficient flexibility and adaptability in the face of superposition OAM beams, and it is difficult to cope with the changing atmospheric conditions and beam propagation characteristics. Therefore, in order to solve the problem of superposition OAM beam identification and turbulence intensity detection under dynamic turbulence intensity conditions, it is necessary to develop a domain-adaptive OAM superposition mode identification and turbulence intensity detection method to achieve faster and more accurate superposition mode identification and turbulence intensity detection. Summary of the invention

[0006] The purpose of the present invention is to provide a domain-adaptive OAM superposition mode recognition and turbulence intensity detection method to overcome the defects of the prior art.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0008] A method for OAM superposition mode recognition and turbulence intensity detection based on domain adaptation comprises the following steps:

[0009] S1, receiving a light intensity picture, and inputting the light intensity picture into a pre-trained multi-task neural network model;

[0010] S2. Identify the OAM beam superposition mode and detect the current turbulence intensity based on the multi-task neural network model.

[0011] Furthermore, the training steps of the multi-task neural network model are:

[0012] The low-frequency compensated phase screen is used to simulate the transmission process of optical signals in atmospheric turbulence to generate a data set of light intensity images with different turbulence intensities, different transmission distances and different OAM modes.

[0013] The data format of the data set is set to a unified RGB image and corresponding normalization and noise reduction preprocessing is performed, and the data set is divided into a training set, a validation set, and a test set according to a set ratio;

[0014] A convolutional neural network is used to train the training set to learn the transmission characteristics of a single topological charge OAM beam, and the network is allowed to learn the transmission characteristics of OAM beams with different topological charges through domain-adaptive transfer learning. The network is then verified and tested using a validation set and a test set.

[0015] Furthermore, the multi-task neural network model includes an input layer, a feature extraction layer, a classifier, a regressor, a domain discriminator and an output layer. The input layer is used to input data including a light intensity image and two sets of labels of its corresponding turbulence intensity and OAM mode. The feature extraction layer adopts a convolutional neural network, including multiple convolution modules, whose convolution kernel size is 3*3, padding=1, stride=1, and each convolution operation in the convolution block is followed by batch normalization BN processing and a ReLU activation function, and a 2*2 maximum pooling layer is used. The MMD distance of the source domain and target domain features is calculated after the convolution layer. The source domain is an OAM receiving spot image under a single topological load, and the target domain is an OAM receiving spot image in a superposition state. The MMD distance is optimized as part of the loss function to minimize the statistical distance between the feature distributions of the two domains and realize the alignment of the feature space. An adaptation layer is set after the feature extraction layer, and the domain discrimination loss is calculated on the output of the adaptation layer to realize adversarial training.

[0016] Furthermore, the step of calculating the MMD distance between the source domain and the target domain features after the convolution layer specifically includes:

[0017] Use the mapping function φ to map the data of the source domain and the target domain into a feature space;

[0018] Calculate the mean of the source domain features and the mean of the target domain features;

[0019] The kernel function calculation formula is used to calculate the MMD values ​​of the source domain and target domain features:

[0020]

[0021] In the formula, P and Q are the distributions of the source domain and the target domain respectively, and x i and j are the samples in the source and target domains, respectively, φ is the mapping function, n and m are the number of samples in the source and target domains, respectively.

[0022] Furthermore, the domain discriminator is used to distinguish source domain and target domain features output by the feature extractor;

[0023] For the source domain feature, the domain discriminator classifies it as the source domain, and the loss function is

[0024]

[0025] For the target domain feature, the domain discriminator classifies it as the target domain, and the loss function is

[0026]

[0027] Among them, D is the domain discriminator, f is the feature extractor, and P s and P t are the distributions of the source domain and the target domain respectively, and the adversarial loss function is expressed as:

[0028]

[0029] Furthermore, the output layer is used to realize feature decoupling, including OAM superposition mode recognition and turbulence intensity detection. For the OAM superposition mode recognition task, a softmax layer is used for classification; for the turbulence intensity detection task, a linear layer is used for regression, and the total loss function includes the cross entropy loss of OAM mode recognition, the mean square error loss of turbulence intensity detection, the MMD distance and the adversarial loss;

[0030] The cross entropy loss for OAM mode identification is:

[0031]

[0032] In the formula, M is the sample category, y ic is a sign function. The true category of the observed sample i is equal to category c and takes 1, otherwise it takes 0. ic is the predicted probability that the observed sample belongs to category c;

[0033] The mean square error loss of turbulence intensity detection is:

[0034]

[0035] In the formula, k is the number of samples, y is i is the probability predicted by the model, is the true value probability;

[0036] The MMD distance and adversarial loss are integrated into the total loss function as follows:

[0037] Loss=Loss cross_entropy +λ1Loss MSE +λ2L MMD +λ3L adv

[0038] In the formula, Loss cross_entropy is the classification loss, Loss MSE is the regression loss, L MMD is the MMD distance, L adv is the adversarial loss, λ1, λ2, and λ3 are hyperparameters used to balance the weights of different loss terms.

[0039] Furthermore, during the training process of the multi-task neural network model, the parameters of the feature extractor, classifier, regressor and domain discriminator are alternately optimized, and the specific steps are as follows: the feature extractor and the domain discriminator are fixed, and the parameters of the classifier and the regressor are optimized to minimize the classification loss and the regression loss; the feature extractor is fixed, and the parameters of the domain discriminator are optimized to maximize its ability to distinguish between the source domain and the target domain features; the domain discriminator is fixed, and the parameters of the feature extractor are optimized to minimize the adversarial loss.

[0040] Compared with the prior art, the advantages of the present invention are: the present invention realizes a network originally suitable for single-topology-charged OAM beams, and there is no need to re-establish a data set for training. Only a small amount of data is needed to identify superimposed OAM modes and analyze the intensity of turbulent channels, which greatly reduces the training cost. It has the advantages of simple operation, high cost-effectiveness, and strong real-time performance, and is expected to be widely used in multiple fields such as channel monitoring, emergency communication, and data backhaul. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] 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.

[0042] Figure 1 It is a flow chart of the domain-adaptive OAM superposition mode recognition and turbulence intensity detection method of the present invention.

[0043] Figure 2 It is the received light intensity image of the present invention under three types of turbulence: weak, medium and strong.

[0044] Figure 3 This is a design idea diagram of the multi-modal OAM beam recognition and turbulence intensity detection of the present invention.

[0045] Figure 4 It is a network model architecture diagram 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] See also Figure 1As shown, this embodiment discloses a method for OAM superposition mode recognition and turbulence intensity detection based on domain adaptation, which is aimed at the OAM mode recognition and turbulence intensity detection problem at the receiving end of the communication system under different communication distances and turbulence intensities of vortex beams superimposed by OAM beams with different topological charges. The present invention designs a multi-task neural network model at the receiving end of the free-space optical communication system to learn the transmission characteristics of the single topological charge OAM beam, and allows the network to learn the transmission characteristics of OAM beams with different topological charges through domain adaptive transfer learning. According to the received light intensity image input into the neural network, the OAM beam superposition mode can be identified and the current turbulence intensity can be detected at the same time, which can be further expanded to transmission conditions such as different turbulence intensities, wavelengths and communication distances.

[0048] The domain-adaptive OAM superposition mode recognition and turbulence intensity detection method of this embodiment includes the following steps:

[0049] Step S1, receiving a light intensity picture, and inputting the light intensity picture into a pre-trained multi-task neural network model;

[0050] Step S2: identifying the OAM beam superposition mode and detecting the current turbulence intensity based on the multi-task neural network model.

[0051] In this embodiment, the training steps of the multi-task neural network model are:

[0052] The transmission process of optical signals in atmospheric turbulence is simulated using a phase screen after low-frequency compensation to generate a data set of light intensity images with different turbulence intensities, different transmission distances, and different OAM modes.

[0053] The dataset is divided into training set, validation set and test set in a ratio of 7:2:1. The image format is unified as 256*256 RGB image, and the corresponding normalization and noise reduction preprocessing is performed. The dataset is divided into target domain and source domain according to the actual communication needs. The source domain data is the OAM receiving spot image under a single topological charge, and the target domain data is the OAM receiving spot image in a superposition state. This patent takes the single topological charge dataset with l=1 as the source domain, and takes the superposition state dataset with l=1, l=5; l=1,3; l=2, l=-3 as the target domain as an example. The received light intensity images under three types of weak, medium and strong turbulence are shown in Table 2.

[0054] A convolutional neural network is used to train the training set to learn the transmission characteristics of a single topological charge OAM beam, and the network is allowed to learn the transmission characteristics of OAM beams with different topological charges through domain-adaptive transfer learning. The network is then verified and tested using a validation set and a test set.

[0055] In this embodiment, Figure 3As shown in the figure, the design ideas for multi-modal OAM beam identification and turbulence intensity detection include multi-task output of convolutional feature extraction, domain adaptation and feature decoupling. The convolutional feature extraction module learns features from the light intensity image, and then migrates the network suitable for single topology OAM beams to the FSO system of superposition OAM beams through domain adaptation technology. Finally, the learned features are decoupled, and the network model simultaneously outputs the turbulence intensity of the channel and identifies the OAM superposition mode of the current received beam.

[0056] The multi-task neural network model includes an input layer, a feature extraction layer, a classifier, a regressor, a domain discriminator and an output layer. The input layer is used to input data including a received light intensity image and two sets of labels of its corresponding turbulence intensity and OAM mode. The feature extraction layer adopts a convolutional neural network, including multiple convolution modules, whose convolution kernel size is 3*3, padding=1, stride=1, and each convolution operation in the convolution block is followed by batch normalization BN processing and the use of ReLU activation function, and a 2*2 maximum pooling layer is used to provide rich feature representation for subsequent tasks.

[0057] Considering that the transmission characteristics of OAM beams with different topological charges may be different, domain adaptation is used to adjust the model to adapt to the new data distribution, including the use of MMD and adversarial training. The specific implementation method is to calculate the MMD distance of the source domain and target domain features after the convolution layer, and optimize it as part of the loss function to minimize the statistical distance between the feature distributions of the two domains and achieve alignment of the feature space. At the same time, a domain discriminator is introduced, an adaptation layer is added after the feature extraction layer, and the domain discrimination loss is calculated on the output of the adaptation layer to implement adversarial training. Through adversarial training, the feature extractor generates feature representations that are difficult to be distinguished by the domain discriminator.

[0058] In this embodiment, the step of calculating the MMD distance between the source domain and the target domain features after the convolution layer specifically includes:

[0059] Use the mapping function φ to map the data of the source domain and the target domain into a feature space;

[0060] Calculate the mean of the source domain features and the mean of the target domain features;

[0061] The kernel function calculation formula is used to calculate the MMD values ​​of the source domain and target domain features:

[0062]

[0063] In the formula, P and Q are the distributions of the source domain and the target domain respectively, and x i and jare the samples in the source and target domains, respectively, φ is the mapping function, n and m are the number of samples in the source and target domains, respectively.

[0064] In this embodiment, the domain discriminator is used to distinguish the source domain and target domain features output by the feature extractor;

[0065] For the source domain feature, the domain discriminator classifies it as the source domain, and the loss function is

[0066]

[0067] For the target domain feature, the domain discriminator classifies it as the target domain, and the loss function is

[0068]

[0069] Among them, D is the domain discriminator, f is the feature extractor, and P s and P t are the distributions of the source domain and the target domain respectively, and the adversarial loss function is expressed as:

[0070]

[0071] In this embodiment, the output layer is used to achieve feature decoupling, including OAM superposition mode recognition and turbulence intensity detection. For the OAM superposition mode recognition task, a softmax layer is used for classification; for the turbulence intensity detection task, a linear layer is used for regression, and the total loss function includes the cross entropy loss of OAM mode recognition, the mean square error loss of turbulence intensity detection, the MMD distance and the adversarial loss;

[0072] The cross entropy loss for OAM mode identification is:

[0073]

[0074] In the formula, M is the sample category, y ic is a sign function. The true category of the observed sample i is equal to category c and takes 1, otherwise it takes 0. ic is the predicted probability that the observed sample belongs to category c;

[0075] The mean square error loss of turbulence intensity detection is:

[0076]

[0077] In the formula, k is the number of samples, y is i is the probability predicted by the model, is the true value probability;

[0078] The MMD distance and adversarial loss are integrated into the total loss function as follows:

[0079] Loss=Loss cross_entropy +λ1Loss MSE +λ2L MMD +λ3L adv (7)

[0080] In the formula, Loss cross_entropy is the classification loss, Loss MSE is the regression loss, L MMD is the MMD distance, L adv is the adversarial loss, λ1, λ2, and λ3 are hyperparameters used to balance the weights of different loss terms.

[0081] In this embodiment, during the training of the multi-task neural network model, the parameters of the feature extractor, classifier, regressor and domain discriminator are optimized alternately, and the specific steps are as follows: fix the feature extractor and domain discriminator, optimize the parameters of the classifier and regressor to minimize the classification loss and regression loss; fix the feature extractor, optimize the parameters of the domain discriminator to maximize its ability to distinguish the source domain and target domain features; fix the domain discriminator, optimize the parameters of the feature extractor to minimize the adversarial loss, so that the features generated by the feature extractor are difficult for the domain discriminator to distinguish their source domain. The feature extractor is forced to generate features that are consistent between different domains, and the domain discriminator provides a mechanism to promote this consistency.

[0082] Through the above process, the network originally applied to single topology-charged OAM beams in this embodiment is trained multiple times and hyperparameters are iterated to continuously minimize the total loss function, and the performance of classification and regression is improved. At the same time, the distribution difference between the source domain and the target domain features is reduced, so that the generalization ability of the model in the target domain is also improved, and finally the recognition of superimposed OAM modes and the detection of turbulence intensity are realized. This embodiment realizes the network originally applied to single topology-charged OAM beams. There is no need to re-establish the data set for training. Only a small amount of data is needed to identify the superimposed OAM mode and analyze the intensity of the turbulent channel, which greatly reduces the training cost. It has the advantages of simple operation, high cost-effectiveness, strong real-time performance, etc., and is expected to be widely used in multiple fields such as channel monitoring, emergency communication and data backhaul.

[0083] 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 method for OAM superposition mode recognition and turbulence intensity detection based on domain adaptation, characterized in that: The following steps are involved: S1, receiving a light intensity picture, and inputting the light intensity picture into a pre-trained multi-task neural network model; S2. Identify the OAM beam superposition mode and detect the current turbulence intensity based on the multi-task neural network model.

2. The method for OAM superposition mode recognition and turbulence intensity detection based on domain adaptation according to claim 1 is characterized in that: The training steps of the multi-task neural network model are: The transmission process of optical signals in atmospheric turbulence is simulated using a phase screen after low-frequency compensation to generate a data set of light intensity images with different turbulence intensities, different transmission distances, and different OAM modes. The data format of the data set is set to a unified RGB image and corresponding normalization and noise reduction preprocessing is performed, and the data set is divided into a training set, a validation set, and a test set according to a set ratio; A convolutional neural network is used to train the training set to learn the transmission characteristics of a single topological charge OAM beam, and the network is allowed to learn the transmission characteristics of OAM beams with different topological charges through domain-adaptive transfer learning. The network is then verified and tested using a validation set and a test set.

3. The method for OAM superposition mode recognition and turbulence intensity detection based on domain adaptation according to claim 2 is characterized in that: The multi-task neural network model includes an input layer, a feature extraction layer, a classifier, a regressor, a domain discriminator and an output layer. The input layer is used to input data including a light intensity image and two sets of labels of its corresponding turbulence intensity and OAM mode. The feature extraction layer adopts a convolutional neural network, including multiple convolution modules, whose convolution kernel size is 3*3, padding=1, stride=1, and each convolution operation in the convolution block is followed by batch normalization BN processing and a ReLU activation function. A 2*2 maximum pooling layer is used, and the MMD distance of source domain and target domain features is calculated after the convolution layer. The source domain is an OAM receiving spot image under a single topological load, and the target domain is an OAM receiving spot image in a superposition state. The MMD distance is optimized as part of the loss function to minimize the statistical distance between the feature distributions of the two domains and realize the alignment of the feature space. An adaptation layer is set after the feature extraction layer, and the domain discrimination loss is calculated on the output of the adaptation layer to realize adversarial training.

4. The method for OAM superposition mode recognition and turbulence intensity detection based on domain adaptation according to claim 3 is characterized in that: The step of calculating the MMD distance between the source domain and the target domain features after the convolution layer specifically includes: Use the mapping function φ to map the data of the source domain and the target domain into a feature space; Calculate the mean of the source domain features and the mean of the target domain features; The kernel function calculation formula is used to calculate the MMD values ​​of the source domain and target domain features: In the formula, P and Q are the distributions of the source domain and the target domain respectively, and x i and j are the samples in the source and target domains, respectively, φ is the mapping function, n and m are the number of samples in the source and target domains, respectively.

5. The method for OAM superposition mode recognition and turbulence intensity detection based on domain adaptation according to claim 3 is characterized in that: The domain discriminator is used to distinguish the source domain and target domain features output by the feature extractor; For the source domain feature, the domain discriminator classifies it as the source domain, and the loss function is For the target domain feature, the domain discriminator classifies it as the target domain, and the loss function is Among them, D is the domain discriminator, f is the feature extractor, and P s and P t are the distributions of the source domain and the target domain respectively, and the adversarial loss function is expressed as:

6. The method for OAM superposition mode recognition and turbulence intensity detection based on domain adaptation according to claim 3 is characterized in that: The output layer is used to achieve feature decoupling, including OAM superposition mode recognition and turbulence intensity detection. For the OAM superposition mode recognition task, a softmax layer is used for classification; for the turbulence intensity detection task, a linear layer is used for regression. The total loss function includes the cross entropy loss of OAM mode recognition, the mean square error loss of turbulence intensity detection, the MMD distance and the adversarial loss. The cross entropy loss for OAM mode identification is: In the formula, M is the sample category, y ic is a sign function. The true category of the observed sample i is equal to category c and takes 1, otherwise it takes 0. ic is the predicted probability that the observed sample belongs to category c; The mean square error loss of turbulence intensity detection is: In the formula, k is the number of samples, y is i is the probability predicted by the model, is the true value probability; The MMD distance and adversarial loss are integrated into the total loss function as follows: Loss=Loss cross_entropy +λ1Loss MSE +λ2L MMD +λ3L adv In the formula, Loss cross_entropy is the classification loss, Loss MSE is the regression loss, L MMD is the MMD distance, L adv is the adversarial loss, λ1, λ2, and λ3 are hyperparameters used to balance the weights of different loss terms.

7. The method for OAM superposition mode recognition and turbulence intensity detection based on domain adaptation according to claim 3 is characterized in that: During the training process of the multi-task neural network model, the parameters of the feature extractor, classifier, regressor and domain discriminator are optimized alternately. The specific steps are as follows: fix the feature extractor and domain discriminator, optimize the parameters of the classifier and regressor to minimize the classification loss and regression loss; fix the feature extractor, optimize the parameters of the domain discriminator to maximize its ability to distinguish between source domain and target domain features; fix the domain discriminator, optimize the parameters of the feature extractor to minimize the adversarial loss.

Images