Method and system for demodulating vortex optical communication based on weighted extreme random forest model

Abstract

The present application relates to the technical field of communication, especially to a vortex light communication demodulation method and system based on a weighted extreme random forest model, the method comprising the following steps: S10, demultiplexing a received vortex light multiplexed beam to obtain received power of different modes and converting the received power into digital signals; S20, inputting the digital signals of the received power of different modes into a pre-trained weighted extreme random forest model for demodulation to obtain demodulation signals; the splitting mode of the weighted extreme random forest model is as follows: extracting a feature from a plurality of features contained in samples with different weights, randomly extracting a sample from the sample set, and splitting based on the extracted feature and sample. The present application regards crosstalk as useful information for signal analysis, without the need for equalization and other operations, and can directly judge the received power information, thereby reducing complexity, greatly simplifying the communication system, shortening the development cycle, and reducing costs, so as to effectively improve the overall performance of vortex light multiplexing communication.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of communication, in particular to a vortex light communication demodulation method and system based on a weighted extreme random forest model. BACKGROUND

[0002] The orbital angular momentum (OAM) carried by vortex beams provides a new dimension of resources for the spatial domain of light waves, attracting more and more researchers' attention. In the transmission process of vortex light multiplexing communication, inter-mode distortion is a common problem, which will affect the stability and quality of the communication signal. In order to deal with this distortion, traditional demodulation methods such as multi-user detection technology and MIMO detection algorithm are widely used, which can effectively cope with this challenge to a certain extent. However, its precision is low, it cannot handle nonlinear interference, it needs to perform channel estimation, carrier recovery, information equalization and other signal preprocessing, and it cannot meet the increasing demand for transmission rate and capacity of communication.

[0003] In recent years, the rapid development of machine learning technology has brought new opportunities to the field of OAM optical communication. With its excellent data analysis and information processing capabilities, compared with traditional digital signal processing algorithms, machine learning algorithms have shown significant advantages in analyzing vortex beam distortion and accurately demodulating signals. Such as deep neural network (DNN), convolutional neural network (CNN), etc. These algorithms can directly analyze the distortion of light intensity to classify the received light beams by OAM mode, and then demodulate them through mapping and decoding methods. However, these algorithms require a large amount of data for training, and the training time is long, which is prone to overfitting and cannot adapt well to new environments. Algorithms such as K-nearest neighbor (KNN), support vector machine (SVM), and random forest (RF) have relatively lower complexity and can better adapt to rapidly changing new environments. However, complexity is still a disadvantage of using these algorithms for demodulation. That is, the current research still has the status quo of improving demodulation accuracy and complexity. SUMMARY

[0004] The purpose of the present application is to provide a vortex light communication demodulation method and system based on a weighted extreme random forest model to reduce the complexity of demodulation while ensuring the accuracy of demodulation.

[0005] In order to achieve the above purpose, the present application provides the following technical solutions:

[0006] In a first aspect, the present application provides a vortex light communication demodulation method based on a weighted extreme random forest model, comprising the following steps:

[0007] S10, demultiplexing the received vortex light multiplexing beam to obtain the received power of different modes and converting it into a digital signal;

[0008] S20, inputting the digital signals of the received power of different modes into a pre-trained weighted extreme random forest model to obtain a demodulation signal; the splitting mode of the weighted extreme random forest model is: extracting one feature from a plurality of features contained in the sample with different weights, and randomly extracting one sample from the sample set, and splitting based on the extracted feature and sample.

[0009] The weighted extreme random forest model is obtained through the following steps of training:

[0010] S201, constructing a sample data set;

[0011] S202, randomly extracting a plurality of samples from the sample data set as a sample set in a way of extracting and then returning;

[0012] S203, when each node of the decision tree needs to be split, extracting one feature from a plurality of features contained in the sample with different weights, randomly extracting one sample from the sample set, and splitting based on the extracted feature and sample until the splitting termination condition is met;

[0013] S204, each decision tree taking the category with the largest number of samples in the sample subset of the leaf node as the output;

[0014] The operations of steps S202-S204 are performed K times to obtain K weighted extreme decision trees, and the K weighted extreme decision trees constitute the weighted extreme random forest model, and K is an integer greater than 1.

[0015] In the S203, the process of extracting one feature from a plurality of features contained in the sample with different weights includes:

[0016] The extraction probability of all features is calculated: for a decision tree, the training of the mode l k The extraction probability P kj of the feature j is: is the complex conjugate of , is the vortex light of the mode l k , is the distorted light beam of the light beam of the mode l m after transmitting a distance z in the turbulent flow, is the distorted light beam of the light beam of the mode l j after transmitting a distance z in the turbulent flow, and M is the total number of features;

[0017] The extraction probability is taken as the weight of the corresponding feature, and all the features with weights form a feature set;

[0018] Randomly extract a feature from the feature set.

[0019] In the above scheme, η kj , η km are crosstalk factors, representing the probability of other mode crosstalk, the probability of feature extraction is related to crosstalk, therefore the weighted extreme random forest model can adaptively adjust the division weight according to the change of environment, can dynamically identify and respond to the influence of different features on communication quality, so as to maintain high performance and reliability in complex and changeable communication environment. Through this adaptive adjustment, it can more effectively handle environmental noise, signal attenuation and other problems, and ensure the stability and efficiency of communication.

[0020] In a second aspect, the present application provides a vortex light communication demodulation system based on a weighted extreme random forest model, comprising:

[0021] A data receiving module is configured to receive the received power of different modes obtained by demultiplexing the vortex light multiplexed light beam and convert it into a digital signal.

[0022] An information demodulation module is configured to input the digital signal of the received power of different modes into a pre-trained weighted extreme random forest model for demodulation to obtain a demodulation signal; the splitting mode of the weighted extreme random forest model is: extracting a feature from a plurality of features contained in the sample with different weights, randomly extracting a sample from the sample set, and splitting based on the extracted feature and sample.

[0023] In a third aspect, the present application provides a computer program product comprising computer readable instructions, characterized in that the computer readable instructions, when executed by a processor, implement the steps of the vortex light communication demodulation method based on the weighted extreme random forest model of the present application.

[0024] In a fourth aspect, the present application provides a computer readable storage medium comprising computer readable instructions, characterized in that the computer readable instructions, when executed by a processor, implement the steps of the vortex light communication demodulation method based on the weighted extreme random forest model of the present application.

[0025] In a fifth aspect, the present application provides an electronic device comprising: a memory storing program instructions; a processor connected to the memory, executing the program instructions in the memory, and implementing the steps of the vortex light communication demodulation method based on the weighted extreme random forest model of the present application.

[0026] Compared with the prior art, the present application has the following technical advantages:

[0027] The application is a low-complexity machine learning demodulation algorithm, which regards crosstalk as useful information for signal analysis, without equalization and other operations, can directly judge after receiving power analog-to-digital conversion, greatly simplifies the communication system, and can prevent overfitting, has higher accuracy in low noise environment, can improve demodulation accuracy and rate, simplify the communication system, shorten the development cycle, reduce the cost, and effectively improve the overall performance of vortex light multiplexing communication.

[0028] When the decision tree node splits, each feature is assigned different weights during feature extraction, the weighting mechanism can highlight important features (generally, the greater the crosstalk factor, the greater the importance) while not ignoring other features. Compared with the average extraction of the extreme random forest algorithm (each feature is extracted with the same probability), the important features can be more easily extracted while not ignoring other features, thereby improving the accuracy. Effectively solve the lack of accuracy and the limitation of adaptability of traditional digital demodulation method, while avoiding the high complexity problem existing in machine learning algorithm. Can better cope with the complex environment of vortex light multiplexing communication.

[0029] As can be seen from the feature extraction probability formula, the extraction probability of the feature is determined by the crosstalk factor, and the crosstalk factor will change with the change of the channel. Therefore, when detecting different modes, the extraction probability of the feature will be different. In practical application, the WERF model can also adaptively adjust the division weight according to the change of the environment, and can dynamically identify and respond to the influence of different features on the communication quality, so as to maintain high performance and reliability in complex and changeable communication environment. Through this adaptive adjustment, environmental noise, signal attenuation and other problems can be more effectively handled, ensuring the stability and efficiency of communication.

[0030] The demodulation method of the application is suitable for marine environment, and provides a better demodulation scheme for underwater communication, and is also suitable for environments such as atmosphere and optical fiber.

[0031] Other advantages of the application are described in the embodiment part. BRIEF DESCRIPTION OF DRAWINGS

[0032] In order to more clearly illustrate the technical solutions in the embodiments of the application, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0033] Figure 1 A flowchart of a vortex light communication demodulation method based on a weighted extreme random forest model provided in the embodiment.

[0034] Figure 2 Training flow chart of the weighted extreme random forest model provided in the examples.

[0035] Figure 3 Structure diagram of the communication system.

[0036] Figure 4a , Figure 4b Structure hierarchy diagram of the weighted extreme decision tree and the weighted extreme random forest model, respectively.

[0037] Figure 5a Distribution diagram of the example sample set under different features.

[0038] Figure 5b Based on the example shown in Figure 5a Splitting process diagram of the decision tree.

[0039] Figure 6a , Figure 6b , Figure 6c The bit error rate results of the algorithm under different numbers of training samples in the sea test example SNR 16dB, SNR 6dB, SNR 12dB environment, respectively.

[0040] Figure 7a , Figure 7b , Figure 7c The bit error rate results of the algorithm under different numbers of decision trees in the sea test example 20 meters SNR 15dB, 20 meters SNR 10dB, 20 meters SNR 14dB environment, respectively.

[0041] Figure 8a , Figure 8b , Figure 8c The bit error rate results of the algorithm under different maximum depths in the sea test example 20 meters SNR 6dB, 20 meters SNR 10dB, 20 meters SNR 16dB environment, respectively.

[0042] Figure 9a , Figure 9b , Figure 9c , Figure 9d , Figure 9e The bit error rate results of the algorithm under different maximum depths in the sea test example 20 meters, 40 meters, 60 meters, 80 meters, Fig. 20 is a graph of the bit error rate results at different SNRs in a 20-meter environment.

[0043] Figure 10a , Figure 10b 、 Figure 10c Fig. 21 is a graph of the bit error rate results at different SNRs in a 1000-meter atmospheric test. 1000 meters, 500 meters, 1000 meters different SNR bit error rate results.

[0044] Figure 11 Fig. 22 is a block diagram of an electronic device. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical solutions and advantages of the present application clearer and more apparent, the present application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.

[0046] Referring to Figure 1 , the vortex light communication demodulation method based on the weighted extreme random forest model provided in the embodiment includes the following steps:

[0047] S10, receiving a vortex light multiplexing beam, and demultiplexing to obtain different mode receiving powers, and converting into digital signals.

[0048] S20, inputting the digital signals of the different mode receiving powers into a pre-trained weighted extreme random forest model for demodulation to obtain a demodulation signal, i.e., different mode sending information.

[0049] The weighted extreme random forest model is an innovation on the extreme random forest algorithm. The training complexity of the random forest algorithm is O(n*log(n)), and the testing complexity is O(n). The training complexity of the extreme random forest algorithm is O(n), and the testing complexity is O(n). However, if the extreme random forest algorithm is directly used for demodulation, the bit error rate will be high, especially in the case of medium-intensity crosstalk, the bit error rate will be higher. The weighted extreme random forest algorithm changes the splitting method of the random forest algorithm, which reduces the bit error rate while having the advantage of low complexity of the extreme random forest algorithm. The weighted extreme random forest is to extract a feature from a subset of samples (M features) with different weights at each node of the decision tree that needs to be split, and then randomly extract a sample from this feature for splitting. Since the feature is extracted from the subset of samples based on different weights, it has a certain pertinence, which in turn can reduce the bit error rate.

[0050] Referring to Figure 2 , the above-mentioned weighted extreme random forest model is trained by the following steps:

[0051] S201, construct a sample dataset.

[0052] Reference can be made to Figure 3 At the transmitting end, the information to be transmitted is loaded onto the laser transmitter through modulation (for example, using on-off keying OOK modulation); after the light beam is converted into different mode vortex beams by a different mode vortex beam converter, the vortex light multiplexing device obtains a multiplexed light beam, which is transmitted by a transmitting antenna and reaches the receiving end through a channel; the receiving end demultiplexes the vortex light multiplexing light beam according to the orthogonality of OAM, obtains the received power of different modes, and converts it into a digital signal.

[0053] The transmitting end generates a sequence of signals, and the digital signal of the received power of different modes obtained by demultiplexing constitutes a sample. The transmitting end generates several sequences of signals to obtain several samples, and all the samples constitute a sample dataset.

[0054] Reference can be made to Figure 3 In actual application, after the receiving end receives the power, the digital signal converted by the analog-to-digital conversion can be directly judged by using the demodulation method (WERF demodulation) of the application, thereby saving a lot of preprocessing operations on the signal, and thus the structure of the communication system can be greatly simplified.

[0055] S202, randomly extract n samples from the sample dataset as a sample set in a bootstrap sampling manner. n is an integer greater than 1.

[0056] A sample set is used to construct (or train) a weighted extreme decision tree (WEDT), and multiple different weighted extreme decision trees can be constructed by repeating the operation multiple times.

[0057] It should be noted that each decision tree is independent of each other, and the process of constructing the decision tree can be executed in parallel or in series. If executed in series, steps S202-S204 are executed in a loop until a predetermined number of decision trees are generated. If executed in parallel, the same operation is executed in parallel and independently according to steps S202-S204.

[0058] S203, when each node of the weighted extreme decision tree needs to be split, a feature is extracted from M features contained in the sample set with different weights, a sample is randomly extracted from the sample set, and the splitting is performed based on the extracted sample and feature until the splitting termination condition is met.

[0059] Reference can be made to Figure 4aA simple structure is exemplified, and the WEDT structure is composed of a root node, an internal node (i.e., a decision node), and a leaf node. Starting from the root node, a "divide and conquer" strategy is adopted to divide the sample set according to a split point selection mechanism. First, a split point is determined, i.e., a specific value selected on a certain feature in the sample is determined. Then, according to the split point, the samples in the sample set are divided into two sample subsets according to the size of the feature value, and the two newly generated sample subsets are recursively iterated by the above method until the split termination condition is met. The position of the sample subset that is no longer split is the leaf node.

[0060] In this embodiment, the split termination condition includes: 1. whether the samples all belong to one kind (each sample has a corresponding result, represented as a high level or a low level in communication, and if all the samples in the node are high / low levels, it means that the samples falling into the node are high / low levels, and no further splitting is needed, for example Figure 5b The first leaf node is shown); 2. whether the maximum depth is reached; and 3. whether the minimum sample number is reached. The traditional random forest algorithm also has a split termination condition, i.e., whether the variance is less than eps, but this split termination condition is deleted in the present embodiment, because through experiments it is found that this parameter plays too low a role, and setting it cannot significantly improve the accuracy, but it needs to spend computational effort to calculate, and the cost performance is low, so this parameter is selected to be removed.

[0061] The traditional random forest algorithm randomly selects m features (m < M) from the sample (M features), and then selects one attribute as the split attribute of the node from the m attributes by using a certain strategy (such as information gain). The extreme random forest algorithm randomly selects one feature from M features (each feature is selected with the same probability), and then randomly selects one sample from the sample set for splitting. Although the complexity of the extreme random forest algorithm is lower, it has strong randomness and thus has a high error rate.

[0062] In the present embodiment, one feature is selected from M features according to the weight, and one sample is randomly selected from the sample set for splitting, so that the complexity is consistent with that of the extreme random forest algorithm, but the feature is selected according to different weights, rather than randomly, so it has a certain pertinence, and thus the error rate can be greatly reduced.

[0063] It is easy to understand that the selection of the feature and the selection of the sample are independent, and there is no priority in the execution order, i.e., one feature can be selected from M features contained in one sample, and then one sample can be selected from the sample set; or one sample can be selected from the sample set, and then one feature can be selected from M features contained in one sample.

[0064] S204, each weighted extreme decision tree takes the class with the most samples in the sample subset of the leaf node as the output. That is, in the leaf node, the class with the most samples is determined as the output class.

[0065] In the test phase, for each weighted extreme decision tree, the unknown sample starts from the root node, reaches a leaf node according to the division of the decision node, and outputs the classification result of the leaf node.

[0066] Reference can be made to Figure 4b The weighted extreme random forest (WERF) model is composed of multiple WEDTs, each of which is independently trained and can process information in parallel. When performing prediction, each WEDT outputs a prediction result for the same data, and then the prediction results of each WEDT are summarized to determine the class with the most votes as the final output.

[0067] As a typical representative of vortex beams, the Laguerre-Gaussian beam is favored for its simplicity and ease of experimental generation. The mathematical expression of its light field is as follows:

[0068]

[0069] In the formula: l is the mode number, r is the radius in cylindrical coordinates; θ represents the azimuth angle; z is the transmission distance; p is the radial index; the Rayleigh length λ is the wavelength of the beam; the wave number ω0 is the waist radius of the Gaussian beam; is the generalized Laguerre polynomial.

[0070] The basic principle of vortex light multiplexing communication is to use vortex beams with different OAM modes as carriers of modulated signals to realize the transmission of multiple signals by superposition. m (t) is a modulated signal, and the multiplexed vortex beam U MUX (r, θ, t) can be represented as:

[0071] When simulating the influence of the ocean or atmospheric environment on beam transmission, the power spectrum inversion method is used to generate a series of random phase screens. The process can be roughly divided into the following three steps: first, a complex Gaussian matrix is constructed, where the elements are complex values randomly generated according to the Gaussian distribution, which represents the initial spatial distribution of the refractive index in the turbulent medium; then, the complex Gaussian matrix is filtered using the channel power spectrum function, which ensures that the generated phase screen accurately reflects the statistical characteristics of the channel; finally, the filtered complex Gaussian matrix is converted into a random phase screen by performing inverse Fourier transform. This phase screen will be used to simulate the propagation of the beam in the channel.

[0072] During the transmission of vortex beams, whether in optical fiber, atmosphere or ocean environment, inter-mode crosstalk will occur, which affects communication. The demodulation algorithm of the application can be applied to the demodulation of all orbital angular momentum multiplexing communication, and the following will be introduced taking the ocean environment as an example.

[0073] Here, the ocean turbulence refractive index fluctuation spatial power spectrum model proposed by Nikishov et al. is adopted, which comprehensively considers temperature, salinity and refractive index fluctuation and other factors, and is widely used to generate ocean turbulence random phase screens. In a homogeneous isotropic seawater environment, the expression of the model is as follows:

[0074]

[0075] In the formula: k x k y are the frequency components in the x-axis and y-axis directions in the spatial frequency domain respectively, and ε is the turbulent kinetic energy dissipation rate of the unit fluid, reflecting the fluctuation intensity of the ocean turbulence; χ T is the mean square temperature dissipation rate, reflecting the influence of temperature fluctuation on the ocean turbulence; ψ is the temperature-salinity gradient rate; ζ is the inner scale of the ocean turbulence; and other parameters are A T = 1.863 x 10 -2 , A S = 1.9 x 10 -4 , A TS = 9.341 x 10 -3 , and δ = 8.284 (k 4 / 3 η) + 12.978 (k 2 η).

[0076] The principle of vortex light multiplexing communication is to regard the vortex light beam as a carrier of the modulated signal, s m (t) is the modulated signal corresponding to the mode l m , and is the Laguerre-Gaussian beam (vortex light) of the mode l m , and the multiplexed vortex light beam U MUX (r, θ, t) carrying information can be represented as

[0077] The multiplexed vortex light beam after turbulent transmission is represented as wherein is the distorted beam of the mode l m light beam after turbulent transmission for a distance z, and n(t) is the additive white Gaussian noise of the channel.

[0078] After the multiplexed light beam after turbulent transmission reaches the receiving end, demultiplexing can be performed according to the orthogonality between different modes of vortex light. Therefore, the mode lk Received signal y k It can be represented as:

[0079]

[0080] in yes The complex conjugate of η km The crosstalk factor represents mode l k The probability that feature m is s interfered with by other patterns k (t) represents pattern l k The corresponding modulation signal, η kk n represents the detection probability. k (t) is the output of additive white Gaussian noise after passing through the k-th distorted OAM-state matched filter, η km With n k The mathematical expression for (t) is as follows:

[0081]

[0082] Using matrix expressions, the signal received by the receiver can be represented as:

[0083]

[0084] It can be simplified to: Y = HS + N.

[0085] Where Y is the received signal, S is the transmitted signal, H is the channel matrix, and N is additive white Gaussian noise.

[0086] In step S203 above, the process of extracting a feature with different weights includes:

[0087] (1) Calculate the extraction probability of all features (the extraction probability is the probability that the feature is extracted): For decision trees, train on pattern l k During detection, the probability P of feature j being extracted is... kj for: η kj The crosstalk factor represents mode l k The probability that feature j is perturbed by other patterns can reflect the different patterns (including pattern l). k ) Vortex beam to mode l k The distortion of the vortex beam. For pattern l j The beam of light is distorted after traveling a distance z through turbulent currents.

[0088] (2) The extraction probability is used as a weight to be assigned to the corresponding feature, and all the features assigned weights form a feature set.

[0089] (3) Randomly extract a feature from the feature set.

[0090] Since each feature is given a different weight when extracting features, the weighting mechanism can highlight the key features while not ignoring other features. Compared to the average extraction of the extreme random forest algorithm (the probability of each feature being extracted is the same), the key features can be more easily extracted, thereby improving the accuracy.

[0091] In practical applications, when obtaining the crosstalk factor, only a certain mode of vortex light can be sent, and then the information received by all receiving ends is normalized to obtain the crosstalk factor corresponding to the mode. By repeatedly sending vortex light of all modes, the crosstalk factors of all modes can be obtained. This weight probability is not absolute, and a slight fluctuation will not cause an impact, so it is not afraid of noise interference, and can be repeated several times to take the average according to the situation. For example, as shown in the above matrix expression, when only mode s1, s2-s n are not sent, the received y1-y n is η 11 *s1+n~η 1n *s1+n, the crosstalk factor can be obtained by normalization to adjust the weight. Next, mode s2 is sent, and all crosstalk factors in the matrix are obtained. n .

[0092] In the training process, detection is performed for all modes in step S203. As can be seen from the extraction probability formula, the extraction probability of the feature is determined by the crosstalk factor, so the extraction probability of the feature will be different when detection is performed for different modes. In practical applications, the WERF model can also adaptively adjust the division weight according to the change of the environment, and can dynamically identify and respond to the influence of different features on the communication quality, thereby maintaining high performance and reliability in complex and variable communication environments. Through this adaptive adjustment, environmental noise, signal attenuation and other problems can be more effectively handled, ensuring the stability and efficiency of communication. Therefore, the weighted extreme random forest model will adaptively adjust with the change of the environment during use, and is not fixed after one training.

[0093] In terms of complexity comparison, Table 1 shows the training and testing complexity of a variety of commonly used and low complexity machine learning demodulation methods. n is the number of training samples, The number of samples is tested. As can be seen, the support vector machine (SVM) algorithm has a high complexity in the training stage, and the K nearest neighbor (KNN) algorithm does not need a training process, but its complexity in the test stage is high. In contrast, the RF algorithm is lower than the first two in maximum complexity. As for the more simplified ERF and WERF algorithms, it is obvious that they have lower complexity than RF. The WERF algorithm changes the probability of random extraction of the split attribute in the ERF algorithm, and the complexity of the two is the same.

[0094] Table 1: Comparison of the complexity of several low-complexity machine learning algorithms

[0095]

[0096] In order to better understand the splitting way of the random forest model, a simple example is given here for illustration.

[0097] Please refer to Figure 5a and Figure 5b , Figure 5a X1 represents the light intensity information received by feature 1, and X2 represents the light intensity information received by feature 2. The higher the received power is, the greater the received power is.

[0098] The circular symbol represents the high level state of mode 1, and the square symbol represents the low level state of mode 1. During the training process, the circle and square are known states. The numbers in the middle of the figure represent the serial numbers of the samples.

[0099] During the training process, first, from feature 1 (X1), it can be judged that Y X1 samples with Y X1 value greater than or equal to 6 are divided into high level (because the data in Figure 5a shows that samples with Y X1 value greater than 6 are all circles, i.e. high level state, so if the newly appeared unknown sample falls within this range, it is most likely to be high level). Similarly, samples with Y X1 value less than or equal to 3 are divided into low level (because the samples in this range are all squares, i.e. low level state). However, for samples with serial numbers 6, 7 and 8, since there are both circles and squares, it cannot be directly judged. At this time, feature 2 (X2) can be analyzed, at this time, only samples with serial numbers 6, 7 and 8 need to be considered, because the level state of other samples has been determined. By observing Figure 5a it can be found that the Y X2 value of samples with serial numbers 7 and 8 is greater than 5, while the Y X2 value of the sample with serial number 6 is less than 5. According to this, Y X2The samples with values greater than 5 are divided into squares, and the samples with values less than 5 are divided into circles. The specific splitting process can be seen in Figure 5b .

[0100] Figure 5a The triangle in the figure represents an unknown level state, and after the model is trained, the unknown level state can be classified. For example, according to the training results above, it can be judged that the sample with serial number 13 corresponds to a circular symbol, representing a high level state of mode 1; and serial numbers 14 and 15 correspond to square symbols, representing a low level state of mode 1. This process shows the importance of feature selection in splitting. If a feature with weak correlation is initially selected for splitting, not only will the computational burden be increased, but also misjudgment may occur. The weighted feature selection method proposed in this embodiment effectively solves this problem and improves the efficiency and accuracy of the algorithm. Although more complex situations may be encountered in actual applications, by constructing a large number of independent WEDTs, high accuracy and robustness can still be maintained.

[0101] From the above splitting process, it can be seen that in the demodulation method of the application, crosstalk is treated as useful information for signal analysis, and without equalization and other operations, the received power after analog-to-digital conversion can be directly judged, reducing the complexity, greatly simplifying the communication system, shortening the development cycle, and reducing the cost, thereby effectively improving the overall performance of vortex light communication.

[0102] In order to test the application effect of the algorithm proposed in the application in the field of underwater vortex light communication, detailed numerical simulation experiments were carried out with the help of the MATLAB platform in the research. A four-channel multiplexing communication system for underwater vortex light was built, and OOK modulation technology was used to directly load the information to be sent onto the laser generator. In this experiment, the influence of communication distance and turbulence intensity on communication performance was focused on. By adjusting the parameter Z to simulate different communication distances, and changing the parameter to control the intensity of turbulence, while keeping other system parameters unchanged (see Table 2) to ensure the accuracy and consistency of the experiment. For the marine environment, the strong turbulence, medium turbulence. In the experiment, 200,000 bits of data stream were transmitted each time, and the transmission experiment was repeated 1000 times under each set of conditions to calculate the average performance index, so as to obtain more representative experimental data.

[0103] Table 2: Parameter configuration table

[0104] Parameter Meaning Value l Vortex beam mode [1,2,3,4] p Vortex beam radial index 0 λ Gaussian beam wavelength 532 nm [CDATA[ω0]] Waist radius of Gaussian beam 3 x 10 -3 ]] ψ Temperature-salinity gradient rate -4 ζ Internal scale of ocean turbulence 1 x 10 -3 ]] - Matrix size of phase screen and laser 512×512 Phase screen separation 10m

[0105] As Figures 6a-6cAs shown in FIG. 4, the performance of RF and WERF algorithms under different number of training samples is compared in terms of BER for 20-meter transmission under various turbulence intensity and SNR levels, with the number of decision trees being 25 and the maximum depth being 10. These figures are used to compare the dependence of the two algorithms on the number of training samples. Figure 6a As shown in FIG. 5, the performance of RF and WERF algorithms under different number of decision trees is compared in terms of BER for 20-meter transmission under various turbulence intensity and SNR levels, with the number of training samples being 1500 and the maximum depth being 10. These figures are used to analyze the dependence of the two algorithms on the number of decision trees. Figure 6b As shown in FIG. 6, the performance of RF and WERF algorithms under different number of decision trees is compared in terms of BER for 20-meter transmission under various turbulence intensity and SNR levels, with the number of training samples being 1500 and the maximum depth being 10. These figures are used to analyze the dependence of the two algorithms on the number of decision trees. Figure 6c As shown in FIG. 7, the performance of RF and WERF algorithms under different number of decision trees is compared in terms of BER for 20-meter transmission under various turbulence intensity and SNR levels, with the number of training samples being 1500 and the maximum depth being 10. These figures are used to analyze the dependence of the two algorithms on the number of decision trees.

[0106] As shown in FIG. 4, the performance of RF and WERF algorithms under different number of training samples is compared in terms of BER for 20-meter transmission under various turbulence intensity and SNR levels, with the number of decision trees being 25 and the maximum depth being 10. These figures are used to compare the dependence of the two algorithms on the number of training samples. Figures 7a-7c As shown in FIG. 5, the performance of RF and WERF algorithms under different number of decision trees is compared in terms of BER for 20-meter transmission under various turbulence intensity and SNR levels, with the number of training samples being 1500 and the maximum depth being 10. These figures are used to analyze the dependence of the two algorithms on the number of decision trees. Figure 7a As shown in FIG. 6, the performance of RF and WERF algorithms under different number of decision trees is compared in terms of BER for 20-meter transmission under various turbulence intensity and SNR levels, with the number of training samples being 1500 and the maximum depth being 10. These figures are used to analyze the dependence of the two algorithms on the number of decision trees. Figure 7b As shown in FIG. 7, the performance of RF and WERF algorithms under different number of decision trees is compared in terms of BER for 20-meter transmission under various turbulence intensity and SNR levels, with the number of training samples being 1500 and the maximum depth being 10. These figures are used to analyze the dependence of the two algorithms on the number of decision trees. Figure 7c As shown in FIG. 6, the performance of RF and WERF algorithms under different number of decision trees is compared in terms of BER for 20-meter transmission under various turbulence intensity and SNR levels, with the number of training samples being 1500 and the maximum depth being 10. These figures are used to analyze the dependence of the two algorithms on the number of decision trees. Figure 7a As shown in FIG. 7, the performance of RF and WERF algorithms under different number of decision trees is compared in terms of BER for 20-meter transmission under various turbulence intensity and SNR levels, with the number of training samples being 1500 and the maximum depth being 10. These figures are used to analyze the dependence of the two algorithms on the number of decision trees. Figure 7b As shown in FIG. 6, the performance of RF and WERF algorithms under different number of decision trees is compared in terms of BER for 20-meter transmission under various turbulence intensity and SNR levels, with the number of training samples being 1500 and the maximum depth being 10. These figures are used to analyze the dependence of the two algorithms on the number of decision trees. As shown in FIG. 7, the performance of RF and WERF algorithms under different number of decision trees is compared in terms of BER for 20-meter transmission under various turbulence intensity and SNR levels, with the number of training samples being 1500 and the maximum depth being 10. These figures are used to analyze the dependence of the two algorithms on the number of decision trees.

[0107] As shown in FIG. 4, the performance of RF and WERF algorithms under different number of training samples is compared in terms of BER for 20-meter transmission under various turbulence intensity and SNR levels, with the number of decision trees being 25 and the maximum depth being 10. These figures are used to compare the dependence of the two algorithms on the number of training samples. Figures 8a-8cAs shown, the graphs present a comparison of the demodulation bit error rate (BER) of the RF and WERF algorithms over a 20-meter transmission distance under various turbulence intensities and noise levels, with different maximum depth settings, assuming a training sample size of 1500 and a decision tree size of 25. These graphs aim to explore the dependence of the two algorithms on the maximum depth of the decision tree. Figure 8a , Figure 8b The results show the signal-to-noise ratios of 6dB and 10dB under moderate turbulence conditions, respectively. Figure 8c This illustrates a noise level of 16 dB under strong turbulence. Combining these three figures, it can be seen that to achieve better demodulation results, the WERF algorithm requires a larger maximum depth than the RF algorithm, a difference of approximately 3. To improve the algorithm's fault tolerance, setting the maximum depth for both algorithms to around 13 can achieve bit error rate convergence. Especially from... Figure 8b It can be observed that when the maximum depth exceeds 10, the demodulation effect of the WERF algorithm is better than that of the RF algorithm, which indicates that in some cases, the demodulation performance of the WERF algorithm is indeed better than that of the RF algorithm.

[0108] By deeply investigating the impact of three key parameters—the number of training samples, the number of decision trees, and the maximum depth—on the convergence of demodulation bit error rate, the results show that the WERF algorithm is more sensitive to the number of decision trees and the maximum depth than the RF algorithm. However, the dependence of the number of training samples on this key parameter is roughly the same, so the overall complexity of the WERF algorithm is still significantly lower than that of the RF algorithm. Due to the large number of parameters involved in the algorithm, the experimental results in this part only reflect the trend of the parameters' impact on algorithm performance and cannot be used for comparing the optimal performance of the algorithms. Specific performance comparisons will be based on… Figures 9a-9e introduce.

[0109] Furthermore, the study also conducted an in-depth comparison of the bit error rate performance of four algorithms—RF, ERF (Extremely Random Forest), WERF, and MMSE (Minimum Mean Square Error)—under different ocean turbulence intensities, transmission distances, and signal-to-noise ratios. Regarding algorithm parameter settings, based on the results of the previous algorithm parameter selection experiments, parameters that achieve the best demodulation effect were selected for different transmission environments. Figures 9a-9d The graphs show the signal-to-noise ratio (SNR) results for transmission distances of 20, 40, 60, and 80 meters under moderate turbulence. Figure 9e The image shows the signal-to-noise ratio (SNR) results for transmission over 20 meters under strong turbulence.

[0110] Depend on Figures 9a-9d It can be seen that the RF algorithm and the WERF algorithm achieve a front-end error correction coding (FEC) score of 3.8 × 10⁻⁶ under basically the same signal-to-noise ratio. -3Tolerance. For this point can be said in the strong turbulence under the transmission of 80 meters, the demodulation effect of the two algorithms is similar, the required signal-to-noise ratio difference will not exceed 0.5dB. Overall, the signal-to-noise ratio required for the bit error rate to reach the FEC tolerance will increase with the increase of turbulence intensity and transmission distance, and the bit error rate can reach the FEC tolerance in the five environments in the figure. In most cases, the RF algorithm is better at demodulating in a strong noise environment, but the difference is not too much, and the WERF algorithm is better at demodulating in a low noise environment, which is particularly pronounced in environments with longer transmission distances or stronger turbulence. The ERF algorithm performs better in long-distance transmission and strong turbulence, especially in strong turbulence, which can achieve the demodulation effect of the RF algorithm. However, overall, the algorithm has limited applicability, and in most cases, the demodulation effect is far lower than that of the WERF algorithm. The MMSE algorithm performs better than RF and WERF in low-noise environments with medium turbulence within 40 meters, but for the FEC tolerance, its effect is far lower than that of RF and WERF, which is more pronounced at longer transmission distances, with a signal-to-noise ratio difference of 2dB to 10dB. Overall, the MMSE algorithm is more limited in its application environment than the RF and WERF algorithms. Figure 9e It can be seen that in a strong turbulence environment, the transmission distance is 20 meters, and the demodulation effect of the RF and WERF algorithms is similar to that of the medium turbulence under the transmission of 80 meters, and the demodulation effect of the MMSE algorithm is significantly lower than that of the other three algorithms. The signal-to-noise ratio required for the RF, ERF, and WERF algorithms to reach the FEC threshold is approximately 15dB, and in a low-noise environment, the WERF algorithm has the best demodulation effect. In summary, the WERF algorithm can meet the accuracy of the RF algorithm and has the low complexity of the ERF algorithm, and can adapt to various transmission environments, making it a more suitable algorithm for underwater vortex light multiplexing communication demodulation.

[0111] In order to test the application effect of the algorithm proposed in the application in the field of atmospheric vortex light communication, detailed numerical simulation experiments were conducted on the MATLAB platform in the study. A four-channel atmospheric vortex light multiplexing communication system was built, and OOK modulation technology was used to load the information to be sent directly to the laser generator. In this experiment, the focus is on the influence of communication distance and turbulence intensity on communication performance. By adjusting the parameter Z to simulate different communication distances, and changing the parameter to control the intensity of turbulence, while keeping other system parameters unchanged to ensure the accuracy and consistency of the experiment. According to the above modeling of the ocean communication system, the ocean power spectrum is replaced by the atmospheric turbulence power spectrum, and the same steps can achieve the conversion of the channel. The atmospheric turbulence power spectrum used is as follows:

[0112]

[0113] wherein, k x k y respectively represent the frequency components in the x-axis and y-axis directions in the spatial frequency domain, κ0=2π / L0, κ l =3.3 / l0, κ L =1 / L0, L0 and l0 respectively represent the outer scale and the inner scale of the turbulence.

[0114] Under the atmospheric turbulence experiment, Figures 10a-10c respectively represent the frequency components in the x-axis and y-axis directions in the spatial frequency domain, κ0=2π / L0, κ transmission of 1000 meters, transmission of 500 meters, transmission of 1000 meters, the bit error rate results of different algorithms and different SNRs. The experimental parameters are set as L0=50m, l0=1mm, the wavelength of the light beam is set as 1550nm, and the phase screen interval is 100 meters. The other parameters involved are consistent with Table 2. The parameters of the RF, ERF and WERF algorithms are all selected, the number of training samples is 1800, the decision tree is 30, and the maximum depth is 10. The comparison of the demodulation bit error rate is shown in the figure. According to the FEC threshold, when the transmission distance is 1000 meters, the WERF algorithm can reach the FEC threshold at about 9dB of SNR, which is basically the same as the RF algorithm, about 3dB less than the MMSE algorithm, and much lower than the ERF algorithm. When the transmission distance is 500 meters, the WERF can reach the FEC threshold at about 14dB of SNR, which is basically the same as the RF algorithm, about 2dB less than the ERF, and about 7dB less than the MMSE algorithm. When the transmission distance is 1000 meters, the WERF can reach the FEC threshold at about 17dB of SNR, which is about 1dB less than the RF, about 5dB less than the ERF, and about 7dB less than the MMSE algorithm. Therefore, it can be concluded that the demodulation method and the WERF provided in the embodiment are also applicable to the vortex light multiplexing communication demodulation in the atmospheric turbulence, and have good demodulation effect. transmission of 1000 meters, transmission of 500 meters, transmission of 1000 meters, the bit error rate results of different algorithms and different SNRs. The experimental parameters are set as L0=50m, l0=1mm, the wavelength of the light beam is set as 1550nm, and the phase screen interval is 100 meters. The other parameters involved are consistent with Table 2. The parameters of the RF, ERF and WERF algorithms are all selected, the number of training samples is 1800, the decision tree is 30, and the maximum depth is 10. The comparison of the demodulation bit error rate is shown in the figure. According to the FEC threshold, when the transmission distance is 1000 meters, the WERF algorithm can reach the FEC threshold at about 9dB of SNR, which is basically the same as the RF algorithm, about 3dB less than the MMSE algorithm, and much lower than the ERF algorithm. When the transmission distance is 500 meters, the WERF can reach the FEC threshold at about 14dB of SNR, which is basically the same as the RF algorithm, about 2dB less than the ERF, and about 7dB less than the MMSE algorithm. When the transmission distance is 1000 meters, the WERF can reach the FEC threshold at about 17dB of SNR, which is about 1dB less than the RF, about 5dB less than the ERF, and about 7dB less than the MMSE algorithm. Therefore, it can be concluded that the demodulation method and the WERF provided in the embodiment are also applicable to the vortex light multiplexing communication demodulation in the atmospheric turbulence, and have good demodulation effect.

[0115] Based on the same inventive concept, the embodiment simultaneously provides a vortex light communication demodulation system based on a weighted extreme random forest model, which comprises a data receiving module and an information demodulation module. The data receiving module is used for demultiplexing the received vortex light multiplexing light beam to obtain received power of different modes and converting the received power into digital signals. The information demodulation module is used for inputting the digital signals of the received power of different modes into a pre-trained weighted extreme random forest model for demodulation to obtain a demodulation signal. The splitting mode of the weighted extreme random forest model is that one feature is extracted from a plurality of features contained in samples with different weights, a sample is randomly extracted from the sample set, and splitting is performed based on the extracted feature and sample.

[0116] The specific processing procedure of the information demodulation module can refer to the related description in the foregoing method, and details are not described herein.

[0117] As shown in Figure 11 the present embodiment also provides an electronic device, which can include a processor 41 and a memory 42, wherein the memory 42 is coupled to the processor 41. It is worth noting that this figure is exemplary, and other types of structures can also be used to supplement or replace this structure to achieve data extraction, report generation, communication or other functions.

[0118] As shown in Figure 11 the electronic device can also include an input unit 43, a display unit 44 and a power supply 45. It is worth noting that the electronic device also does not necessarily include all the components shown in Figure 11 In addition, the electronic device can also include components not shown in Figure 11 can refer to prior art.

[0119] The processor 41, also known as a controller or operation control, can include a microprocessor or other processor device and / or logic device, which receives input and controls the operation of each component of the electronic device.

[0120] The memory 42, for example, can be one or more of a cache, a flash memory, a hard drive, a removable media, a volatile memory, a non-volatile memory or other suitable device, which can store configuration information of the processor 41, instructions executed by the processor 41 and other information. The processor 41 can execute the program stored in the memory 42 to achieve information storage or processing, etc. In one embodiment, the memory 42 also includes a buffer memory, i.e. a buffer, to store intermediate information.

[0121] The present embodiment also provides a computer program product, which includes computer readable instructions, when the computer readable instructions are executed in the electronic device, the program product causes the electronic device to perform the operation steps contained in the method of the present application.

[0122] The present embodiment also provides a storage medium storing computer readable instructions, which causes the electronic device to perform the operation steps contained in the method of the present application.

[0123] Those skilled in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware, computer software or a combination of both. In order to clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been described in the above description in a general manner. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. A skilled person can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0124] The integrated units, if realized in the form of software functional units and sold or used as independent products, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application or the part that contributes to the prior art essentially, or the whole or part of the technical solutions can be embodied in the form of a software product, which is stored in a storage medium and includes a number of instructions for making a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0125] The above-described embodiments are merely specific embodiments of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of various equivalent modifications, replacements and improvements within the technical scope disclosed by the present application, and these modifications, replacements and improvements should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A vortex optical communication demodulation method based on a weighted extreme random forest model, characterized in that, The method comprises the following steps: S10, the received vortex light multiplexed beam is demultiplexed to obtain different mode receiving power, and converted into a digital signal; S20, the digital signal of the different mode receiving power is input into a pre-trained weighted extreme random forest model for demodulation to obtain a demodulation signal; the splitting mode of the weighted extreme random forest model is: one feature is extracted from a plurality of features contained in the sample with different weights, one sample is randomly extracted from the sample set, and splitting is performed based on the extracted feature and sample; S201, a sample data set is constructed; S202, a plurality of samples are randomly extracted from the sample data set as a sample set in an extracted and then returned manner; S203, when each node of the decision tree needs to be split, one feature is extracted from a plurality of features contained in the sample with different weights, one sample is randomly extracted from the sample set, and splitting is performed based on the extracted feature and sample until the splitting termination condition is met; In the S203, the processing of extracting one feature from a plurality of features contained in the sample with different weights comprises: The extraction probability of all features is calculated: for the decision tree, the training pattern is detected, the extraction probability of feature j is : , , , is the complex conjugate of the vortex light of the pattern , the light beam distorted after the turbulent transmission distance of the light beam of the pattern , , the light beam distorted after the turbulent transmission distance of the light beam of the pattern , M is the total number of features; The extracted probability is used as a weight for the corresponding feature, and all the weighted features form a feature set; One feature is randomly extracted from the feature set; S204, each decision tree takes the category with the largest number of samples in the sample subset of the leaf node as the output; The operation of steps S202-S204 is performed K times to obtain K weighted extreme decision trees, and the K weighted extreme decision trees form the weighted extreme random forest model, and K is an integer greater than 1.

2. The vortex optical communication demodulation method based on the weighted extreme random forest model according to claim 1, characterized in that, In the S201, the processing of constructing the sample data set comprises: The information to be sent is loaded on the laser transmitter through modulation at the transmitting end; After the light beam passes through the different modal vortex beam converter to generate different mode vortex beams, the vortex light multiplexer obtains a multiplexed beam, the multiplexed beam is sent through the sending antenna, and reaches the receiving end through the channel; The receiving end demultiplexes the vortex light multiplexed beam to obtain different mode receiving power according to the orthogonality of OAM, and converts it into a digital signal to form a sample; The above operation is repeatedly performed until a certain number of samples are obtained, and all the samples form a sample data set. 3.The vortex optical communication demodulation method based on the weighted extreme random forest model according to claim 1, wherein, In the S203, the splitting termination condition is one of the following three: 1, whether the samples belong to the same category; 2, whether the maximum depth is reached; 3, whether the minimum number of samples is reached.

4. A vortex optical communication demodulation system based on a weighted extreme random forest model, characterized in that, It comprises: The data receiving module is configured to demultiplex the received vortex light multiplexed beam to obtain different mode receiving power, and convert it into a digital signal; The information demodulation module is configured to input the digital signal of the different mode receiving power into a pre-trained weighted extreme random forest model for demodulation to obtain a demodulation signal; the splitting mode of the weighted extreme random forest model is: one feature is extracted from a plurality of features contained in the sample with different weights, one sample is randomly extracted from the sample set, and splitting is performed based on the extracted feature and sample; The information demodulation module performs the following operation when extracting features: calculating extraction probabilities of all features, training a pattern for a decision tree The extraction probability of feature j when detecting is , , , is the complex conjugate of , is a vortex light of a pattern , is a distorted light beam of a pattern after a turbulent transmission distance , is a distorted light beam of a pattern after a turbulent transmission distance , and M is the total number of features; the extraction probability is used as a weight of the corresponding feature, and all weighted features form a feature set;​ One feature is randomly extracted from the feature set.

5. A computer program product comprising computer readable instructions, characterized in that, The computer readable instructions, when executed by the processor, implement the steps of the vortex light communication demodulation method based on the weighted extreme random forest model in any one of claims 1-3.

6. A computer-readable storage medium comprising computer-readable instructions, wherein, The computer readable instructions, when executed by the processor, implement the steps in the vortex optical communication demodulation method based on the weighted extreme random forest model according to any one of claims 1-3.

7. An electronic device, comprising: Comprise: A memory, storing program instructions; A processor, connected with the memory, executing the program instructions in the memory, and implementing the steps in the vortex optical communication demodulation method based on the weighted extreme random forest model according to any one of claims 1-3.

Images