Reverse modulation optical communication bit synchronization method and system based on support vector machine

Abstract

The invention belongs to the technical field of communication, and relates to a reverse modulation optical communication bit synchronization method and system based on a support vector machine. The method comprises the steps that a system inputs feature vectors with labels; taking a label of the feature vector as a classification target, constructing a classifier, learning a mapping relation between the feature vector and the classification target by using a support vector machine, determining an optimal separation hyperplane, and obtaining a trained classifier; and the system receives the online transmitted feature vector, predicts the level value by using the trained classifier, and obtains a bit-synchronized target signal according to the prediction result of the level value and the position of the jump edge. According to the method for recognizing the jump edge based on the support vector machine, bit synchronization of reverse modulation wireless optical communication is achieved, signal features can be learned, the jump edge can be detected, and along with increase of turbulence intensity, the performance advantage is more obvious compared with a traditional method.

Description

Technical Field

[0001] The present invention belongs to the field of communication technology, and in particular, relates to a reverse modulation optical communication bit synchronization method and system based on a support vector machine. Background Art

[0002] Synchronization is the premise and foundation of communication. Reverse modulation wireless optical communication is affected by atmospheric turbulence. The use of traditional bit synchronization methods has problems such as poor synchronization performance and low accuracy.

[0003] Traditional free space optical communication (FSO) requires symmetrical installation of laser transmitting and receiving systems at both ends. Due to the good directivity and small divergence angle of laser (milli-arc level), the slight vibration of the platform will cause the deviation of the signal light, increase the power loss of the system, and increase the bit error rate. In severe cases, it will directly lead to communication interruption. Therefore, a complex Acquisition Tracking and Pointing (ATP) system is required at both the transmitting and receiving ends, which makes the traditional FSO system heavy, large in size, high in power consumption and increased in technical complexity, which restricts the application of traditional FSO on small platforms such as drones and small satellites.

[0004] Compared with the traditional FSO system, the Modulating Retro-reflector (MRR) wireless optical communication system is asymmetric. The system is divided into three parts: the interrogation end (including the optical transmitter and receiver), the reverse modulation end (including the reverse modulator and the reflection system), and the atmospheric channel. The reverse modulation end cleverly uses the principle of reflection to make the incident light beam return along the original path. The system model is shown in the attached figure. Figure 1 As shown. An unmodulated DC interrogation beam with stable light intensity is emitted from the optical transmitter, passes through the forward link and reaches the reverse modulation end. The reverse modulator modulates the data to be transmitted into the laser beam, and reflects the modulated laser beam through the reflection structure and returns to the optical receiver at the interrogation end through the reverse link. The optical signal is converted into an electrical signal by photoelectric conversion, and the electrical signal is demodulated to restore the data sent from the reverse modulation end. However, the atmospheric transmission channel is random and unpredictable. The laser beam transmitted in free space is affected by the energy attenuation effect of the atmospheric channel and the atmospheric turbulence effect, and has non-stationary characteristics. In particular, the beam will be distorted after being affected by atmospheric turbulence.

[0005] At present, the modulation technology widely used in reverse modulation wireless optical communication is OOK, which relies on the combination of NRZ to realize the modulation of the transmission signal. The transmission signal has only two states, "0" and "1". When the signal quality is good, the low level and high level are received accordingly. The initial phase of the code element sequence can be obtained by making a simple decision based on the decision threshold. However, when there is strong interference or signal fading, the level jitter at the receiving end is large, the level tracking is difficult, and the probability of misjudging the jump edge will greatly increase, resulting in the phase-locked loop being unable to obtain the comparison reference pulse for bit synchronization, and bit synchronization is difficult to achieve. The structure block diagram of the phase-locked loop is shown in the attached figure. Figure 2 Turbulence can cause fluctuations in the phase and irradiance of the propagating light beam. Under the influence of turbulence, it is difficult to track and judge the level when using traditional synchronization methods, and it is difficult to locate the rising and falling edges, which makes it difficult to determine the symbol decision time. Summary of the invention

[0006] In order to solve the above technical problems, the present invention provides a method and system for reverse modulation optical communication bit synchronization based on support vector machine.

[0007] In a first aspect, the present invention provides a method for bit synchronization of reverse modulation optical communication based on a support vector machine, comprising an offline training phase and an online discrimination phase: Offline training phase, including: The system inputs a labeled feature vector; the feature vector is a set of level values ​​damaged by turbulence received by the receiving end; the value of the label is determined according to the jump state of the middle position of the set of level values; Taking the labels of feature vectors as classification targets, constructing classifiers, using support vector machines to learn the mapping relationship between feature vectors and classification targets, determining the optimal separation hyperplane, and obtaining the trained classifiers; The online discrimination stage includes: The system receives the feature vector transmitted online, uses the trained classifier to predict the level value, and obtains the bit-synchronized target signal according to the predicted result of the level value and the position of the transition edge.

[0008] In a second aspect, the present invention provides a reverse modulation optical communication bit synchronization system based on a support vector machine, comprising an offline training unit and an online discrimination unit; Offline training unit, including input unit and model building and training unit: An input unit, used for inputting a labeled feature vector into the system; the feature vector is a group of level values ​​damaged by turbulence received by the receiving end; the value of the label is determined according to the jump state of the middle position of the group of level values; The model building and training unit is used to build a classifier using the label of the feature vector as the classification target, use the support vector machine to learn the mapping relationship between the feature vector and the classification target, determine the optimal separation hyperplane, and obtain the trained classifier; Online identification unit, including receiving unit and processing unit: A receiving unit, used for the system to receive the feature vector transmitted online; The processing unit is used to predict the level value by using the trained classifier, and obtain the bit-synchronized target signal according to the predicted result of the level value and the position of the transition edge.

[0009] Based on the above technical solution, the present invention can also be improved as follows.

[0010] Furthermore, a number of continuous sampling points are constructed into a feature vector, and the step of each feature vector is one sampling point.

[0011] Furthermore, when the system inputs a labeled feature vector, the feature vectors whose positions of the middle sampling points among all the sampling points jump are labeled as first label values, and the feature vectors whose positions do not jump are labeled as second label values.

[0012] Furthermore, when determining the optimal separating hyperplane, two boundary lines of the classified data with the same direction as the separating hyperplane are determined, and the vertical distances from the boundary points on the two boundary lines to the separating hyperplane are equal; the separating hyperplane with the largest distance between the boundary lines is determined to obtain the optimal separating hyperplane.

[0013] Furthermore, let the total number of input feature vectors be m, the dimension of the feature vector be n, and k be the number of feature vectors. is the kth eigenvector, is the mth eigenvector; is a set of n-dimensional feature vectors; the input of the system is m n-dimensional feature vectors, expressed as: , ; set up is the label corresponding to the kth eigenvector, is the label corresponding to the mth eigenvector, then the labels corresponding to each eigenvector are expressed as: , ; set up is the weight coefficient of the eigenvector, is the transposed matrix of weight coefficients, is the bias of the eigenvector, , the equation defining the optimal separating hyperplane is: ; The optimal separating hyperplane divides the feature vectors into two categories, and the dividing line is represented as: ; According to the mapping relationship between the feature vector and the classification target, the objective function is expressed as: .

[0014] Furthermore, using Python tools, we introduced the support vector machine of the sklearn library to build a classification model, and used the penalty factor and Gaussian kernel parameters as the classification model parameters; Perform stratified sampling on the data set, divide the data set into several mutually exclusive subsets of equal size, divide the mutually exclusive subsets into training set and validation set, train the classification model with the training set, and then use the validation set to validate the trained classification model to obtain the mean of several groups of test results; analyze the performance of the classification model through cross-validation method.

[0015] Furthermore, the penalty factor and Gaussian kernel parameters are selected from the array: .

[0016] Furthermore, the performance of the classification model is analyzed by cross-validation. Assume that the precision is P, the recall is R, and the harmonic mean of the precision and recall is , the positive class prediction is expressed as , the positive class prediction is expressed as negative class , the negative class prediction is expressed as positive class ,but: ; ; .

[0017] The beneficial effects of the present invention are: the present invention proposes a method for identifying transition edges based on a support vector machine to achieve bit synchronization of reverse modulation wireless optical communication, which can learn signal characteristics and detect transition edges, and as the turbulence intensity increases, the performance advantage is more obvious compared to traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a model diagram of a reverse modulation wireless optical communication system; Figure 2 A schematic diagram of a method for bit synchronization of reverse modulation optical communication based on a support vector machine provided in Example 1 of the present invention; Figure 3 The schematic diagram of the bit synchronization method of reverse modulation optical communication based on support vector machine is shown in FIG. Figure 4 The invention simulates a reverse modulation wireless optical communication system affected by turbulence in a real atmospheric channel; Figure 5 Schematic diagram of the optimal separating hyperplane; Figure 6 The light intensity waveforms with different flicker indices were measured for the experiment; Figure 7 The simulation diagram shows the performance comparison between the classifiers generated by the two methods and the traditional method for distinguishing the transition edge under different turbulence conditions; Figure 8 This is a simulation diagram comparing the performance of the SVM model with the traditional method; Fig. 9 The schematic diagram of the reverse modulation optical communication bit synchronization system based on support vector machine. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0020] Example 1 As an example, Figure 3 As shown, in order to solve the above technical problems, this embodiment provides a reverse modulation optical communication bit synchronization method based on a support vector machine (SVM), including an offline training stage and an online discrimination stage: Offline training phase, including: The system inputs a labeled feature vector; the feature vector is a set of level values ​​damaged by turbulence received by the receiving end; the value of the label is determined according to the jump state of the middle position of the set of level values; Taking the labels of feature vectors as classification targets, constructing classifiers, using support vector machines to learn the mapping relationship between feature vectors and classification targets, determining the optimal separation hyperplane, and obtaining the trained classifiers; The online discrimination stage includes: The system receives the feature vector transmitted online, uses the trained classifier to predict the level value, and obtains the bit-synchronized target signal according to the predicted result of the level value and the position of the transition edge.

[0021] Specifically, in the offline training stage, the input labeled feature vector is a set of turbulence-damaged level values ​​received by the receiving end. The value of the label is based on whether the middle position of the level value is a transition edge. If so, it is marked as +1, otherwise it is marked as -1. The label of the feature vector is used as the classification target, and the support vector machine is used to learn the mapping relationship between the two, and an optimal separation hyperplane is found under the condition of maximizing the boundary line interval. In the online discrimination stage, the system receives the feature vector transmitted online and predicts it using the classifier trained in the first stage. The class predicted as +1 indicates that there is a transition edge in the middle position of the level value. Once the position of the transition edge is found, bit synchronization can be achieved.

[0022] Although affected by turbulence and noise, the received symbol sequence still retains some characteristics of the transition edge. The present invention acquires the characteristics of the transition edge in the received symbol sequence by collecting a large amount of data through high-multiple sampling, introduces machine learning into data processing and analysis, and uses the advantages of machine learning to extract the potential characteristics of the data to realize the method of identifying the transition edge. The present invention proposes a signal transition edge judgment method based on a support vector machine, and solves the bit synchronization problem in reverse modulation wireless optical communication under the influence of turbulence through a machine learning method.

[0023] By building an indoor experimental platform, such as Figure 4 As shown in the figure, a reverse modulation wireless optical communication system affected by turbulence in a real atmospheric channel is simulated to collect experimental data. The sending signal uses a data signal with a frequency of 100KHz, and the receiving end is set to a 2.5MHz frequency for sampling.

[0024] Optionally, a plurality of consecutive sampling points are constructed into a feature vector, and the step of each feature vector is one sampling point.

[0025] In actual application, 25 consecutive sampling points are constructed into a feature vector, and the step of each feature vector is one sampling point.

[0026] Optionally, when the system inputs a labeled feature vector, the feature vectors whose intermediate sampling points among all the sampling points have a jump position are labeled as first label values, and the feature vectors whose intermediate sampling points have no jump position are labeled as second label values.

[0027] Specifically, when marking feature vectors, the feature vectors with a jump edge at the middle of the 25 consecutive sampling points are marked as "+1", and those that do not meet this condition are marked as "-1". By learning the features, a classifier is built using a support vector machine, and the feature vectors to be classified are predicted. After classification, the position of the jump edge is successfully found, and the start and end times of each code element are determined.

[0028] Optionally, when determining the optimal separating hyperplane, two boundary lines of the classified data in the same direction as the separating hyperplane are determined, and the vertical distances from the boundary points on the two boundary lines to the separating hyperplane are equal; the separating hyperplane with the largest distance between the boundary lines is determined to obtain the optimal separating hyperplane.

[0029] As attached Figure 5 The optimal separation hyperplane diagram is shown in the figure. The horizontal axis is the value of the feature vector in the x direction, the vertical axis is the label value corresponding to the feature vector, and N1 represents the label value. The feature vector of N2 represents the label value. The feature vector of , L1 and L2 boundary lines, S is the separating hyperplane. The direction of the boundary line is consistent with the direction of the separating hyperplane. The points on the boundary line are the boundary points of the two types of data. The separating hyperplane needs to ensure that the distances from the two types of boundary points to the separating hyperplane are equal. The distance between the boundary lines is called the boundary line interval. There is and only one separating hyperplane that can maximize the interval. Such a separating hyperplane is called the optimal separating hyperplane.

[0030] Optionally, let the total number of input feature vectors be m, the dimension of the feature vector be n, and k be the number of feature vectors. is the kth eigenvector, is the mth eigenvector; is a set of n-dimensional feature vectors; the input of the system is m n-dimensional feature vectors, expressed as: , ; set up is the label corresponding to the kth eigenvector, is the label corresponding to the mth eigenvector, then the labels corresponding to each eigenvector are expressed as: , ; set up is the weight coefficient of the eigenvector, is the transposed matrix of weight coefficients, is the bias of the eigenvector, , the equation defining the optimal separating hyperplane is: ; The optimal separating hyperplane divides the feature vectors into two categories, and the dividing line is represented as: ; According to the mapping relationship between the feature vector and the classification target, the objective function is expressed as: .

[0031] Optionally, use Python tools to introduce the support vector machine of the sklearn library to build a classification model, and use the penalty factor and Gaussian kernel parameters as classification model parameters; Perform stratified sampling on the data set, divide the data set into several mutually exclusive subsets of equal size, divide the mutually exclusive subsets into training set and validation set, train the classification model with the training set, and then use the validation set to validate the trained classification model to obtain the mean of several groups of test results; analyze the performance of the classification model through cross-validation method.

[0032] Suppose the data set is D, and perform stratified sampling on the data set D, dividing D into s mutually exclusive subsets of basically the same size, and ensuring that the data distribution between each subset is consistent. Then, use s-1 subsets as training sets each time, and the remaining subsets as validation sets. Use the training set to train the classification model, and then use the validation set to validate the trained classification model to obtain the mean of the s groups of test results. Measuring the performance of the classification model by the cross-validation method can effectively avoid overfitting and underfitting, making the classification results more convincing.

[0033] Optionally, the penalty factor and Gaussian kernel parameters are selected from an array: .

[0034] The problem of imbalanced positive and negative samples is solved by setting classification weights. Different categories contain different numbers of samples. By assigning different weights to different categories, the goal is to increase the weight of the minority sample category and reduce the weight of the majority sample category. This method eliminates the need to process the sample itself, and because of the support of the support vector machine in the sklearn library, it only needs to set the balance of parameters and specify the weights of different categories. You can also use its default balance parameters. At this time, the weight is inversely proportional to the number of samples in different categories. The calculation formula is: number of samples / number of categories*number of classification samples.

[0035] Optionally, the performance of the classification model can be analyzed by cross-validation. Let the precision be P, the recall be R, and the harmonic mean of the precision and recall be , the positive class prediction is expressed as , the positive class prediction is expressed as negative class , the negative class prediction is expressed as positive class ,but: ; ; .

[0036] Through experimental simulation and verification, hot air is sent into the two ports of the air slot by using a hot air blower. By combining the different gears of the two hot air blowers, turbulence of different intensities is simulated. Three sets of experimental data are selected, and the flicker index is used to describe the intensity of the turbulence. Assume that the received light intensity is , <> represents the ensemble average, and the scintillation index is , the calculation formula of the flicker index is: .

[0037] Through calculation, the flicker index of the three cases is obtained as follows: 0.03, 0.048 and 0.064. The waveform of the received data is shown in the attached figure. Figure 6 As shown, the horizontal axis is the number of sampling points, and the vertical axis is the level value (unit: V). Figure 6 (a) is the light intensity waveform with a flicker index of 0.03, (b) is the light intensity waveform with a flicker index of 0.048, and (c) is the light intensity waveform with a flicker index of 0.064. Since the data was collected in an indoor simulated environment, the focus was on simulating the influence of turbulence. In the indoor environment, the length of the air slot is 4 meters, the communication distance is short, and the signal-to-noise ratio is larger than that in the real atmospheric environment. Therefore, the awgn function in the matlab tool is used to further add noise to the experimental data to generate feature vectors under different signal-to-noise ratios. In order to accelerate the convergence speed and improve the classification accuracy, the data is normalized. The three sets of data measured in the experiment are denoised (SNR=5, 10, 15, 20) and normalized for training. Two methods are used respectively. One is to use the data of each signal-to-noise ratio to train and generate a classifier for the corresponding signal-to-noise ratio data. The other is to integrate the four different signal-to-noise ratio data into a training set to generate an overall classifier. Attached Figure 7 This is a simulation diagram comparing the performance of the classifiers generated by the two methods and the traditional method for distinguishing the transition edge under different turbulence conditions. The horizontal axis is the signal-to-noise ratio (in dB), and the vertical axis is the harmonic mean of the precision and recall rate. Attached Figure 7 (d) is a performance comparison simulation diagram for a flicker index of 0.03, (e) is a performance comparison simulation diagram for a flicker index of 0.048, and (f) is a performance comparison simulation diagram for a flicker index of 0.064.

[0038] The simulation results show that the performance of the SVM classifiers of the two methods is similar, and compared with the traditional method, the performance in the recognition of the jump edge is better. Therefore, the method of integrating data with different signal-to-noise ratios into a training set to generate a classifier is selected, which avoids the limitation of calling different classifiers for different signal-to-noise ratios. Next, the classifier trained by this method is used to test data with other signal-to-noise ratios that have not been trained. Three sets of data with SNR of 8, 12 and 18 are selected. The results are shown in Table 1.

[0039] Table 1 Comparison of fitting performance on untrained data:

[0040] From the results in the table, we can see that although the data with SNR of 8, 12 and 18 were not trained when generating the classifier, the classifier can still identify the transition edge of the data well. Therefore, it is not necessary to traverse the data under all signal-to-noise ratios when generating the classifier, which greatly reduces the amount of training and expands the scope of application and practicality.

[0041] Attached Figure 8 It is the fitting curve of using SVM classifier and traditional method to identify the jump edge, with the horizontal axis being the signal-to-noise ratio and the vertical axis being the signal-to-noise ratio. Figure 8 It can be seen that under the same signal-to-noise ratio, the stronger the turbulence, the easier it is to misjudge the transition edge, because as the turbulence increases, the random fluctuation of the signal will increase, and the characteristics of the transition edge will become less obvious. When the turbulence intensity is the same, the smaller the noise, the larger the signal-to-noise ratio, and the higher the probability of the transition edge being correctly identified. As the noise increases, the level value changes more, and the possibility of a position that was not a transition edge being misjudged as a transition edge increases. The value decreases accordingly. When the signal-to-noise ratio is large (SNR=20 in the figure), turbulence is the main influencing factor. Comparing the performance of the traditional method and the SVM classifier, observing the difference between the two red lines, the two green lines and the two blue lines, we can see that when the turbulence is small, the performance of the SVM method is similar to that of the traditional method. As the turbulence becomes stronger, the superiority of the SVM method becomes more prominent in terms of performance. Therefore, the introduction of the support vector machine has greatly improved the accuracy of the jump edge recognition.

[0042] The simulation results show that when the channel is in a weak turbulence environment, regardless of the size of the turbulence, the SVM method can learn signal features and detect transition edges, and as the turbulence intensity increases, the performance advantage of the SVM method becomes more obvious. Since the construction of the classifier is completed in the offline stage, it is not at a disadvantage in terms of the time required for online prediction. Therefore, when the SVM method is used in reverse modulation wireless optical communication to distinguish the transition edge of the OOK signal to achieve signal synchronization, the influence of turbulence can be overcome. The present invention realizes bit synchronization of MRR wireless optical communication through a support vector machine, and uses a support vector machine to realize the judgment of the transition edge.

[0043] The present invention proposes a method for identifying transition edges based on a support vector machine to achieve bit synchronization of reverse modulation wireless optical communication. The method can learn signal characteristics and detect transition edges, and as the turbulence intensity increases, the performance advantage is more obvious than that of traditional methods.

[0044] Example 2 Based on the same principle as the method shown in Example 1 of the present invention, as shown in the attached Fig. 9 As shown, an embodiment of the present invention also provides a reverse modulation optical communication bit synchronization system based on a support vector machine, including an offline training unit and an online discrimination unit; Offline training unit, including input unit and model building and training unit: An input unit, used for inputting a labeled feature vector into the system; the feature vector is a group of level values ​​damaged by turbulence received by the receiving end; the value of the label is determined according to the jump state of the middle position of the group of level values; The model building and training unit is used to build a classifier using the label of the feature vector as the classification target, use the support vector machine to learn the mapping relationship between the feature vector and the classification target, determine the optimal separation hyperplane, and obtain the trained classifier; Online identification unit, including receiving unit and processing unit: A receiving unit, used for the system to receive the feature vector transmitted online; The processing unit is used to predict the level value by using the trained classifier, and obtain the bit-synchronized target signal according to the predicted result of the level value and the position of the transition edge.

[0045] Optionally, a plurality of consecutive sampling points are constructed into a feature vector, and the step of each feature vector is one sampling point.

[0046] Optionally, when the system inputs a labeled feature vector, the feature vectors whose intermediate sampling points among all the sampling points have a jump position are labeled as first label values, and the feature vectors whose intermediate sampling points have no jump position are labeled as second label values.

[0047] Optionally, when determining the optimal separating hyperplane, two boundary lines of the classified data in the same direction as the separating hyperplane are determined, and the vertical distances from the boundary points on the two boundary lines to the separating hyperplane are equal; the separating hyperplane with the largest distance between the boundary lines is determined to obtain the optimal separating hyperplane.

[0048] Optionally, let the total number of input feature vectors be m, the dimension of the feature vector be n, and k be the number of feature vectors. is the kth eigenvector, is the mth eigenvector; is a set of n-dimensional feature vectors; the input of the system is m n-dimensional feature vectors, expressed as: , ; set up is the label corresponding to the kth eigenvector, is the label corresponding to the mth eigenvector, then the labels corresponding to each eigenvector are expressed as: , ; set up is the weight coefficient of the eigenvector, is the transposed matrix of weight coefficients, is the bias of the eigenvector, , the equation defining the optimal separating hyperplane is: ; The optimal separating hyperplane divides the feature vectors into two categories, and the dividing line is represented as: ; According to the mapping relationship between the feature vector and the classification target, the objective function is expressed as: .

[0049] Optionally, use Python tools to introduce the support vector machine of the sklearn library to build a classification model, and use the penalty factor and Gaussian kernel parameters as classification model parameters; Perform stratified sampling on the data set, divide the data set into several mutually exclusive subsets of equal size, divide the mutually exclusive subsets into training set and validation set, train the classification model with the training set, and then use the validation set to validate the trained classification model to obtain the mean of several groups of test results; analyze the performance of the classification model through cross-validation method.

[0050] Optionally, the penalty factor and Gaussian kernel parameters are selected from an array: .

[0051] Optionally, the performance of the classification model can be analyzed by cross-validation. Let the precision be P, the recall be R, and the harmonic mean of the precision and recall be , the positive class prediction is expressed as , the positive class prediction is expressed as negative class , the negative class prediction is expressed as positive class ,but: ; ; .

[0052] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A bit synchronization method for reverse modulation optical communication based on support vector machine, characterized in that: Including offline training stage and online discrimination stage: Offline training phase, including: The system inputs a labeled feature vector; the feature vector is a set of level values ​​damaged by turbulence received by the receiving end; the value of the label is determined according to the jump state of the middle position of the set of level values; Taking the labels of feature vectors as classification targets, constructing classifiers, using support vector machines to learn the mapping relationship between feature vectors and classification targets, determining the optimal separation hyperplane, and obtaining the trained classifiers; The online discrimination stage includes: The system receives the feature vector transmitted online, uses the trained classifier to predict the level value, and obtains the bit-synchronized target signal according to the predicted result of the level value and the position of the transition edge.

2. According to claim 1, the reverse modulation optical communication bit synchronization method based on support vector machine is characterized in that: Several consecutive sampling points are constructed into a feature vector, and the step of each feature vector is one sampling point.

3. According to claim 1, the reverse modulation optical communication bit synchronization method based on support vector machine is characterized in that: When the system inputs a labeled feature vector, the feature vectors whose intermediate sampling points have a jump position among all the sampling points are labeled as the first label value, and the feature vectors whose intermediate sampling points have no jump position are labeled as the second label value.

4. According to claim 1, the reverse modulation optical communication bit synchronization method based on support vector machine is characterized in that: When determining the optimal separating hyperplane, two boundary lines of the classified data with the same direction as the separating hyperplane are determined, and the vertical distances from the boundary points on the two boundary lines to the separating hyperplane are equal; the separating hyperplane with the largest distance between the boundary lines is determined to obtain the optimal separating hyperplane.

5. According to claim 1, the reverse modulation optical communication bit synchronization method based on support vector machine is characterized in that: Assume that the total number of input feature vectors is m, the dimension of the feature vector is n, and k is the number of feature vectors. is the kth eigenvector, is the mth eigenvector; is a set of n-dimensional feature vectors; the input of the system is m n-dimensional feature vectors, expressed as: ， ； set up is the label corresponding to the kth eigenvector, is the label corresponding to the mth eigenvector, then the labels corresponding to each eigenvector are expressed as: ， ； set up is the weight coefficient of the eigenvector, is the transposed matrix of weight coefficients, is the bias of the eigenvector, , the equation defining the optimal separating hyperplane is: ; The optimal separating hyperplane divides the feature vectors into two categories, and the dividing line is represented as: ； According to the mapping relationship between the feature vector and the classification target, the objective function is expressed as: 。 6. The method for bit synchronization of reverse modulation optical communication based on support vector machine according to claim 1, characterized in that: Using Python tools, we introduced the support vector machine of the sklearn library to build a classification model, and used the penalty factor and Gaussian kernel parameters as the classification model parameters; Perform stratified sampling on the data set, divide the data set into several mutually exclusive subsets of equal size, divide the mutually exclusive subsets into training set and validation set, train the classification model with the training set, and then use the validation set to validate the trained classification model to obtain the mean of several groups of test results; analyze the performance of the classification model through cross-validation method.

7. The method for bit synchronization of reverse modulation optical communication based on support vector machine according to claim 6, characterized in that: The penalty factor and Gaussian kernel parameters are selected from the array: 。 8. The method for bit synchronization of reverse modulation optical communication based on support vector machine according to claim 6, characterized in that: The performance of the classification model is analyzed by cross-validation. Assume that the precision is P, the recall is R, and the harmonic mean of the precision and recall is , the positive class prediction is expressed as , the positive class prediction is expressed as negative class , the negative class prediction is expressed as positive class ,but: ； ； 。 9. A reverse modulation optical communication bit synchronization system based on support vector machine, characterized in that: Includes offline training unit and online discrimination unit; Offline training unit, including input unit and model building and training unit: An input unit, used for inputting a labeled feature vector into the system; the feature vector is a group of level values ​​damaged by turbulence received by the receiving end; the value of the label is determined according to the jump state of the middle position of the group of level values; The model building and training unit is used to build a classifier using the label of the feature vector as the classification target, use the support vector machine to learn the mapping relationship between the feature vector and the classification target, determine the optimal separation hyperplane, and obtain the trained classifier; Online identification unit, including receiving unit and processing unit: A receiving unit, used for the system to receive the feature vector transmitted online; The processing unit is used to predict the level value by using the trained classifier, and obtain the bit-synchronized target signal according to the predicted result of the level value and the position of the transition edge.

Images