Cloud wireless access network modeling method based on machine learning

Abstract

The invention relates to a cloud wireless access network modeling method based on machine learning, and the method comprises the steps: obtaining a remote cloud wireless access network transmitter, improving the remote cloud wireless access network transmitter, replacing a conventional microwave source of the remote cloud wireless access network transmitter with a double-ring photoelectric oscillator, and suppressing a side mode through single-mode oscillation, a direct modulation laser is introduced behind the arbitrary waveform generator and is used for generating power gain by utilizing a chirp effect and optical fiber dispersion cooperative gain; and based on a lightweight gradient elevator model, modeling a millimeter wave signal spectrum of the improved remote cloud wireless access network transmitter, and matching input feature dimensions and output feature dimensions. According to the invention, the dual-ring photoelectric oscillator is adopted to replace a traditional microwave source, so that phase noise is reduced; a cooperative gain behavior of directly modulating laser chirp and optical fiber dispersion is utilized to improve millimeter wave signal transmission gain; the spectral characteristics of millimeter wave signals are modeled through a lightweight gradient elevator model, and spectrum prediction is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of resource allocation, and particularly relates to a cloud radio access network modeling method based on machine learning. BACKGROUND

[0002] The cloud radio access network realizes flexible allocation and efficient management of resources through the separation of the centralized baseband processing unit and the distributed remote radio head, and is one of the core architectures of the next generation communication system. In the cloud radio access network, direct modulation lasers are widely used in optical front-haul links due to their low cost and convenient deployment, and optical wireless technology supports high-speed communication in the millimeter wave frequency band by transmitting radio frequency signals through optical fibers.

[0003] However, there are two key challenges in the performance optimization of the cloud radio access network system: on the one hand, the traditional microwave source has the problem of high phase noise, which limits the stability of the millimeter wave signal; on the other hand, the chirp effect of the direct modulation laser and the synergistic effect of the fiber dispersion will produce complex gain behavior, accompanied by high harmonic distortion, and the traditional analytical model is difficult to accurately capture the interaction of these nonlinear, dynamic characteristics and dispersion effects.

[0004] Specifically, the existing modeling method has the following limitations: (1) the analytical model proposed by the prior art needs to rely on complex mathematical derivation, and the modeling precision of the coupling effect of the nonlinear direct modulation laser, the dynamic characteristics of the electro-optical oscillator and the dispersion effect is insufficient, and the computational complexity is high; (2) the traditional machine learning model is prone to underfitting or unstable prediction when dealing with the dimension mismatch problem of input and output, and has long training time and limited generalization ability. SUMMARY

[0005] In order to solve the problems existing in the prior art, the purpose of the present application is to provide a cloud radio access network modeling method based on machine learning, which is based on a data-driven modeling framework of a lightweight gradient boosting machine, combined with a double-ring electro-optical oscillator technology, to realize efficient and high-precision modeling of the millimeter wave signal spectrum characteristics in the remote cloud radio access network architecture. The double-ring electro-optical oscillator realizes low phase noise and high spurious suppression ratio through the phase matching mechanism of the double-fiber loop; the lightweight gradient boosting machine solves the input-output dimension mismatch problem by automatically learning feature interaction, while retaining complete spectral information, improving modeling efficiency and accuracy.

[0006] To achieve the above purpose, the present application provides the following scheme:

[0007] A cloud radio access network modeling method based on machine learning, comprising:

[0008] acquiring a remote cloud radio access network transmitter, and improving the remote cloud radio access network transmitter:

[0009] The double-ring optoelectronic oscillator replaces the traditional microwave source of the remote cloud radio access network transmitter, and is used for suppressing side modes by single-mode oscillation and introducing a directly modulated laser after an arbitrary waveform generator, for generating power gain by using the chirp effect and fiber dispersion synergy gain;

[0010] The improved remote cloud radio access network transmitter is modeled based on a lightweight gradient boosting machine model, and input features and output feature dimensions are matched.

[0011] Optionally, suppressing side modes by the single-mode oscillation includes:

[0012] The double-ring optoelectronic oscillator uses a long optical fiber loop and a short optical fiber loop for single-mode oscillation and suppressing side modes:

[0013]

[0014] Wherein, f osc is the oscillation frequency, k and m are integers, τ l is the long loop delay, and τ s is the short loop delay.

[0015] Optionally, the optical field expression of the directly modulated laser is:

[0016]

[0017] Wherein, m a is the amplitude modulation index, ω IF is the intermediate frequency signal angular frequency, ω o is the laser carrier frequency angular frequency, is the maximum phase shift caused by intermediate frequency modulation, E DML (t) is the optical field of the directly modulated laser, and t is time.

[0018] Optionally, the Mach-Zehnder modulator of the improved remote cloud radio access network transmitter works in a carrier suppression mode, enhances sideband power, and the optical signal is transmitted through a single-mode optical fiber, loss is compensated by an erbium-doped fiber amplifier, and the optical signal is converted into a millimeter wave electrical signal by a photodetector at the remote radio head;

[0019] The electric field expression of the photodetector at the remote radio head is:

[0020]

[0021] Wherein, P o is the laser output power, is the Mach-Zehnder modulator DC bias phase shift, β is the fiber dispersion parameter, L is the single-mode optical fiber length, m RF is the radio frequency modulation index, and m+ , m - are up and down sideband modulation indices, respectively, E R (t) is the electric field of the photodetector at the remote radio head, e is the natural constant, ω IF is the angular modulation frequency of the intermediate frequency signal, ω osc is the angular modulation frequency of the microwave signal.

[0022] Optionally, the up and down sideband modulation indices comprise:

[0023]

[0024] wherein a is a linewidth enhancement factor, k is an adiabatic chirp parameter, m a is an amplitude modulation index, j is an imaginary unit.

[0025] Optionally, matching the input feature and the output feature dimensions comprises:

[0026] Collecting the input feature and the output feature, and constructing a training set by using the input feature and the output feature;

[0027] Training the light gradient boosting machine model by using the training set comprises:

[0028] Taking an initial prediction value of each output dimension as a constant of a minimized loss function, and calculating a first-order derivative and a second-order derivative of the loss function by using each training sample in the training set;

[0029] Taking the first-order derivative as a gradient, and reserving a high gradient sample set of a first target number of absolute values of the gradient, randomly selecting a low gradient sample set of a second target number from the remaining samples, and further compensating the gradient of the low gradient samples;

[0030] Using histogram aggregation and feature bundling to accelerate splitting, and constructing a decision tree; in the process of constructing the decision tree, stopping splitting when a splitting gain is lower than a first target value or a second-order derivative sum of sub-nodes is less than a second target value, proving that the training is completed, thereby obtaining the trained light gradient boosting machine model;

[0031] Modeling the millimeter wave signal spectrum of the improved remote cloud radio access network transmitter by using the trained light gradient boosting machine model, and matching the input feature and the output feature dimensions.

[0032] Optionally, the loss function comprises:

[0033]

[0034] wherein l(·) is a mean square error loss function, y ij is a real spectrum label, For model prediction output, u is the number of spectral sampling points.

[0035] Optionally, the first-order derivative and the second-order derivative of the loss function are calculated by:

[0036]

[0037] Wherein, g i is the first-order derivative, h i is the second-order derivative, is the predicted value of the m-1th iteration, y i is the true label, is the partial derivative symbol.

[0038] Optionally, the split gain includes:

[0039]

[0040] Wherein, is the split gain, I L , I R is the left and right child node sample set after splitting, I is the current node sample set, and λ is the L2 regularization coefficient.

[0041] The beneficial effects of the present application are:

[0042] The present application adopts a double-ring optoelectronic oscillator to replace the traditional microwave source, which significantly reduces the phase noise. The double-ring optoelectronic oscillator realizes single-mode oscillation and suppresses side modes through the phase matching mechanism of the long and short fiber loops. Compared with the traditional microwave source, the double-ring optoelectronic oscillator realizes a 27dBc / Hz reduction in phase noise at a 10kHz frequency offset.

[0043] The present application utilizes the synergistic gain behavior of the chirp of the directly modulated laser and the dispersion of the optical fiber to improve the transmission performance of the millimeter wave signal. The synergistic effect of the chirp effect of the directly modulated laser and the dispersion of the optical fiber will produce significant power gain.

[0044] The present application proposes a millimeter wave signal spectrum modeling framework based on a light gradient boosting machine, which solves the problem of input and output dimension mismatch. The light gradient boosting machine automatically learns the complex interaction between the intermediate frequency frequency, the intermediate frequency amplitude and the single-mode optical fiber length and the spectrum profile through the Leaf-wise growth strategy of gradient boosting decision tree, histogram acceleration splitting and gradient-oriented sampling, without the need for manual feature expansion to retain complete spectrum information. Compared with traditional models, the framework greatly improves the training efficiency while maintaining high accuracy.

[0045] The lightweight gradient boosting machine model of the application has excellent accuracy, robustness and high efficiency. The average absolute error of the lightweight gradient boosting machine model to the intermediate frequency sideband is not more than 1.70dB, and the average absolute error to the second harmonic distortion sideband is not more than 0.81dB, which is better than the model of feedforward neural network, recurrent neural network, long short-term memory network and the like; when the intermediate frequency frequency, amplitude and single-mode optical fiber length change, the average absolute error fluctuation is small. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0047] Figure 1 A schematic diagram of a cloud radio access network modeling method based on machine learning for an embodiment of the application;

[0048] Figure 2 A phase noise comparison between a double-ring optoelectronic oscillator and a conventional microwave source for an embodiment of the application;

[0049] Figure 3 A side mode suppression comparison between a double-ring optoelectronic oscillator and a single-ring optoelectronic oscillator for an embodiment of the application; (a) is a side mode suppression comparison between a 2km long ring in a double-ring optoelectronic oscillator and a single-ring optoelectronic oscillator, and (b) is a side mode suppression comparison between a 170m short ring in a double-ring optoelectronic oscillator and a single-ring optoelectronic oscillator;

[0050] Figure 4 A comparison of the frequency spectrum curves of the millimeter wave signal modulated by a 60MHz intermediate frequency signal under a 5km and 35km single-mode optical fiber link for an embodiment of the application;

[0051] Figure 5 A training structure of the lightweight gradient boosting machine model for an embodiment of the application;

[0052] Figure 6 A comparison of the convergence performance during model training and verification for an embodiment of the application; (a) is a training set convergence curve, and (b) is a verification set convergence curve;

[0053] Figure 7 A comparison of the predicted frequency spectrum curve and the actual frequency spectrum curve under the conditions of a 60MHz intermediate frequency frequency, a 440mV pp intermediate frequency amplitude and a 35km single-mode optical fiber link length for an embodiment of the application;

[0054] Figure 8This is a comparison of the mean absolute error of the upper sideband spectrum curves on the test set in this embodiment of the invention;

[0055] Figure 9 This is a comparison of the computation time for the model training and testing processes in an embodiment of the present invention. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0058] like Figure 1 As shown, this embodiment discloses a cloud wireless access network modeling method based on machine learning, including: acquiring a remote cloud wireless access network transmitter; improving the remote cloud wireless access network transmitter by replacing the traditional microwave source of the remote cloud wireless access network transmitter with a dual-ring opto-oscillator to suppress side modes through single-mode oscillation; and introducing a direct-modulation laser after the arbitrary waveform generator to generate power gain by utilizing the chirp effect and fiber dispersion to achieve synergistic gain. The millimeter-wave signal spectrum of the improved remote cloud wireless access network transmitter is modeled based on a lightweight gradient booster model to match the dimensions of input and output features.

[0059] Figure 1Figure 1 is a schematic diagram of a remote cloud wireless access network architecture assisted by a direct modulation laser and an optoelectronic oscillator; CO - central office, ODN - optical distribution network, RRH - remote radio head, DML - direct modulation laser, AWG - arbitrary waveform generator, PC - polarization controller, MZM - Mach-Zehnder modulator, OC1 - first optical coupler, OC2 - second optical coupler, OC3 - third optical coupler, SMF - single mode fiber, PD1 - first photodetector, PD2 - second photodetector, PD3 - third photodetector, LNA - low noise amplifier, BPF - bandpass filter, PA1 - first power amplifier, PA2 - second power amplifier, PNA - phase noise analyzer, EDFA - erbium-doped fiber amplifier, OSA - optical spectrum analyzer, SA - spectrum analyzer. The optical distribution network is responsible for transmitting signals from the central office to the end user. The arbitrary waveform generator generates intermediate frequency signal waveforms, which are then used to modulate the direct modulation laser. The polarization controller adjusts the polarization state of the optical signal to ensure stability and compatibility throughout the transmission process. The bias voltage of the Mach-Zehnder modulator is controlled to achieve a carrier suppression state, thereby obtaining a frequency-doubled radio frequency signal at the third photodetector. The first photodetector and the second photodetector convert the circulating optical signal into an electrical signal. The low noise amplifier amplifies the signal and minimizes the introduction of noise, thereby maintaining signal quality. The bandpass filter ensures signal integrity by filtering out unnecessary frequency components. The first power amplifier further provides gain for the oscillator circuit. The phase noise analyzer is used to observe the phase noise. The erbium-doped fiber amplifier is used to amplify the optical signal to compensate for any losses incurred during fiber transmission. The third photodetector in the remote radio head completes the photoelectric signal conversion for spectral analysis. The second power amplifier further amplifies the millimeter wave signal to meet the requirements of transmission distance and signal strength. The spectrum analyzer is used to analyze the spectral characteristics of the millimeter wave signal, facilitating monitoring and optimizing signal transmission quality. The spectrum analyzer is used to observe the spectrum and carrier suppression.

[0060] Further, suppressing side modes by single-mode oscillation includes: a double-ring optoelectronic oscillator uses a long optical fiber loop and a short optical fiber loop for single-mode oscillation and suppresses side modes.

[0061] Specifically, the core of the remote cloud wireless access network transmitter architecture of the present application includes a central office, an optical distribution network and a remote radio head, and the key components and designs are as follows:

[0062] Double-ring optoelectronic oscillator: single-mode oscillation is achieved by using a long optical fiber loop and a short optical fiber loop, and the formula is as follows:

[0063]

[0064] Where, f osc is the oscillation frequency, k and m are integers, and τl τ is long ring delay s τ is short ring delay. Short ring determines mode spacing, which is raised to simplify mode competition; long ring dominates phase noise suppression, which utilizes high Q value of fiber to reduce phase noise. Experimentally, the phase noise at 10 kHz frequency offset is reduced by 27 dBc / Hz compared with traditional microwave source, and the side mode suppression is raised by 40.83 dB and 12.66 dB for long ring and short ring optoelectronic oscillators, respectively.

[0065] Further, the improved remote cloud radio access network transmitter is operated in a carrier suppression mode, the sideband power is enhanced, the optical signal is transmitted through a single-mode fiber, the loss is compensated by an erbium-doped fiber amplifier, and the optical signal is converted into a millimeter wave electrical signal by a photodetector at a remote radio head.

[0066] Specifically, the signal generation and transmission link: an arbitrary waveform generator generates an intermediate frequency signal, which drives a directly modulated laser to generate an optical modulation signal, and the optical field expression of the directly modulated laser is:

[0067]

[0068] wherein m is the amplitude modulation index, ω is the intermediate frequency signal angular frequency, ω is the laser carrier frequency angular frequency, a IF o is the maximum phase shift caused by intermediate frequency modulation. The Mach-Zehnder modulator is operated in a carrier suppression mode, the sideband power is enhanced, the optical signal is transmitted through a single-mode fiber, the loss is compensated by an erbium-doped fiber amplifier, and the optical signal is converted into a millimeter wave electrical signal by a photodetector at a remote radio head, and the electric field expression of the photodetector in the remote architecture is:

[0069]

[0070] wherein P is the laser output power, o is the Mach-Zehnder modulator DC bias phase shift (set to π / 2 to achieve carrier suppression), β is the fiber dispersion parameter, L is the single-mode fiber length, m is the radio frequency modulation index, m and m are the upper and lower sideband modulation indices, respectively, and the expressions are: RF + -

[0071]

[0072] wherein α is the linewidth enhancement factor, and κ is the adiabatic chirp parameter. After amplification by a power amplifier, it is monitored by a spectrum analyzer.

[0073] ​​​​​​​Further, the matching of the input feature and the output feature dimension includes: collecting the input feature and the output feature, and constructing a training set by using the input feature and the output feature; training the light gradient boosting machine model by using the training set includes: taking the initial prediction value of each output dimension as a constant of a minimum loss function, and calculating the first-order derivative and the second-order derivative of the loss function by using each training sample in the training set; taking the first-order derivative as a gradient, and retaining a first target number of high-gradient sample sets in the absolute value of the gradient; a second target number of low-gradient sample sets are randomly selected from the remaining samples, and the gradient of the low-gradient samples is further compensated; histogram aggregation and feature bundling are used to accelerate splitting to construct a decision tree; in the process of constructing the decision tree, when the splitting gain is lower than the first target value or the second-order derivative of the sub-node is less than the second target value, the splitting is stopped, and the trained light gradient boosting machine model is obtained; the millimeter wave signal spectrum of the improved remote cloud radio access network transmitter is modeled by using the trained light gradient boosting machine model, and the input feature and the output feature dimension are matched.

[0074] Specifically, the millimeter wave signal spectrum modeling based on the light gradient boosting machine is as shown in the following formula: Figure 5

[0075] The light gradient boosting machine model is used to model the spectrum profile of the millimeter wave signal, and the specific process is as follows:

[0076] Input and output feature definition: input feature: intermediate frequency signal frequency 11-60MHz, intermediate frequency signal amplitude 210-500mV pp , single-mode optical fiber link length 5-35km, which is organized into an input matrix:

[0077] x i =[F i ,A i ,L i ] T i=1,2,…,n (6)

[0078] Wherein, F i is the intermediate frequency of the i-th sample, A i is the intermediate frequency amplitude of the i-th sample, and L i is the optical fiber length of the i-th sample.

[0079] Output feature: millimeter wave spectrum profile collected by a spectrum analyzer, including sampling points of intermediate frequency upper and lower sidebands, second harmonic distortion upper and lower sidebands and background noise, which is organized into an output matrix:

[0080] y ij =[s i1 ,s i2 ,…,s ij ] T ​i = 1, 2, …, n; j = 1, 2, …, u (7)

[0081] where s ij is the jth spectral sample point of the ith sample, and u is the total number of sample points.

[0082] Dataset construction: 10500 samples were collected by a spectrum analyzer, and divided into training set, validation set and test set according to the ratio of 7:2:1.

[0083] Lightweight gradient boosting machine model training:

[0084] Initialization: for each output dimension, the initial prediction value is a constant that minimizes the loss function, and the loss function uses mean square error:

[0085]

[0086] The initial prediction value calculation formula is:

[0087]

[0088] where γ is the constant to be optimized, and l(·) is the above-mentioned mean square error loss function.

[0089] Hyperparameter setting: the number of boosting iterations M = 100, the maximum number of leaf nodes K = 20, the learning rate η = 0.1, the gradient-oriented sampling ratio a = 0.2, b = 0.1, the L2 regularization coefficient λ = 0.5, the minimum split gain γ min = 0.1, and the minimum Hessian H min = 1.0.

[0090] Boosting iteration:

[0091] Gradient and Hessian calculation: for each sample, calculate the first-order derivative (gradient) and second-order derivative (Hessian) of the loss function:

[0092]

[0093] where, is the prediction value of the m-1th iteration.

[0094] Sample sampling: select samples by gradient-oriented sampling strategy: keep a|D|high gradient sample set A with the largest gradient absolute value, randomly select b|D|low gradient sample set R from the remaining samples, and compensate the gradient of low gradient samples:

[0095]

[0096] where g r is the scaled gradient value of sample r, and gˊ rThe original gradient value of the sample r, The final iteration sample set is S m for the compensation coefficient, r is the sample in the random subset, is a random subset, a is the high gradient sample retention ratio, b is the random subset sampling ratio, and a is the high gradient sample subset.

[0097] Decision tree construction: histogram aggregation and feature bundling are used to accelerate splitting, and the histogram calculation formula of the feature bundle is:

[0098] hist b = ∑ f∈b hist f (13)

[0099] where hist f is the histogram of a single feature f, and b is the feature bundle. The split gain calculation formula is:

[0100]

[0101] where I is the current node sample set, I L and I R are the sample sets of the left and right child nodes after splitting.

[0102] Leaf node value calculation: when the split gain is lower than γ min or the child node Hessian sum is less than H min , stop splitting, and the calculation formula of the leaf node value v is:

[0103]

[0104] Prediction update: after each iteration, the prediction value update formula is:

[0105]

[0106] where v (m) (x i ) is the leaf node value assigned by the mth decision tree to the sample x i , and η is the learning rate.

[0107] Model performance verification:

[0108] (1) Accuracy verification: on the test set, the average absolute error of the lightweight gradient boosting machine on the intermediate frequency sideband is 0.98-1.30 dB, and the average absolute error on the second harmonic distortion sideband is 0.75-1.55 dB, which is significantly better than the feedforward neural network model.

[0109] (2) Robustness verification: When the intermediate frequency frequency, amplitude, and fiber length change, the average absolute error of the light gradient boosting machine fluctuates by 0.32 dB, 0.29 dB, and 0.80 dB, respectively, and the stability is better than that of the comparative model.

[0110] (3) Efficiency verification: The training time is 26.09 seconds, which is much faster than bidirectional long short-term memory network, extreme gradient boosting, and other models; the test time is 447 milliseconds, which meets the actual deployment requirements.

[0111] The double-ring optoelectronic oscillator of the application replaces the traditional microwave source, and significantly reduces the phase noise. The double-ring optoelectronic oscillator realizes single-mode oscillation and suppresses side modes through the phase matching mechanism of the long and short fiber loops.

[0112] As shown in Figure 2 Compared with the traditional microwave source, the double-ring optoelectronic oscillator realizes a 27dBc / Hz phase noise reduction at a 10kHz frequency offset.

[0113] The application is significantly better than the single-loop structure in side mode suppression performance, as shown in Figure 3 (a)-(b), under the configuration of a 2km long loop, the side mode suppression is improved by 40.83dB compared with the single-loop optoelectronic oscillator; under the configuration of a 170m short loop, the side mode suppression is improved by 12.66dB.

[0114] The application utilizes the synergistic gain behavior of the chirp of the directly modulated laser and the dispersion of the optical fiber to improve the transmission performance of the millimeter wave signal. The synergistic effect of the chirp of the directly modulated laser and the dispersion of the optical fiber will produce significant power gain. As shown in Figure 4 The millimeter wave signal gain under 35km single-mode fiber transmission is improved by 19dB compared with 5km, which is better than the attenuation characteristics of the local architecture.

[0115] The application proposes a millimeter wave signal spectrum modeling framework based on a light gradient boosting machine, as shown in Figure 5 The light gradient boosting machine automatically learns the complex interaction between the intermediate frequency frequency, the intermediate frequency amplitude, and the single-mode fiber length and the spectrum profile through the Leaf-wise growth strategy of gradient boosting decision trees, histogram acceleration splitting, and gradient-oriented sampling, without the need for manual feature expansion to retain complete spectrum information. Compared with traditional models, the framework greatly improves the training efficiency while maintaining high accuracy.

[0116] The light gradient boosting machine model used in the application exhibits excellent and stable convergence performance during the training process. As shown in Figure 6As shown in (a)-(b), the model loss decreases rapidly and tends to be stable with the increase of the number of iterations, whether in the training set or the validation set, and no overfitting phenomenon occurs, proving the effectiveness of the training strategy and the generalization ability of the model.

[0117] The present application has high-precision prediction capability, for example, Figure 7 As shown in the figure, under the conditions of 60MHz intermediate frequency frequency, 440mVpp intermediate frequency amplitude and 35km single-mode optical fiber link length, the spectrum curve predicted by the lightweight gradient boosting machine model is highly consistent with the actually measured spectrum curve, and the effect is obviously better than that of the feedforward neural network.

[0118] The lightweight gradient boosting machine model of the present application has excellent accuracy, robustness and high efficiency. Figure 8 As shown in the figure, the average absolute error of the model for the intermediate frequency sideband is not more than 1.70dB, and the average absolute error for the second harmonic distortion sideband is not more than 0.81dB, which is better than the feedforward neural network, recurrent neural network, long short-term memory network and other models; when the intermediate frequency, amplitude and single-mode optical fiber length change, the average absolute error fluctuation is small.

[0119] The lightweight gradient boosting machine modeling framework proposed by the present application has advantages in efficiency, for example, Figure 9 As shown in the figure, the model training time is only 26.09 seconds, and the test time is only 447 milliseconds, and the comprehensive time consumption is lower than other traditional machine learning models and neural network models, meeting the requirements of modeling efficiency in actual engineering applications.

[0120] The above-described embodiments are only descriptions of the preferred modes of the present application, and do not limit the scope of the present application, and various modifications and improvements to the technical solutions of the present application made by those skilled in the art without departing from the design spirit of the present application shall fall within the protection scope determined by the claims of the present application.

Claims

1.A method for cloud radio access network modeling based on machine learning, characterized in that, The application relates to a remote cloud radio access network transmitter, and improvement of the remote cloud radio access network transmitter. A double-ring optical-electric oscillator is used to replace a traditional microwave source of the remote cloud radio access network transmitter, for suppressing side modes through single-mode oscillation, and a direct modulation laser is introduced after an arbitrary waveform generator, for generating power gain by using a chirp effect and fiber dispersion synergy gain. A millimeter wave signal spectrum of the improved remote cloud radio access network transmitter is modeled based on a light gradient boosting machine model, and input features and output features are matched in dimension. Suppression of side modes through single-mode oscillation includes: 2.The method of claim 1, wherein, The double-ring optical-electric oscillator uses a long optical fiber loop and a short optical fiber loop for single-mode oscillation and suppression of side modes. An optical field expression of the direct modulation laser is: where f osc is the oscillation frequency, k, m are integers, τ l is the long loop delay, τ s is the short loop delay. 3.The method of claim 1, wherein, A Mach-Zehnder modulator of the improved remote cloud radio access network transmitter works in a carrier suppression mode, sideband power is enhanced, and an optical signal is transmitted through a single-mode optical fiber, loss is compensated by an erbium-doped fiber amplifier, and the optical signal is converted into a millimeter wave electrical signal by an optical-electric detector at a remote radio head; where m a is the amplitude modulation index, ω IF is the intermediate frequency signal angular frequency, ω o is the laser carrier frequency angular frequency, is the maximum phase shift due to the intermediate frequency modulation, E DML (t) is the optical field of the directly modulated laser, t is time. 4.The method of claim 1, wherein, An electrical field expression of the optical-electric detector at the remote radio head is: The upper and lower sideband modulation indexes include: where P o is the laser output power, is the Mach-Zehnder modulator DC bias phase shift, β is the fiber dispersion parameter, L is the single mode fiber length, m RF is the radio frequency modulation index, m + , m - are the upper and lower sideband modulation indices, E R (t) is the electric field of the photodetector at the remote radio head, e is the natural constant, ω IP is the angular modulation frequency of the intermediate frequency signal, ω OSC is the angular modulation frequency of the microwave signal. 5.The method of claim 4, wherein, Matching the input features and the output features in dimension includes: where a is the linewidth enhancement factor and k is the adiabatic chirp parameter a is the amplitude modulation index and j is the imaginary unit. 6.The method of claim 1, wherein, Input features and output features are collected, and the input features and the output features are used to build a training set; Training the light gradient boosting machine model using the training set includes: An initial prediction value of each output dimension is taken as a constant of a minimum loss function, and a first-order derivative and a second-order derivative of the loss function are calculated using each training sample in the training set; The first-order derivative is taken as a gradient, a high-gradient sample set of a first target number is reserved from the gradient absolute value, a low-gradient sample set of a second target number is randomly selected from the remaining samples, and the gradient of the low-gradient samples is further compensated; A decision tree is built by using histogram aggregation and feature bundling to accelerate splitting; in the process of building the decision tree, splitting is stopped when a splitting gain is lower than a first target value or a second-order derivative and of a child node is less than a second target value, training is proved to be completed, and thus a trained light gradient boosting machine model is obtained; The millimeter wave signal spectrum of the improved remote cloud radio access network transmitter is modeled using the trained light gradient boosting machine model, and the input features and the output features are matched in dimension. The loss function includes: 7.The method of claim 6, wherein, The first-order derivative and the second-order derivative of the loss function are calculated; where l(·) is the mean square error loss function, y ij is the true spectral label, is the model prediction output, and u is the number of spectral sample points. 8.The method of claim 6, wherein, The splitting gain includes: where g i is the first derivative, h i is the second derivative, is the prediction value of the m-1th iteration, y i is the true label, is the partial derivative sign. 9.The method of claim 6, wherein, ​ wherein, is the split gain, I L , R are the left and right child sample sets after splitting, I is the current node sample set, and λ is the L2 regularization coefficient.

Images