CN114978313B - Compensation method of visible light CAP system based on Bayesian neurons - Google Patents
Compensation method of visible light CAP system based on Bayesian neurons Download PDFInfo
- Publication number
- CN114978313B CN114978313B CN202210539304.XA CN202210539304A CN114978313B CN 114978313 B CN114978313 B CN 114978313B CN 202210539304 A CN202210539304 A CN 202210539304A CN 114978313 B CN114978313 B CN 114978313B
- Authority
- CN
- China
- Prior art keywords
- network parameter
- bayesian
- network
- signal
- neurons
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 210000002569 neuron Anatomy 0.000 title claims abstract description 66
- 238000000034 method Methods 0.000 title claims abstract description 25
- 238000013135 deep learning Methods 0.000 claims abstract description 13
- 238000012549 training Methods 0.000 claims description 53
- 230000006870 function Effects 0.000 claims description 41
- 238000003062 neural network model Methods 0.000 claims description 37
- 238000009826 distribution Methods 0.000 claims description 21
- 238000012360 testing method Methods 0.000 claims description 19
- 230000008569 process Effects 0.000 claims description 7
- 230000004913 activation Effects 0.000 claims description 6
- 238000007781 pre-processing Methods 0.000 claims description 5
- 238000013528 artificial neural network Methods 0.000 abstract description 12
- 238000004891 communication Methods 0.000 abstract description 10
- 230000005540 biological transmission Effects 0.000 abstract description 9
- 230000003287 optical effect Effects 0.000 description 5
- 238000005070 sampling Methods 0.000 description 5
- 238000005516 engineering process Methods 0.000 description 4
- 238000013531 bayesian neural network Methods 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 108010076504 Protein Sorting Signals Proteins 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 2
- 238000012937 correction Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 230000003446 memory effect Effects 0.000 description 2
- 230000004075 alteration Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000009795 derivation Methods 0.000 description 1
- 230000004807 localization Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 230000009022 nonlinear effect Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000001105 regulatory effect Effects 0.000 description 1
- 238000007493 shaping process Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 230000001629 suppression Effects 0.000 description 1
- 238000013076 uncertainty analysis Methods 0.000 description 1
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04B—TRANSMISSION
- H04B10/00—Transmission systems employing electromagnetic waves other than radio-waves, e.g. infrared, visible or ultraviolet light, or employing corpuscular radiation, e.g. quantum communication
- H04B10/11—Arrangements specific to free-space transmission, i.e. transmission through air or vacuum
- H04B10/114—Indoor or close-range type systems
- H04B10/116—Visible light communication
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/243—Classification techniques relating to the number of classes
- G06F18/2431—Multiple classes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/047—Probabilistic or stochastic networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04B—TRANSMISSION
- H04B10/00—Transmission systems employing electromagnetic waves other than radio-waves, e.g. infrared, visible or ultraviolet light, or employing corpuscular radiation, e.g. quantum communication
- H04B10/50—Transmitters
- H04B10/516—Details of coding or modulation
- H04B10/54—Intensity modulation
- H04B10/541—Digital intensity or amplitude modulation
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L25/00—Baseband systems
- H04L25/02—Details ; arrangements for supplying electrical power along data transmission lines
- H04L25/03—Shaping networks in transmitter or receiver, e.g. adaptive shaping networks
- H04L25/03006—Arrangements for removing intersymbol interference
- H04L25/03165—Arrangements for removing intersymbol interference using neural networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D30/00—Reducing energy consumption in communication networks
- Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Computer Networks & Wireless Communication (AREA)
- Signal Processing (AREA)
- Probability & Statistics with Applications (AREA)
- Electromagnetism (AREA)
- Power Engineering (AREA)
- Optical Communication System (AREA)
Abstract
The invention belongs to the technical field of visible light communication, and particularly relates to a compensation method of a visible light CAP system based on Bayesian neurons; the method comprises the following steps: in a visible light CPA system, a signal received by downsampling at a receiving end is input into a deep learning nonlinear compensation module based on Bayesian neurons for nonlinear compensation, and a compensated QAM signal is obtained; performing QAM demapping on the compensated QAM signal to obtain an equalization signal, and realizing nonlinear compensation on a visible light CAP system; the invention can compensate the received distortion signal into the normal signal before transmission, improves the transmission speed of the VLC system and improves the overall performance of the system, and the weight parameter of the neural network in the invention is a random variable instead of a determined value, can give out the uncertainty of prediction, prevents over fitting and has very strong robustness on nonlinear compensation.
Description
Technical Field
The invention belongs to the technical field of visible light communication, and particularly relates to a compensation method of a visible light CAP system based on Bayesian neurons.
Background
A visible light communication technology (Visible Light Communication, VLC) is used as a novel communication mode in optical wireless communication, electromagnetic waves in a visible light band are used as an information carrier, and high-speed bright-dark changing light signals which are emitted by LEDs and are indistinguishable by naked eyes are used for transmitting information. VLC has the advantages of no license operation, relatively low cost, high space diversity, high bandwidth efficiency, no electromagnetic interference transmission and the like, and can make up for the defect of wireless frequency bandwidth of overcrowded wireless communication. However, in VLC systems, there is a wide range of nonlinear distortion problems that severely compromise the overall performance of the system. The equalization technology is a technology for effectively reducing intersymbol interference and improving communication quality by compensating and correcting the transmission characteristics of a channel. The equalization technology of the existing VLC is developed in several stages, wherein the first stage is to utilize a hardware equalizer to improve the LED modulation bandwidth; the second stage is a traditional adaptive-based linear equalization technique, such as Constant Modulus Algorithm (CMA), cascaded multimode algorithm (CMMA), least mean square error (LMS), recursive Least Squares (RLS), etc.; the third stage is to add non-linear compensation in the adaptive equalizer, such as a Volterra-based non-linear post equalizer and a look-up table (LUT) of shorter Taps (Taps), etc. However, none of the above methods can well mitigate nonlinear distortion. Until the fourth stage, deep learning and conventional machine learning have begun to be applied to VLC systems, such as K-means, DBSCAN, ANN, LSTM, etc., such that significant breakthroughs have been made in solving the problem of non-linearities in VLC systems.
The patent "a nonlinear suppression method of visible light communication system based on LSTM (application No. cn202110906991. X)" proposes to suppress nonlinear effects and memory effects of the visible light system by using the memory effect of the LSTM predistortion network, but does not consider using the statistical rule of the received signal to compensate the signal. The patent 'a visible light communication method, a device, a system and a computer readable storage medium (application number CN 202110528061.5)' proposes a method for compensating nonlinear distortion of an electric signal by utilizing a post-equalization algorithm of an artificial neural network. The neural networks proposed in the above patents all adopt fixed weight values during training, only the predicted signal values can be given, and the confidence of the predicted signals cannot be given. A large amount of training signal data is often required, and there is a serious overfitting phenomenon with a small amount of signal data. When compensation is performed on the signal which does not appear, the uncertainty in the training signal data cannot be accurately estimated, so that excessive confidence is caused, the signal compensation effect is reduced, and the generalization capability is poor.
Disclosure of Invention
Aiming at the defects existing in the prior art, the invention provides a compensation method of a visible light CAP system based on Bayesian neurons, which comprises the following steps: in a visible light CAP system, a signal received by downsampling at a receiving end is input into a deep learning nonlinear compensation module based on Bayesian neurons for nonlinear compensation, and a compensated QAM signal is obtained; performing QAM demapping on the compensated QAM signal to obtain an equalization signal, and realizing nonlinear compensation on a visible light CAP system;
the deep learning nonlinear compensation module based on the Bayesian neuron comprises a trained neural network model based on the Bayesian neuron, and the process for training the neural network model based on the Bayesian neuron comprises the following steps:
s1: acquiring an original signal data set, preprocessing the original signal data set, and obtaining a training set and a label set of the training set;
s2: initializing network parameters and setting ideal values of a loss function; wherein the network parameters include a first network parameter, a second network parameter, and a third network parameter;
s3: initializing a variation parameter according to the second network parameter and the third network parameter, and defining a minimized variation distribution according to the variation parameter and the first network parameter;
s4: inputting the training set and the label set of the training set into a neural network model based on Bayesian neurons for training;
s5: calculating a loss function of a neural network model based on Bayesian neurons according to the first network parameters and the minimized variation distribution;
s6: judging whether the loss function reaches an ideal value of the loss function, if so, storing network parameters and outputting a neural network model corresponding to the network parameters, otherwise, updating the network parameters by adopting Bayesian back propagation, and returning to the step S4.
Preferably, the Bayesian neuron-based neural network model comprises an input layer, a hidden layer and an output layer; the input layer is an 11-dimensional vector; the hidden layer contains a 2-layer structure and is composed of 64 neurons and 32 neurons which are connected completely; the output layer is a 16-dimensional vector.
Furthermore, the hiding layer adopts a ReLU function as an activation function, and the output layer adopts a softmax function as an activation function;
the formula of the ReLU function is:
the softmax function is formulated as:
wherein y represents the output of the hidden layer, x represents the input of the hidden layer, and P i Representing the probability of the output signal level of the ith neuron, y i Representing the output of the ith neuron of the hidden layer, y k The output of the kth neuron of the hidden layer is represented, and n represents the number of decision classifications.
Preferably, the preprocessing of the raw signal data set comprises:
acquiring signal data of a transmitting end and signal data of a receiving end of a CAP visible light system; receiving end signal data according to 7:3, dividing the ratio into a training set sample and a test set sample; transmitting end signal data according to 7:3 into a label signal sample of a training set and a label signal sample of a test set;
dividing the training set sample and the test set sample by adopting a sliding window to obtain a training set and a test set;
dividing a label signal sample of the training set and a label signal sample of the test set by adopting a sliding window to obtain the label signal sample set of the training set and the label signal sample set of the test set;
and taking the middle number of the label signal sample set of the training set to form a label set of the training set.
Preferably, the first network parameters are:
ω＝μ+log(1+e ρ )*ε，ε～N(0，1)
wherein ω represents a first network parameter, μ represents a second network parameter, ρ represents a third network parameter, ε -N (0, 1) represents a random variable ε obeys a standard normal distribution.
Preferably, the loss function of the bayesian neuron-based neural network model is:
L(ω|θ)＝H+αKL[q(ω|θ)||P(ω)]
wherein L (omega|theta) represents the loss of the neural network model, H represents the cross entropy loss of the model, alpha represents the dynamic attenuation coefficient, KL represents the divergence loss, q (omega|theta) represents the variation distribution, P (omega) represents the prior probability of the first network parameter, and P i Representing the probability of the output signal level of the ith neuron, t i Representing the target classification level.
Preferably, the process of updating the network parameters includes:
calculating a gradient of the second network parameter and a gradient of the third network parameter;
updating the second network parameter according to the gradient of the second network parameter, and updating the third network parameter according to the gradient of the third network parameter;
the first network parameter is updated according to the new second network parameter and the third network parameter.
Further, the formula for updating the second network parameter is:
μ * ＝μ-ηΔ μ
wherein delta is μ Representing the gradient of the second network parameter, L (ω, θ) representing the loss of the neural network model, ω representing the first network parameter, μ representing the second network parameter, μ * Representing the updated second network parameter, η representing the learning rate.
Further, the formula for updating the third network parameter is:
ρ * ＝ρ-ηΔ ρ
wherein delta is ρ A gradient representing a third network parameter, ρ representing the third network parameter, ε -N (0, 1) representing that the random variable ε obeys a standard normal distribution, L (ω, θ) representing the loss of the neural network model, ω representing the first network parameter, η representing the learning rate, ρ * Representing the updated third network parameter.
The beneficial effects of the invention are as follows: the invention designs a Bayesian neuron-based visible light CAP system compensation method, which adopts a Bayesian neuron-based deep learning nonlinear compensation module to realize the correction effect on received signals, and a Bayesian neuron-based neural network model in the module can perform feature extraction on the received signals and the relation among the signals and give out the confidence coefficient of the prediction result of the received signals, so that the input is correctly classified as corresponding CAP-16 signals to the greatest extent; in a visible light CAP system with serious linear and nonlinear damage, the invention can compensate the received distortion signal into a normal signal before transmission, thereby improving the transmission speed of the VLC system and improving the overall performance of the system; different from the existing simple neural network nonlinear compensation model, the weight parameter of the neural network, namely the first network parameter, is a random variable instead of a determined value, so that uncertainty of prediction can be given, overfitting is prevented, and the neural network nonlinear compensation model has very strong robustness in nonlinear compensation.
Drawings
FIG. 1 is a block diagram of a Bayesian neuron-based deep learning visible light CAP system in accordance with the present invention;
FIG. 2 is a schematic diagram of a neural network model based on Bayesian neurons in the present invention;
FIG. 3 is a flowchart of a Bayesian neuron-based neural network model training in accordance with the present invention.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
The invention provides a compensation method of a visible light CAP system based on Bayesian neurons, which considers the statistical rule of received signals and the over-fitting problem of a neural network, and can improve the generalization capability of the neural network; the method comprises the following steps:
in a visible light CAP system, a signal received by downsampling at a receiving end is input into a deep learning nonlinear compensation module based on Bayesian neurons for nonlinear compensation, and a compensated QAM signal is obtained; performing QAM demapping on the compensated QAM signal to obtain an equalization signal, and realizing nonlinear compensation on a visible light CAP system; wherein QAM is quadrature amplitude modulation.
As shown in FIG. 1, at the transmitting end of the visible light CAP system, firstly, the input original binary bit stream is QAM mapped, then the generated complex sequence takes real part and imaginary part through up-sampling and I/Q modulator, and the real part data is taken as in-phase component (I path), the imaginary part data is taken as quadrature componentQuantity (Q path), two paths of signals are respectively passed through shaping filter f I(t) And f Q(t) The signal obtained by subtracting the output signals of the filter is subjected to a hardware pre-equalizer and an amplifier, then DC bias is increased through a bias device, then the LED is used for intensity modulation, the electric signal is converted into an optical signal, and the optical intensity is regulated through a diaphragm and a filter. At the receiving end, the photo diode PIN receives the optical signal, the optical signal is converted into an electric signal in the photo diode, and then the electric signal passes through the amplifier and the low-pass filter and then passes through the two paths of matched filters m I(t) And m Q(t) And down sampling, inputting the signals into a post-equalization module of the deep learning visible light system based on Bayesian neurons for signal equalization, and finally converting the output result of the module into a binary bit stream through QAM demapping so as to realize data transmission.
The deep learning nonlinear compensation module based on the Bayes neurons comprises a trained neural network model based on the Bayes neurons; the construction of the Bayesian neuron-based neural network model comprises the following contents:
bayesian neural networks use probabilistic estimates instead of point estimates, enabling bayesian localization of network parameters while providing the possibility of uncertainty analysis. Therefore, the Bayesian neural network fits posterior distribution, which is different from the traditional neural network which fits label values by using cross entropy, mean square error and other loss functions, so that the overfitting can be reduced. Based on the idea of probability estimation, as shown in fig. 2, the invention constructs a neural network model based on Bayesian neurons, wherein the neural network model comprises: an input layer, a hidden layer and an output layer; the input layer is an 11-dimensional sequence vector; the hidden layer contains a 2-layer structure with parameter prior applied, and consists of 64 neurons and 32 neurons which are connected completely; the output layer is a 16-dimensional vector, which is a result of obtaining the corresponding CAP-16 level classification by a softmax function after full connection, and is classified as "+ -15, + -13, + -11, + -9, + -7, + -5, + -3, + -1", i.e. the neural network model is a 16-multi-classifier.
In a neural network model based on Bayesian neurons, a hidden layer adopts a ReLU function as an activation function, and an output layer adopts a softmax function as an activation function; the formula of the ReLU function is:
the softmax function is formulated as:
wherein y represents the output of the hidden layer, x represents the input of the hidden layer, and P i Representing the output of the ith neuron, i.e. the signal level probability, y i Representing the output of the ith neuron of the hidden layer, y k Representing the output of the kth neuron of the hidden layer; n represents the number of neurons to be finally output, namely the number of final decision classifications, and in the present invention, n=16 is adopted by the CAP-16 modulation scheme.
As shown in fig. 3, the process of training the bayesian neuron-based neural network model includes:
s1: the method comprises the steps of obtaining an original signal data set, and preprocessing the original signal data set to obtain a training set and a label set of the training set.
Acquiring signal data of a transmitting end and signal data of a receiving end of a CAP visible light system; receiving end signal data according to 7:3, dividing the ratio into a training set sample and a test set sample; transmitting end signal data according to 7:3 into a label signal sample of a training set and a label signal sample of a test set;
dividing the training set sample and the test set sample by adopting a sliding window to obtain a training set and a test set; specifically, the signal sequence of the training set sample is { x } 1 ，x 2 ，...，x n The data length is n, the number of taps t=11 is selected as the sliding window size, and the data of the ith division is { x } i ，x i+1 ，…，x i+t-1 Obtaining a data vector; similarly, the test set samples are divided accordingly;
by slidingDividing a label signal sample of the training set and a label signal sample of the test set by a dynamic window to obtain the label signal sample set of the training set and the label signal sample set of the test set; taking the middle number of the label signal sample set of the training set to form a label set of the training set; specifically, the signal sequence of the label signal sample of the training set is { y } 1 ，y 2 ，...，y n Data length n, data of ith sub-division { y }, data of i ，y i+1 ，…，y i+t-1 Intermediate number y of the ith divided data } i+(t-1)/2 As training tag values for the set of cut data, the training tag values for each set of data constitute a tag set.
S2: initializing network parameters and setting ideal values of a loss function; wherein the network parameters include a first network parameter, a second network parameter, and a third network parameter.
Setting a first network parameter omega to obey Gaussian distribution, namely omega-N (mu, sigma); wherein σ=log (1+e ρ ) μ is a second network parameter, ρ is a third network parameter; specifically, random sampling (point-by-point multiplication) is performed on the network parameters to obtain ω, namely:
ω＝μ+log(1+e ρ )*ε，ε～N(0，1)
wherein ε -N (0, 1) represents that the random variable ε obeys a standard normal distribution.
The ideal value of the loss function is set, for example, the ideal value of the loss function may be 0.001.
S3: initializing a variation parameter according to the second network parameter and the third network parameter, and defining a minimized variation distribution according to the variation parameter and the first network parameter.
The variation parameter is expressed as θ= (μ, ρ) according to the second network parameter and the third network parameter; defining a minimized variation distribution q (omega|theta) to approximate the true posterior probability P (omega|d) based on the variation parameters and the first network parameters; wherein, D= { X, Y }, X represents input data, namely receiving end data needing to be balanced in the CAP-VLC system; y represents the label data, i.e., the label value corresponding to each set of input data in the CAP-VLC system.
S4: the training set and the label set of the training set are input into a neural network model based on Bayesian neurons for training.
And inputting the data of the training set and the label value corresponding to the training set in the label set into a neural network model based on Bayesian neurons for training.
S5: calculating a loss function of a neural network model based on Bayesian neurons according to the first network parameters and the minimized variation distribution; determining the loss function includes the following:
the probability of predicting the equalized signal level is expressed as:
P(Y * |X * ，D)＝∫P(Y * |X * ，ω)P(ω|D)dω
wherein Y is * Representing the predicted value, X * Representing the input data.
Since ω is a random variable, the predicted value is also a random variable; wherein:
where P (ω|d) represents the posterior distribution, P (D) represents the edge likelihood, and P (d|ω) represents the likelihood function.
The core of probability prediction by adopting a Bayesian neural network is to make efficient approximate posterior inference, and the invention adopts variation inference to fit posterior probability.
Let the variation parameter θ= (μ, σ), each weight ω i From normal distribution (mu) i ，σ i ) And (3) sampling. It is desirable to minimize that the variation distribution q (ωθ) is close to the true posterior probability P (ωd) and to use the KL divergence to measure the distance between these two distributions. That is, optimizing θ:
wherein θ * Representing the optimized variation parameter theta.
Further derivation:
from the above equation, the optimization is divided into two parts, one part is to minimize the KL divergence of the estimation posterior and the prior; another part is the desire to maximize the log-likelihood with respect to the estimation posterior.
Therefore, the present invention designs the loss function as:
L(ω|θ)＝H+αKL[q(ω|θ)||P(ω)]
wherein L (ω|θ) represents the loss of the neural network model, and H represents the cross entropy loss of the model; alpha represents a dynamic attenuation coefficient, preferably 0.01; KL represents the loss of divergence, q (ω|θ) represents the variation distribution, and P (ω) represents the prior probability of the first network parameter; p (P) i Representing the output signal level probability of the ith neuron; t is t i Representing the target classification level, t of the target class i Equal to 1, t of other classes i Equal to 0.
S6: judging whether the loss function reaches an ideal value of the loss function, if so, storing network parameters and outputting a neural network model corresponding to the network parameters, otherwise, updating the network parameters by adopting Bayesian back propagation, and returning to the step S4.
Judging whether the loss function reaches an ideal value of the loss function, if so, storing network parameters and outputting a neural network model corresponding to the network parameters, wherein the model is a trained neural network model based on Bayesian neurons; otherwise, adopting Bayesian back propagation to update the network parameters, wherein the process for updating the network parameters comprises the following steps:
calculating a gradient of the second network parameter and a gradient of the third network parameter; the formula is:
wherein delta is μ Representing the gradient of the second network parameter, L (ω, θ) representing the loss of the neural network model, Δ ρ Representing the gradient of the third network parameter.
Updating the second network parameter according to the gradient of the second network parameter, and updating the third network parameter according to the gradient of the third network parameter; the formula is:
μ * ＝μ-ηΔ μ
ρ * ＝ρ-ηΔ ρ
wherein μ represents the second network parameter, μ * Represents the updated second network parameter, eta represents the learning rate, ρ represents the third network parameter, ρ * Representing the updated third network parameter.
According to the new second network parameter mu * And a third network parameter ρ * Updating the first network parameters:
and inputting the signal data received by the downsampling of the receiving end into a deep learning nonlinear compensation module based on Bayesian neurons for nonlinear compensation, obtaining a compensated QAM signal, performing QAM demapping on the compensated QAM signal, obtaining an equalization signal, and realizing nonlinear compensation on a visible light CAP system.
The invention applies the post-deep learning visible light system equalization module based on Bayesian neurons to the down-sampling module to realize the correction effect on the received signals, thereby realizing the nonlinear compensation of the CAP-VLC system; the neural network model based on Bayesian neurons in the module can perform feature extraction on received signals and the relation between the signals, give out confidence level of the prediction result of the received signals, and correctly classify the input into corresponding CAP-16 signals as far as possible; in a visible light CAP system with serious linear and nonlinear damage, the invention can compensate the received distortion signal into a normal signal before transmission, thereby improving the transmission speed of the VLC system and improving the overall performance of the system; different from the existing simple neural network nonlinear compensation model, the weight parameters of the neural network in the invention are random variables, but not determined values, so that uncertainty of prediction can be given, overfitting is prevented, and the neural network nonlinear compensation model has very strong robustness in nonlinear compensation.
While the foregoing is directed to embodiments, aspects and advantages of the present invention, other and further details of the invention may be had by the foregoing description, it will be understood that the foregoing embodiments are merely exemplary of the invention, and that any changes, substitutions, alterations, etc. which may be made herein without departing from the spirit and principles of the invention.
Claims (3)
1. A method for compensating a bayesian neuron-based visible light CAP system, comprising: in a visible light CAP system, a signal received by downsampling at a receiving end is input into a deep learning nonlinear compensation module based on Bayesian neurons for nonlinear compensation, and a compensated QAM signal is obtained; performing QAM demapping on the compensated QAM signal to obtain an equalization signal, and realizing nonlinear compensation on a visible light CAP system;
the deep learning nonlinear compensation module based on the Bayesian neurons comprises a trained neural network model based on the Bayesian neurons, wherein the neural network model based on the Bayesian neurons comprises an input layer, a hidden layer and an output layer; the input layer is an 11-dimensional vector; the hidden layer contains a 2-layer structure and is composed of 64 neurons and 32 neurons which are connected completely; the output layer is a 16-dimensional vector; the process of training the bayesian neuron-based neural network model includes:
s1: acquiring an original signal data set, preprocessing the original signal data set, and obtaining a training set and a label set of the training set;
s2: initializing network parameters and setting ideal values of a loss function; wherein the network parameters include a first network parameter, a second network parameter, and a third network parameter; the first network parameters are:
ω＝μ+log(1+e ρ )*ε,ε～N(0,1)
wherein ω represents a first network parameter, μ represents a second network parameter, ρ represents a third network parameter, ε -N (0, 1) represents a random variable ε obeying a standard normal distribution;
s3: initializing a variation parameter according to the second network parameter and the third network parameter, and defining a minimized variation distribution according to the variation parameter and the first network parameter; the specific process comprises the following steps: the variation parameter is expressed as θ= (μ, ρ) according to the second network parameter and the third network parameter; defining a minimized variation distribution q (omega|theta) to approximate the true posterior probability P (omega|d) based on the variation parameters and the first network parameters; wherein, D= { X, Y }, X represents input data, namely receiving end data needing to be balanced in the CAP-VLC system; y represents tag data, namely a tag value corresponding to each group of input data in the CAP-VLC system;
s4: inputting the training set and the label set of the training set into a neural network model based on Bayesian neurons for training;
s5: calculating a loss function of a neural network model based on Bayesian neurons according to the first network parameters and the minimized variation distribution; the loss function of the neural network model based on Bayesian neurons is as follows:
L(ω|θ)＝H+αKL[q(ω|θ)||P(ω)]
wherein L (omega, theta) represents the loss of the neural network model, H represents the cross entropy loss of the model, alpha represents the dynamic attenuation coefficient, KL represents the divergence loss, q (omega|theta) represents the variation distribution, P (omega) represents the prior probability of the first network parameter, and P i Representing the probability of the output signal level of the ith neuron, t i Representing a target classification level;
s6: judging whether the loss function reaches an ideal value of the loss function, if so, storing network parameters and outputting a neural network model corresponding to the network parameters, otherwise, updating the network parameters by adopting Bayesian back propagation, and returning to the step S4; the process of updating network parameters includes:
calculating a gradient of the second network parameter and a gradient of the third network parameter;
updating the second network parameter according to the gradient of the second network parameter, and updating the third network parameter according to the gradient of the third network parameter; the formula for updating the second network parameter is:
μ * ＝μ-ηΔ μ
wherein delta is μ Represents the gradient, μ of the second network parameter * Representing the updated second network parameter, η representing the learning rate;
the formula for updating the third network parameter is:
ρ * ＝ρ-ηΔ ρ
wherein delta is ρ Representing the gradient of the third network parameter ρ * Representing the updated third network parameter;
the first network parameter is updated according to the new second network parameter and the third network parameter.
2. The method for compensating a visible light CAP system based on bayesian neurons according to claim 1, wherein the hidden layer uses a ReLU function as an activation function, and the output layer uses a softmax function as an activation function;
the formula of the ReLU function is:
the softmax function is formulated as:
wherein y represents the output of the hidden layer, x represents the input of the hidden layer, and P i Representing the probability of the output signal level of the ith neuron, y i Representing the output of the ith neuron of the hidden layer, y k The output of the kth neuron of the hidden layer is represented, and n represents the number of decision classifications.
3. A method of compensating a bayesian neuron-based visible light CAP system according to claim 1, wherein the preprocessing of the raw signal data set comprises:
acquiring signal data of a transmitting end and signal data of a receiving end of a CAP visible light system; receiving end signal data according to 7:3, dividing the ratio into a training set sample and a test set sample; transmitting end signal data according to 7:3 into a label signal sample of a training set and a label signal sample of a test set;
dividing the training set sample and the test set sample by adopting a sliding window to obtain a training set and a test set;
dividing a label signal sample of the training set and a label signal sample of the test set by adopting a sliding window to obtain the label signal sample set of the training set and the label signal sample set of the test set;
and taking the middle number of the label signal sample set of the training set to form a label set of the training set.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210539304.XA CN114978313B (en) | 2022-05-18 | 2022-05-18 | Compensation method of visible light CAP system based on Bayesian neurons |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210539304.XA CN114978313B (en) | 2022-05-18 | 2022-05-18 | Compensation method of visible light CAP system based on Bayesian neurons |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114978313A CN114978313A (en) | 2022-08-30 |
| CN114978313B true CN114978313B (en) | 2023-10-24 |
Family
ID=82982648
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210539304.XA Active CN114978313B (en) | 2022-05-18 | 2022-05-18 | Compensation method of visible light CAP system based on Bayesian neurons |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114978313B (en) |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106454350A (en) * | 2016-06-28 | 2017-02-22 | 中国人民解放军陆军军官学院 | Non-reference evaluation method for infrared image |
| CN112865866A (en) * | 2021-01-20 | 2021-05-28 | 重庆邮电大学 | Visible light PAM system nonlinear compensation method based on GSN |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP3364342A1 (en) * | 2017-02-17 | 2018-08-22 | Cogisen SRL | Method for image processing and video compression |
| US10905337B2 (en) * | 2019-02-26 | 2021-02-02 | Bao Tran | Hearing and monitoring system |
-
2022
- 2022-05-18 CN CN202210539304.XA patent/CN114978313B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106454350A (en) * | 2016-06-28 | 2017-02-22 | 中国人民解放军陆军军官学院 | Non-reference evaluation method for infrared image |
| CN112865866A (en) * | 2021-01-20 | 2021-05-28 | 重庆邮电大学 | Visible light PAM system nonlinear compensation method based on GSN |
Non-Patent Citations (3)
| Title |
|---|
| LED Nonlinearity Estimation and Compensation in VLC Systems Using Probabilistic Bayesian Learning;Chen Chen;《MDPI》;全文 * |
| 基于稀疏贝叶斯学习的可见光通信系统中LED非线性补偿技术;董作霖;《电视技术》;全文 * |
| 基于稀疏贝叶斯学习的可见光通信系统中LED非线性补偿技术;董作霖;吴松;李明;;电视技术(第02期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114978313A (en) | 2022-08-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| WO2019246605A1 (en) | Optical fiber nonlinearity compensation using neural networks | |
| Chuang et al. | Employing deep neural network for high speed 4-PAM optical interconnect | |
| CN112865866B (en) | Visible light PAM system nonlinear compensation method based on GSN | |
| Aref et al. | End-to-end learning of joint geometric and probabilistic constellation shaping | |
| CN109905170B (en) | K-DNN-based nonlinear distortion compensation algorithm and visible light communication device | |
| CN109039472A (en) | A kind of data center's optic communication dispersive estimates and management method based on deep learning | |
| CN109347555A (en) | A kind of visible light communication equalization methods based on radial basis function neural network | |
| CN112598072B (en) | Equalization method of improved Volterra filter based on weight coefficient migration of SVM training | |
| CN110392006B (en) | Self-adaptive channel equalizer and method based on integrated learning and neural network | |
| KR101979394B1 (en) | Adaptive transmission scheme determination apparatus based on MIMO-OFDM System using machine learning model and adaptive transmission method the same | |
| CN111917474B (en) | Implicit triple neural network and optical fiber nonlinear damage balancing method | |
| CN109818889B (en) | Equalization algorithm for SVM classifier optimization in high-order PAM optical transmission system | |
| Niu et al. | End-to-end deep learning for long-haul fiber transmission using differentiable surrogate channel | |
| CN111565160A (en) | Combined channel classification, estimation and detection method for ocean communication system | |
| CN104410593B (en) | Numerical chracter nonlinearity erron amendment equalization methods based on decision-feedback model | |
| CN114513394B (en) | Signal modulation format identification method, system and device based on attention mechanism diagram neural network and storage medium | |
| CN114978313B (en) | Compensation method of visible light CAP system based on Bayesian neurons | |
| CN111988249A (en) | Receiving end equalization method based on adaptive neural network and receiving end | |
| CN114124223B (en) | Convolutional neural network optical fiber equalizer generation method and system | |
| Yıldırım et al. | Deep receiver design for multi-carrier waveforms using cnns | |
| CN113938198B (en) | Optical fiber transmission system, LDA-based method and module for simplifying nonlinear equalizer | |
| Wang et al. | Low-complexity nonlinear equalizer based on artificial neural network for 112 Gbit/s PAM-4 transmission using DML | |
| CN114500189B (en) | Direct pre-equalization method, system, device and medium for visible light communication | |
| Lu et al. | Patterns quantization with noise using Gaussian features and Bayesian learning in VLC systems | |
| CN114204993B (en) | Nonlinear equalization method and system based on polynomial mapping feature construction |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |