CN109660297B - A Machine Learning-Based Physical Layer Visible Light Communication Method - Google Patents

Abstract

The invention discloses a physical layer visible light communication method based on machine learning (ML), establishes an end-to-end visible light communication system, and has a transmission distance of 0 cm to 140 cm. On this basis, the received data is sampled and converted into a vector and collected into the dataset. The performance of ML-based demodulation methods, including Convolutional Neural Networks (CNN), Deep Belief Networks (DBN), and Adaptive Boosting (AdaBoost), is investigated. In particular, the present invention proposes a method for representing modulated signals with a two-dimensional image. In addition, the effects of modulation method, data vector dimension and training set size on performance are also investigated. Finally, the effective rates of different modulation schemes are calculated, and the results show that different modulation schemes should be selected according to the signal-to-noise ratio.

Description

A Machine Learning-Based Physical Layer Visible Light Communication Method

technical field

The present invention relates to the field of visible light communication, in particular to a physical layer visible light communication method based on machine learning (ML).

Background technique

With the rapid development of mobile digital devices and wireless data services, the demand for high-speed wireless transmission has grown exponentially. The traditional radio frequency (Radio Frequency, RF) system is facing the challenge of overcrowded spectrum, and the enhancement of network capacity has encountered a bottleneck (Reference [1]). Visible Light Communication (VLC) has emerged as a potential solution for short-range wireless communication due to its advantages of huge unmanaged spectrum, high data rate, strong security, and strong anti-electromagnetic interference capability. has received extensive attention (Ref. [2]). With the large-scale deployment of light-emitting diodes (LEDs), VLCs employ intensity modulation and direct detection techniques for a dual purpose: illumination and data transmission (refs [3]–[7]).

The modulation and demodulation of wireless signals play a fundamental role in the VLC system. In most existing studies, the VLC channel is assumed to be a direct channel with additive white Gaussian noise (AWGN) (References [8]-[10]). However, practical VLC channels face the challenges of non-linearity, multipath dispersion, impulse noise, spurious or continuous interference, and low sensitivity of photodetectors of commercial LEDs. These challenges always coexist and interact in practice, making the demodulation problem challenging.

So far, many works have applied ML to various communication domains, such as channel estimation (refs [13], [14]), medium access control (refs [15], [16]), automatic modulation classification ( [17], [18]), interference management (ref. [19]), pilot allocation (ref. [20]), antenna selection (ref. [21]), channel decoding (ref. [22] )). A recent literature [23] proposes the concept of interpreting an end-to-end communication system as an autoencoder, and its feasibility is verified using software-defined radio in literature [24]. Since then, potential applications of deep learning (DL) at the physical layer have also received increasing attention, mainly due to the new features of future communications, such as complex scenarios with unknown channel models (ref. [25]), high-speed and Precise handling requirements.

In [26], the authors used an artificial neural network as a demodulator to demodulate 16-QAM signals, and experiments proved that the demodulator outperformed the traditional method using linear filters. In [27], the authors propose an artificial neural network (ANN) based demodulator for frequency shift keying (FSK) signals. Compared with the traditional receiver, the demodulator has better anti-interference ability. By utilizing one-dimensional convolutional neural network (1-D CNN), literature [28] proposed a binary phase shift keying demodulator to deal with carrier frequency offset and sampling frequency error. In [29], a deep convolutional network (DCNN) demodulator was proposed to demodulate the sequence of symbols separately from the mixed signal. In [30], the authors demonstrate the excellent performance of DCNN in demodulating Rayleigh decayed signals. Feature extraction methods based on Deep Belief Network (DBN) are also used to solve this challenging problem. In [31], a DBN-based method is proposed for signal demodulation in different types of communication channels. The authors demonstrate that the DBN-based demodulator can be used for AWGN channels with a certain channel impulse response and for Rayleigh non-frequency selective flat fading channels. In [32], a short-distance multipath channel signal demodulation method based on DL (Deep Learning) is proposed. In [33], a neural network-based software radio receiver is proposed to process demodulated signals on unknown channels. However, these works are based on simulated datasets rather than real datasets.

DL has certain applications in the VLC system. Reference [34] adopts an unsupervised DL-based autoencoder to design the transceiver for multicolor VLC system, which outperforms the state-of-the-art in terms of average symbol error rate. In [35], an autoencoder based on a DL method is used to combat the complex optical properties of dimming control and channel imperfections. However, research on ML-based demodulators in VLC systems is insufficient, and there are no open measured datasets.

references:

[1] V. Chandrasekhar, J. Andrews, and A. Gatherer, "Femtocell networks: asurvey," IEEE Commun. Mag., vol. 46, no. 9, pp. 59–67, Sep. 2008.

[2] “IEEE standard for local and metropolitan area networks–part 15.7:Short-range wireless optical communication using visible light,” IEEE Std802.15.7-2011, pp.1–309, Sep.2011.

[3] T.Komine and M.Nakagawa, "Fundamental analysis for visible-lightcommunication system using LED lights," IEEE Trans.Consum.Electron.,vol.50,no.1,pp.100–107,Feb.2004.

[4] H. Elgala, R. Mesleh, and H. Haas, "Indoor optical wireless communication: Potential and state-of-the-art," IEEE Commun. Mag., vol. 49, no. 9, pp. 56– 62, Dec. 2011.

[5] S. Arnon, J. Barry, G. Karagiannidis, R. Schober, and M. Uysal, Advanced Optical Wireless Communication Systems, 1st ed, Cambridge, U.K.: Cambridge Univ, 2012.

[6] A. Jovicic, J. Li, and T. Richardson, “Visible light communication: opportunities, challenges and the path to market,” IEEE Commun. Mag., vol. 51, no. 12, pp. 26–32 , Dec.2013.

[7] P.H.Pathak, X.Feng, P.Hu, and P.Mohapatra, "Visible lightcommunication, networking, and sensing: a survey, potential and challenges," IEEECommun.Surveys Tuts.,vol.17,no.4, pp. 2047–2077, Sep. 2015.

[8] T.Fath and H.Haas, "Performance comparison of MIMO techniques for optical wireless communications in indoor environments," IEEE Trans. Commun., vol. 61, no. 2, pp. 733–742, Feb. 2013.

[9] T.Q.Wang, Y.A.Sekercioglu, and J.Armstrong, "Analysis of an opticalwireless receiver using a hemispherical lens with application in MIMO visiblelight communications," J.Lightw.Technol., vol.31, no.11, pp.1744 –1754, Jun. 2013. [10] K. Ying, H. Qian, R. J. Baxley, and S. Yao, “Joint optimization of precoder and equalizer in MIMO VLC systems,” IEEE J. Sel. Areas in Comm., vol. 33, no.9, pp.1949–1958, Sep.2015

[11] T. Mitchell, B. Buchanan, G. Dejong, T. Dietterich, P. Rosenbloom, and A. Waibel, Machine Learning, China Machine Press, 2003.

[12] K. Hornik, M. Stinchcombe, and H. White, "Multilayer feedforward networks are universal approximators," Neural Networks, vol. 2, no. 5, pp. 359–366, Jul. 1989.

[13] M.N.Seyman and N.Tapnar, "Channel estimation based on neuralnetwork in space time block coded MIMO-OFDM system," Digital Signal Process., vol.23, no.1, pp.275–280, Jan.2013.

[14] H.Ye, G.Y.Li, and B.H.Juang, “Power of deep learning for channelestimation and signal detection in OFDM systems,” IEEE Wireless Commun.Lett., vol.7, no.1, pp.114–117, Apr.2018.

[15] S. Hu, Y. Yao, and Z. Yang, “Mac protocol identification using supportvector machines for cognitive radio networks,” IEEE Wireless Commun., vol. 21, no. 1, pp. 52–60, Feb. 2014.

[16] S. Wang, H. Liu, P. H. Gomes, and B. Krishnamachari, "Deep reinforcement learning for dynamic multichannel access in wireless networks," IEEE Trans. Cogn. Commun. Netw., vol. 4, no. 2, pp. 257-265, Jun. 2018.

[17] M.W.Aslam, Z.Zhu, and A.K.Nandi, “Automatic modulationclassification using combination of genetic programming and KNN,” IEEETrans.Wireless Commun.,vol.11,no.8,pp.2742–2750,Jun.2012.

[18] A.S.Liu and Z.Qi, “Automatic modulation classification based on the combination of clustering and neural network,” The Journal of ChinaUniversities of Posts and Telecommunications, vol.18, no.4, pp.13,38–19,38 , Apr. 2011.

[19]H.Sun,X.Chen,M.Hong,Q.Shi,X.Fu,and N.D.Sidiropoulos,“Learning tooptimize:Training deep neural networks for interference management,”IEEETrans.Signal Process.,vol.66, no.20,pp.5438-5453,Oct.2018.

[20] K. Kim, J. Lee, and J. Choi, "Deep learning based pilot allocation scheme (DL-PAS) for 5G massive mimo system," IEEE Commun. Lett., vol. 22, no. 4, pp. 828–831, Feb. 2018.

[21] J. Joung, "Machine learning-based antenna selection in wireless communications," IEEE Commun. Lett., vol. 20, no. 11, pp. 2241–2244, Jul. 2016.

[22] F.Liang, C.Shen, and F.Wu, “An iterative BP-CNN architecture for channel decoding,” IEEE J.Sel.Topics Signal Process., vol.12, no.1, pp.144–159 , Jan. 2018.

[23] T.O.Shea and J.Hoydis, "An introduction to deep learning for the physical layer," IEEE Trans.Cogn.Commun.Netw., vol.3, no.4, pp.563–575, Oct.2017.

[24] S. Dorner, S. Cammerer, J. Hoydis, and S.T. Brink, "Deep learning based communication over the air," IEEE J.Select.Topics Signal Process., vol.12, no.1, pp.132– 143, Feb. 2018.

[25] V.Raj and S.Kalyani, "Backpropagating through the air: Deeplearning at physical layer without channel models," IEEE Commun.Lett., vol.22, no.11, pp.2278-2281, Nov.2018.

[26] A.N.Milad, M.A.M, and Rahmadwati, “Neural network demodulator for quadrature amplitude modulation (QAM),” International Journal of Advanced Studies in Computer Science and Engineering, vol.5, no.7, pp.10–14, 2016.

[27] M.Li, H.Zhong, and M.Li, “Neural network demodulator for frequencyshift keying,” in 2008 International Conference on Computer Science and Software Engineering, vol.4, pp.843–846, Dec.2008.

[28] M. Zhang, Z. Liu, L. Li, and H. Wang, “Enhanced efficiency BPSK demodulator based on one-dimensional convolutional neural network,” IEEE Access, vol. 6, pp. 26939-26948, 2018.

[29] X.Lin, R.Liu, W.Hu, Y.Li, X.Zhou, and X.He, “A deep convolutionalnetwork demodulator for mixed signals with different modulation types,” inProc.IEEE 15th Intl Conf on Dependable , Autonomic and Secure Computing, pp.893–896, Nov.2017.

[30] A.S.Mohammad, N.Reddy, F.James, and C.Beard, “Demodulation of fadedwireless signals using deep convolutional neural networks,” in 2018 IEEE 8thAnnual Computing and Communication Workshop and Conference (CCWC), pp.969–975 , Jan. 2018.

[31] M.Fan and L.Wu, “Demodulator based on deep belief networks incommunication system,” in International Conference on Communication, Control, Computing and Electronics Engineering, pp.1–5, Jan.2017.

[32] L.Fang and L.Wu, “Deep learning detection method for signal demodulation in short range multipath channel,” in IEEE International Conference on Opto-Electronic Information Processing, pp.16–20, Jul.2017.

A.Akan,and H.

“Neural network based receiver designfor software defined radio over unknown channels,”in Proc.8^thInt.Conf.Electr.Electron.Eng.,pp.297–300,Nov.2013.[33] M.

A. Akan, and H.

“Neural network based receiver design for software defined radio over unknown channels,” in Proc.8 ^th Int.Conf.Electr.Electron.Eng., pp.297–300, Nov.2013.

[34] H. Lee, I. Lee, and S. H. Lee, “Deep learning based transceiver design for multi-colored VLC systems,” Optics Express, vol.26, no.5, pp.6222-6238, 2018.

[35] H.Lee, I.Lee, T.Q.S.Quek, and H.L.Sang, “Binary signaling design forvisible light communication: a deep learning framework,” Optics Express, vol.26, no.14, pp.18131-18142, 2018 .

[36] J.Sola and J.Sevilla, "Importance of input data normalization for the application of neural networks to complex industrial problems," Nucl.Sci., vol.44, no.3, pp.1464–1468, Jun.1997 .

[37] RC Gonzalez and REWoods, Digital Image Processing ( ^3rd Edition), Prentice-Hall, Inc., 2007.

[38] G.E.Hinton and R.R.Salakhutdinov, "Reducing the dimensionality of data with neural networks," Science, vol.313, no.5786, pp.504–507, 2006

[39] D.E. Rumelhart and J.L. McClelland, Information Processing in Dynamical Systems: Foundations of Harmony Theory, MIT Press, 1987.

[40] S. Haykin, Neural Networks: A Comprehensive Foundation (3rd Edi-tion), Macmillan, 1998

[41] G.E.Hinton, “Training products of experts by minimizing contrastive divergence,” Neural Computation, vol.14, no.8, pp.1771–1800, Aug.2002.

[42] L.J.Buturovic and L.T.Citkusev, "Back propagation and forwardpropagation," in Int.Joint Conf.Neural Networks, vol.4, pp.486–491, Jun.1992.

[43] Y.Freund and R.E.Schapire, "A decision-theoretic generalization of on-line learning and an application to boosting," Comput.Syst.Sci.,vol.55,no.1,pp.119–139,1997.

[44] I.Mukherjee, C.Rudin, and R.E.Schapire, "The rate of convergence of Adaboost," J.Mach.Learn.Res., vol.14, no.3, pp.2315–2347, 2011.

T.Onoda,and K.R.Müller,“Soft margins for Adaboost,”Machine Learning,vol.42,no.3,pp.287–320,2001.[45] G.

T. Onoda, and KRMüller, "Soft margins for Adaboost," Machine Learning, vol. 42, no. 3, pp. 287–320, 2001.

[46] M. Riesenhuber and T. Poggio, "Hierarchical models of objectrecognition in cortex," Nature neuroscience, vol.2, no.11, pp.1019-1025, 1999.

SUMMARY OF THE INVENTION

Aiming at the deficiencies of the prior art, the present invention provides a physical layer visible light communication method based on machine learning, comprising the following steps:

Step 1, establish a VLC system model;

Step 2, adopt any one of the demodulators based on convolutional neural network (CNN), deep belief network (DBN), adaptive boosting (adaptive boosting, AdaBoost) to establish the VLC. system model for demodulation.

Step 1 includes:

Step 1-1, build an end-to-end VLC system, which includes a single light emitting diode (LEDs) emitter and a single photodetector, the emission signal x(t) is as follows:

Where, t is the time, s(t) is the baseband signal, j is the imaginary unit, s(t) is the baseband signal, fc is the carrier frequency, p( _t ) is the signal pulse, and T is the signal period;

Let g denote the channel between the LED and the photodetector, including the direct path and the multi-reflection path, at the receiver, the received signal y(t) is as follows:

y(t)=gx(t)+n(t) (2)

表示第i个周期内的信号采样向量，其中

表示第n个采样点：

n取值为1～N，N是一个周期内的采样数；where n(t) is the received noise, and the received analog signal y(t) is sampled to a digital signal through a digital-to-analog converter; set

represents the vector of signal samples in the ith cycle, where

Represents the nth sampling point:

n ranges from 1 to N, where N is the number of samples in a cycle;

标准化到[0,1]区间，如下：Step 1-2, set the training data set to include K received sampling data periods, and i takes a value from 1 to K. The received sample sequence

Normalize to the [0,1] interval, as follows:

表示归一化后的第i个采样点的值，

表示采样序列的最小值，

表示采样序列的最大值；in,

represents the value of the i-th sampling point after normalization,

represents the minimum value of the sampling sequence,

Represents the maximum value of the sampling sequence;

设定其对应标签z_i，其中1≤i≤K，令

表示标记的训练数据集。令

表示所有标签的集合，由使用的调制方式决定，且

Steps 1-3, for the normalized ith vector

Set its corresponding label z _i , where 1≤i≤K, let

Represents the labeled training dataset. make

represents the set of all tags, determined by the modulation used, and

In step 2, the CNN-based demodulator includes a visualization module and a CNN, and when the CNN-based demodulator is used to demodulate the established VLC system model, the following steps are included:

通过可视化模块转化成二维的图像格式，以输入到专为二维数据设计的CNN中进行图像分类，可视化模块的输出图像表示为X，X是28×28大小的矩阵：

为实数集；Step a1, the received data vector

It is converted into a two-dimensional image format by the visualization module and input into a CNN designed for two-dimensional data for image classification. The output image of the visualization module is represented as X, where X is a 28×28 matrix:

is the set of real numbers;

表示第一个卷积层的第i个卷积核，

用

表示由

得到的特征图，由下式获得：In step a2, the CNN processing X includes two convolutional layers, two pooling layers and a fully connected layer, using

represents the ith convolution kernel of the first convolutional layer,

use

represented by

The resulting feature map is obtained by:

的偏置，*表示卷积操作，

为激活函数，

表示第一次卷积之后得到的矩阵的元素，

表示

的元素，p＝1,2,...,24，q＝1,2,...,24；where b _i represents

The bias of , * denotes the convolution operation,

is the activation function,

represents the elements of the matrix obtained after the first convolution,

express

elements of , p=1,2,...,24, q=1,2,...,24;

表示对第i个特征图的池化结果，通过下式得到：In step a3, the convolutional layer is followed by a pooling layer, which performs a downsampling operation on the feature map output by the previous layer, and uses the maximum pooling method for pooling. The size of the receptive field is 2×2.

Represents the pooling result of the i-th feature map, which is obtained by the following formula:

表示第i个输入的特征图；where pooling( ) represents the downsampling function,

represents the feature map of the ith input;

表示第二个卷积层中的第j个卷积核，

设定

是第二层卷积层的输出，

由下式得到：Step a4, use

represents the jth convolution kernel in the second convolutional layer,

set up

is the output of the second convolutional layer,

It is obtained by the following formula:

表示第二次卷积之后得到的矩阵的元素，

是

中的元素，p＝1,2,...,10，q＝1,2,...,10。in

represents the elements of the matrix obtained after the second convolution,

Yes

Elements in , p=1,2,...,10, q=1,2,...,10.

经过全连接层转化为一个一维的标签向量y³，该向量的神经元数目由调制方式决定，CNN输出的标签

表示为：Step a5, after the second pooling layer with a receptive field of 2 × 2, its output

After the fully connected layer, it is converted into a one-dimensional label vector y ³ . The number of neurons in this vector is determined by the modulation method. The label output by CNN

Expressed as:

并能够进一步映射到解调结果

Where [y ³ ] _i represents the value of the i-th dimension element in y ³ , and the dimension of y ³ is determined by the modulation method used. The subscript corresponding to the largest element in y ³ is the label

and can be further mapped to the demodulation result

中的每一个元素首先被转化为二维平面上的一个点，将这些点用折线连接起来，得到接收信号的波形图，调制信号的幅度信息和相位信息都被保存在这张灰度图片中，使用双三次插值法缩小该灰度图的尺寸，采用全局阈值算法将缩小后的图像转化为二值图像；用X表示可视化模块最终的输出的二值图像，X是一个28×28大小的矩阵。Step a1 includes: the visualization module performs the following processing:

Each element in is first converted into a point on a two-dimensional plane, and these points are connected with a broken line to obtain the waveform diagram of the received signal. The amplitude information and phase information of the modulating signal are stored in this grayscale image. , use the bicubic interpolation method to reduce the size of the grayscale image, and use the global threshold algorithm to convert the reduced image into a binary image; use X to represent the final output binary image of the visualization module, X is a 28×28 size matrix.

In step 2, when using a DBN-based demodulator to demodulate the established VLC system model, the following steps are included:

Step b1, establish a deep belief network (DBN) with three restricted Boltzmann machines (Restricted boltzmann machines, RBM). RBM is an implementation of an undirected graphical model. ₁ ,v ₂ ,...,v _m ] ^T and the hidden layer h=[h ₁ ,h ₂ ,...,h _n ] ^T , where v _i and h _j represent the i-th unit of the display layer respectively and the value of the jth unit of the hidden layer, i is 1～m, j is 1～n; set W=[w ₁ ,w ₂ ,...,w _n ] ^T represents v and The connection weight matrix between h, where w _j =[w _j1 ,w _j2 ,...,w _jm ] ^T , where w _ji represents the connection weight between vi and _{h j ;} a=[a ₁ ,a ₂ ,..., _am ] ^T and b=[b ₁ ,b ₂ ,...,b _n ] ^T represent the bias of v and the bias of h, respectively, where a _i represents the bias of v _i , b _j represents the offset of h _j ;

中的归一化信号

第一个RBM的能量函数E(v,h)如下：Step b2, introduce the energy function to represent the state of the RBM, and use the training data set

normalized signal in

The energy function E(v,h) of the first RBM is as follows:

E(v,h)=-a ^T vb ^T hh ^T Wv, (8)

in,

The marginal distribution p(v) of the display layer v is expressed as follows:

是归一化因子；in,

is the normalization factor;

In step b3, the optimal parameters W, a, b are obtained by maximizing the following unconstrained log-maximum-likelihood function:

In step b4, the gradient descent method is used to solve the optimization problem in step b3, and the variables W, a, and b are updated as follows:

Among them, ε represents the learning rate, which is taken as 0.1, and ΔW, Δa and Δb represent the partial derivative of the objective function to W, the partial derivative of a and the partial derivative of b, respectively;

In step b5, the partial derivatives of the variables W, a, and b are respectively approximated as:

代表重构的显层数据，其中

为重构的第i个显层神经元的值，i取值为1～m。p(h_j＝1|v)表示对于给定的显层v，隐层的第j个神经元被激活的概率。

表示给定重构之后的显层

隐层的第j个神经元被激活的概率。in,

represents the reconstructed explicit layer data, where

is the value of the reconstructed i-th explicit layer neuron, where i is 1 to m. p(h _j =1|v) represents the probability that the jth neuron of the hidden layer is activated for a given visible layer v.

Represents the explicit layer after a given reconstruction

The probability that the jth neuron of the hidden layer is activated.

可由如下方法得到：

It can be obtained by the following methods:

Given the explicit layer v, the distribution of the units in the hidden layer h is as follows:

其中

是隐层的第j个神经元的值，j＝1,2,...,n，则：According to the distribution of formula (13), the hidden layer data is generated according to the following formula

in

is the value of the jth neuron of the hidden layer, j=1,2,...,n, then:

Among them, p(h|v) represents the probability of obtaining the hidden layer state h given the visible layer v.

显层的第i个神经元被激活的概率

由下式给出：for a given hidden layer

The probability that the ith neuron of the explicit layer is activated

is given by:

由(15)分布产生，如下：Step b6, reconstructed display layer data

Produced by the (15) distribution, as follows:

In step b7, after obtaining the optimal parameters W, a, b of the first RBM by using the gradient descent method, the hidden layer h of the first RBM is regarded as the visible layer of the second RBM, and h ⁽¹⁾ is the first RBM. The hidden layer of two RBMs; after training the weight matrix and bias of the second RBM, consider h ⁽¹⁾ as the explicit layer of the third RBM, and let h ⁽²⁾ be the hidden layer of the third RBM; After training the third RBM, all parameters of the RBM are fine-tuned by a supervised back-propagation algorithm; in the testing phase, DBN is applied to the signal demodulation, and the signal demodulation result is output

In step 2, when adopting the demodulator based on adaptive enhancement AdaBoost to demodulate the established VLC system model, the following steps are included:

中所有样本的权重用d_q＝[d_q,1,d_q,2,...,d_q,K]^T来表示，其中d_q,i表示第i个样本的权重，q＝1,2,...,Q；当q＝1，d_q,i＝1/K，i＝1,2,...,K；步骤c2，根据d_q对

进行重采样，得到的

为第q个KNN分类器的训练集；设定

其中(x_q,i,z_q,i)是重采样之后的第i个样本，x_q,i为数据向量，z_q,i为其对应的标签，有

为每个KNN分类器的测试集；用

表示训练数据集中离测试样本

最近的样本，即：Step c1, set the strong classifier to be composed of Q k-nearest neighbor (KNN) classifiers, let k=1. For the qth KNN classifier, the training dataset

_{The weights of} ^all the _{samples in} the 2,...,Q; when q=1, d _q,i =1/K, i=1,2,...,K; step c2, according to d _q pair

After resampling, we get

is the training set of the qth KNN classifier; set

where (x _q,i ,z _q,i ) is the ith sample after resampling, x _q,i is the data vector, z _q,i is the corresponding label, there are

is the test set for each KNN classifier; use

Indicates that the training data set is far from the test sample

The most recent sample, namely:

是x_q,i与

之间的欧氏距离，设定

的标签是

KNN分类器就将

归为类别

in

is x _q,i and

Euclidean distance between, set

The label is

KNN classifier will

Classify

表示第q个KNN分类器，即第q个KNN分类器对样本

的分类结果为

第q个KNN分类器的误差e_q定义为误分类样本的权重之和：Step c3, use

Represents the qth KNN classifier, that is, the qth KNN classifier pair sample

The classification result is

The error e _q of the qth KNN classifier is defined as the sum of the weights of the misclassified samples:

where I(a,b) is the indicator function, defined as follows:

中样本的权重，其中d_q+1,i代表第i个样本的权重，i＝1,2,...,K。它通过下式得到：Let d _q+1 =[d _q+1,1 ,d _q+1,2 ,...,d _q+1,K ] ^T denotes the training data set corresponding to the q+1th KNN classifier

The weight of the middle sample, where d _q+1,i represents the weight of the ith sample, i=1,2,...,K. It is obtained by:

计算得到；在约束e_q＜0.5下，β_q＜1；如果

被分类正确，有

如果

被分类错误，

where β _q is given by the function

Calculated; β _q < 1 under constraint e _q <0.5; if

is classified correctly, there are

if

misclassified,

To re-evaluate the weights of the samples, redefine d _q+1,i by the following normalization formula:

Step c4, after generating Q KNN classifiers, the strong classifier is defined by the following formula:

表示；

是所有标签的集合，z是标签。

是G_q的系数；对于

I(G_q(y),z)作为将y标记为z的投票值；如果I(G_q(y),z)＝1，G_q将样本y分类为标签z，否则y不属于标签z；对于所有KNN分类器，具有最大加权投票值

的类别是这个Adaboost分类器的输出结果

进一步得到解调结果

Among them, H(y) represents the classification result of the test sample y by the strong classifier.

express;

is the set of all labels, and z is the label.

is the coefficient of G _q ; for

I(G _q (y), z) as the vote to label y as z; if I (G _q (y), z) = 1, G _q classifies sample y as label z, otherwise y does not belong to label z ; for all KNN classifiers, with the maximum weighted vote value

The category of is the output of this Adaboost classifier

Further get demodulation result

The present invention proposes a flexible data-driven end-to-end VLC system prototype. Using the proposed VLC system prototype, eight modulated signals, namely OOK, QPSK, 4-PPM, 16-QAM, 32-QAM, 64-QAM, 128-QAM, and 256-QAM, were collected in a real physical environment. Furthermore, an open online modulation dataset was established, in which eight modulated signals were measured at distances ranging from 0 cm to 140 cm.

The present invention proposes a demodulator based on three data drives. Specifically, a CNN-based demodulator is proposed with two convolutional layers and two pooling layers. The classifier first converts the modulated signal into an image, and then demodulates the signal through image classification. Furthermore, a DBN-based demodulator consisting of three restricted Boltzmann machines (RBMs) is designed to extract modulation features. Finally, an AdaBoost signal demodulator is proposed, which uses the weak classifier KNN to construct a strong classifier.

Based on the modulation dataset, the present invention investigates the performance of three proposed data-driven demodulators. Specifically, as the transmission distance increases, the performance of the three demodulators decreases. Given the transmission distance, the higher the order of the modulation method used, the lower the accuracy of demodulation. The experimental results show that the demodulator based on AdaBoost performs the best among all modulation schemes. In addition to this, higher order modulation should be used to achieve the maximum effective rate at short distances or with high signal-to-noise ratios.

Description of drawings

The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the above or other aspects of the present invention will become clearer.

Figure 1 is a schematic diagram of the VLC system.

Figure 2 is a schematic diagram of the visualization module.

Figure 3 is a schematic diagram of CNN.

Figure 4 is a schematic diagram of max pooling.

FIG. 5 is a schematic diagram of a DBN structure.

Figure 6 is an RBM with m visible layer neurons and n hidden layer neurons.

Figure 7 is the generation process of the strong classifier.

Figure 8a shows the demodulation accuracy of the OOK modulated signal with respect to the distance d.

Figure 8b shows the demodulation accuracy of the 32-QAM modulated signal with respect to distance d.

Figure 8c shows the demodulation accuracy of the 256-QAM modulated signal with respect to distance d.

Figure 9 shows the accuracy of the AdaBoost-based demodulator with respect to distance d.

Figure 10 shows the demodulation accuracy of 16-QAM and 32-QAM modulated signals as a function of training period K.

Figure 11a shows the relationship between the distance d of the demodulation accuracy of the eight modulated signals of OOK, QPSK, 4-PPM, 32-QAM, 64-QAM, 128-QAM and 256-QAM when N=40.

Figure 11b shows the effective rate R _eff for the eight modulation schemes.

Detailed ways

The present invention will be further described below with reference to the accompanying drawings and embodiments.

(1) Experimental device:

This embodiment establishes a flexible end-to-end visible light communication prototype platform in a real physical environment, which can demodulate various modulated signals. Based on the established prototype platform, this embodiment establishes measurement modulation data sets from actual communication systems, including training data sets and test data sets, which are available to all researchers. The experimental setup is shown in Figure 1.

(2) System model:

Consider an end-to-end VLC system containing a single LED emitter and a single photodetector, as shown in Figure 1.

Using digital modulation schemes, such as M-ary quadrature amplitude modulation (M-QAM), and M-ary pulse position modulation (M-PPM), the transmit signal x(t) is as follows:

是基带信号，t是时间，j是虚数单位，f_c是载波频率，p(t)是信号脉冲，T是信号周期。in,

is the baseband signal, t is the time, j is the imaginary unit, f _c is the carrier frequency, p(t) is the signal pulse, and T is the signal period.

Let g denote the channel between the LED and the photodetector, including the direct path and the multi-reflection path. At the receiver, the received signal y(t) is given as:

y(t)=gx(t)+n(t) (2)

表示一个周期内的信号采样向量，其中

表示第n个采样点：

n取值为1～N，N是一个周期内的采样数。where n(t) is the received noise. The received analog signal y(t) is sampled into a digital signal through a digital-to-analog converter. Assume

represents a vector of signal samples in one cycle, where

Represents the nth sampling point:

n ranges from 1 to N, where N is the number of samples in one cycle.

标准化到[0,1]区间，如下：It is assumed that the training data set contains K received sampling data periods, and i takes a value from 1 to K. Proper normalization can significantly reduce computation time before processing by ML-based demodulators (ref. [36]). Therefore, first the received sample sequence is

Normalize to the [0,1] interval, as follows:

in,

表示归一化后的第i个采样点的值，

表示采样序列的最小值，

表示采样序列的最大值；

represents the value of the i-th sampling point after normalization,

represents the minimum value of the sampling sequence,

Represents the maximum value of the sampling sequence;

设定其对应标签z_i，其中1≤i≤K。令

表示标记的训练数据集。用

表示标签集，由使用的调制方式决定。例如当调制方式为正交相移键控(Quadrature Phase Shift Keying，QPSK)，

Also, for the normalized i-th vector

Set its corresponding label z _i , where 1≤i≤K. make

Represents the labeled training dataset. use

Represents a tag set, determined by the modulation used. For example, when the modulation method is Quadrature Phase Shift Keying (QPSK),

(3) CNN-based demodulator

通过可视化模块转化成二维的图像格式，以输入到专为二维数据设计的CNN中进行图像分类。可视化模块如图2所示。如图2所示，

中的每一个元素首先被转化为二维平面上的一个点。将这些点用折线连接起来，就可以得到接收信号的波形图。调制信号的幅度信息和相位信息都被保存在这张具有很多像素点的图片中。本发明进一步使用双三次插值法来缩小该灰度图的尺寸，以在保存有用信息的同时减少计算负担。另外，为了进一步强调有用信息，本发明采用全局阈值算法来将缩小的图像转化为二值图像([参考文献37])。用X表示可视化模块的输出图像(矩阵)，X是28×28大小的矩阵：

Compared with general neural networks, convolutional neural networks (CNNs) require less preprocessing. In addition, it has the characteristics of sparse connection, weight sharing, etc., the structure is simpler, and the robustness is stronger. The CNN-based demodulator proposed by the present invention includes a visualization module and a CNN. First convert the received data vector

It is converted into a two-dimensional image format by the visualization module and input into a CNN specially designed for two-dimensional data for image classification. The visualization module is shown in Figure 2. as shown in picture 2,

Each element in is first transformed into a point on a two-dimensional plane. Connect these points with a broken line to get the waveform of the received signal. Both the amplitude and phase information of the modulated signal are stored in this picture with many pixels. The present invention further uses bicubic interpolation to reduce the size of the grayscale image to reduce computational burden while preserving useful information. Additionally, to further emphasize useful information, the present invention employs a global thresholding algorithm to convert the downscaled image into a binary image ([Ref. 37]). Let X denote the output image (matrix) of the visualization module, where X is a matrix of size 28×28:

表示第一个卷积层的第i个卷积核，

用

表示由

得到的特征图，由下式获得：The CNN that processes X includes two convolutional layers, two pooling layers and one fully connected layer, the structure is shown in Figure 3, and the parameters are shown in Table 1. use

represents the ith convolution kernel of the first convolutional layer,

use

represented by

The resulting feature map is obtained by:

的偏置，*表示卷积操作，

为激活函数，

表示第一次卷积之后得到的矩阵的元素，

表示

的元素，p＝1,2,...,24，q＝1,2,...,24。where b _i represents

The bias of , * denotes the convolution operation,

is the activation function,

represents the elements of the matrix obtained after the first convolution,

express

The elements of , p=1,2,...,24, q=1,2,...,24.

The convolutional layer is followed by a pooling layer that performs a downsampling operation on the feature maps output by the previous layer. Pooling can reduce the number of parameters, reduce the computational burden, and improve the robustness of the network. The pooling method used in the present invention is maximum pooling, and the size of the receptive field is 2×2, as shown in FIG. 4 .

表示对第i个特征图的池化结果，通过下式得到：use

Represents the pooling result of the i-th feature map, which is obtained by the following formula:

表示第i个输入的特征图。where pooling( ) represents the downsampling function,

Represents the feature map of the ith input.

表示第二个卷积层中的第j个卷积核，

设定

是第二层卷积层的输出，

由下式得到：use

represents the jth convolution kernel in the second convolutional layer,

set up

is the output of the second convolutional layer,

It is obtained by the following formula:

表示第二次卷积之后得到的矩阵的元素，

是

中的元素，p＝1,2,...,10，q＝1,2,...,10。in

represents the elements of the matrix obtained after the second convolution,

Yes

Elements in , p=1,2,...,10, q=1,2,...,10.

经过全连接层转化为一个一维的标签向量y³，该向量的神经元数目由调制方式决定。CNN输出的标签

表示为：After the second pooling layer with a receptive field of 2×2, its output

After the fully connected layer, it is converted into a one-dimensional label vector y ³ , and the number of neurons in this vector is determined by the modulation method. Labels output by CNN

Expressed as:

进一步地，最大值的标签

被映射到解调结果

[y ³ ] _i represents the value of the i-th dimension element in y ³ , and the dimension of y ³ is determined by the modulation method used. The subscript corresponding to the largest element in y ³ is the label

Further, the label of the maximum value

is mapped to the demodulation result

Table 1 CNN parameter settings

Network layer Size of convolution kernel/receptive field step size output size input layer 28×28 Convolutional layer 1 5×5 1 24×24 Pooling layer 1 2×2 2 12×12 Convolutional layer 2 3×3 1 10×10 Pooling Layer 2 2×2 2 5×5

(4) DBN-based demodulator

Since DBN can effectively utilize multiple nonlinear transformations to extract high-level features from data, it is widely used to solve practical problems (Ref. [38]). The present invention considers a DBN with three restricted Boltzmann machines (RBMs), each RBM consisting of an explicit layer v=[v ₁ ,v ₂ ,..., _vm ] ^T and a hidden layer h=[ h ₁ , h ₂ ,...,h _n ] ^T , as shown in Figure 6, where vi and h _j represent the _i -th unit of the display layer and the j-th unit of the hidden layer, respectively. Let W=[w ₁ ,w ₂ ,...,w _n ] ^T represent the connection weight matrix between v and h, and its w _j =[w _j1 ,w _j2 ,...,w _jm ] ^T ,j =1,2,...,n, w _ji represents the connection weight between vi and _{h j .} In addition, a=[a ₁ ,a ₂ ,..., _am ] ^T and b=[b ₁ ,b ₂ ,...,b _n ] ^T represent the biases of v and h, respectively, where a _i represents The bias of v _i , and b _j represents the bias of h _j .

中的归一化信号

第一个RBM的能量函数如下：Inspired by statistical mechanics, an energy function is introduced to represent the state of the RBM (Ref. [39]). use

normalized signal in

The energy function of the first RBM is as follows:

E(v,h)=-a ^T vb ^T hh ^T Wv, (8)

in,

The edge distribution of the display layer v is as follows:

是归一化因子，v表示显层，h表示隐层。in,

is the normalization factor, v represents the visible layer, and h represents the hidden layer.

Then, the optimal parameters W, a, b can be obtained by minimizing the energy of the RBM (Ref. [40]) as follows:

The present invention adopts the gradient descent method to solve the optimization problem (10). Specifically, the variables W, a, and b are updated as follows:

Among them, ε represents the learning rate, which is taken as 0.1, and ΔW, Δa and Δb represent the gradients of W, a, and b, respectively.

Furthermore, the partial derivatives of the variables W, a, and b can be approximated as

表示给定重构之后的显层

隐层的第j个神经元被激活的概率。

Represents the explicit layer after a given reconstruction

The probability that the jth neuron of the hidden layer is activated.

代表重构的显层数据，(参考文献[41])。其中

为重构的第i个显层神经元的值，i取值为1～m。in,

Represents reconstructed display layer data, (Ref. [41]). in

is the value of the reconstructed i-th explicit layer neuron, where i is 1 to m.

Given the explicit layer v, the probability p(h _j = 1|v) of the jth neuron being activated in the hidden layer h is calculated as follows:

如下：Then, according to the distribution of formula (13), we get

as follows:

Likewise, the probability that the ith neuron of the display layer v is activated is given by:

可以由(15)分布产生，如下：Then, the reconstructed data

can be generated from the (15) distribution, as follows:

Using the gradient descent method, the optimal parameters W, a, b of the first RBM can be obtained. Then, the hidden layer h of the first RBM is regarded as the visible layer of the second RBM, and its hidden layer is h( ¹ ), as shown in FIG. 6 . After training the weight matrix and bias of the second RBM, consider h( ¹ ) as the explicit layer of the third RBM, and let h( ² ) be the hidden layer of the third RBM. After training the third RBM, all parameters of the RBM (weights and biases) are fine-tuned by a supervised back-propagation algorithm (Ref. [42]). The parameters of the DBN model are updated to approximate the optimal classifier. After training, apply the DBN network in the test phase to signal demodulation, and output the signal demodulation result

(5) ADABOOST-based demodulator

The Adaboost algorithm is able to integrate multiple independent weak classifiers into a high-performance strong classifier (Reference [43]). The present invention uses the AdaBoost algorithm to demodulate the signal, uses the k-nearest neighbor (KNN) algorithm as a weak classifier, and sets k=1. The final strong classifier H is the combination of all KNNs, as shown in Figure 7.

中所有样本的权重用d_q＝[d_q,1,d_q,2,...,d_q,K]^T来表示，其中d_q,i表示第i个样本的权重，q＝1,2,...,Q。当q＝1，d_q,i＝1/K，i＝1,2,...,K。It is assumed that the strong classifier is composed of Q KNNs (Ref. [44]). For the qth KNN classifier, the sample set

_{The weights of} ^all the _{samples in} the 2,...,Q. When q=1, d _q,i =1/K, i=1,2,...,K.

进行重采样(参考文献[45])，得到的

为第q个KNN分类器的训练集。假设

有

为每个KNN分类器的测试集。用

表示训练集中离测试样本

最近的样本，即According to d _q pair

Resampling (Ref. [45]) yields

is the training set of the qth KNN classifier. Assumption

Have

Test set for each KNN classifier. use

Indicates that the training set is far from the test sample

The most recent sample, i.e.

是x_q,i与

之间的欧氏距离。假设

的标签是

KNN分类器就将

归为类别

in

is x _q,i and

Euclidean distance between . Assumption

The label is

KNN classifier will

Classify

表示第q个分类器，即第q个分类器对样本

的分类结果为

G_q的误差被定义为误分类样本的权重之和：use

Represents the qth classifier, that is, the qth classifier pairs the sample

The classification result is

The error of G _q is defined as the sum of the weights of the misclassified samples:

where I(a,b) is the indicator function, defined as follows:

中样本的权重，它通过下式得到：Similarly, let d _q+1 =[d _q+1,1 ,d _q+1,2 ,...,d _q+1,K ] ^T denote the corresponding value of the q+1th KNN

The weight of the sample in , which is obtained by:

计算而来。在约束e_q＜0.5下，β_q＜1。如果

被分类正确，有

如果

被分类错误，

where β _q is given by the function

calculated. β _q &lt; 1 under the constraint e _q &lt; 0.5. if

is classified correctly, there are

if

misclassified,

In order to re-evaluate the weights of the samples, d _q+1,i is redefined by the following normalization formula:

After generating Q KNN classifiers, a strong classifier is defined by:

是G_q的系数。对于

I(G_q(y),z)可作为将y标记为z的投票值。如果I(G_q(y),z)＝1，G_q将样本y分类为z类，否则y不属于z类。对于所有分类器，具有最大加权投票值

的类别是这个Adaboost分类器的输出结果

进一步映射到解调结果

where H(y) is the strong classifier, y is the test sample,

is the coefficient of _Gq . for

I(G _q (y), z) can be used as a vote to label y as z. If I(G _q (y), z) = 1, G _q classifies the sample y as class z, otherwise y does not belong to class z. For all classifiers, have the maximum weighted vote value

The category of is the output of this Adaboost classifier

further map to demodulation result

(6) Experimental results and discussion

End-to-end VLC system prototype:

An end-to-end VLC system prototype is proposed to generate an actual VLC modulation dataset and validate the proposed DL-based demodulation method including source computer, arbitrary function generator, amplifier, DC bias, LED, sliding rail, optoelectronic detector and a mixed-domain oscilloscope, as shown in Figure 8. Table 2 lists the device parameters of the end-to-end VLC system prototype.

Using the proposed end-to-end VLC system prototype, modulation data are collected and an open shared database for ML processing is established. There are eight modulation types (i.e. OOK, QPSK, 4-PPM, 16QAM, 32QAM, 64QAM, 128QAM and 256QAM, as listed in Table 3. For each modulation method, there are four sampling points per cycle, i.e. N = 10, 20, 40, 80. In particular, when using 4-PPM modulation, N=8, 16, 32, 64. Let d represent the distance between the LED and the photodetector. From d=0cm to d=140cm Data is collected every 5cm. The illumination of the ambient light in the dataset is about 85Lux. At d=100cm, the illumination of the LED is 492Lux. In order to reduce the generalization error, 2/3 of the dataset is used as the training set , 1/3 as the test set. The database can be found at https://pan.baidu.com/s/1ThsN3tQaTtFgHryZjqpTuw.

Table 2 Device Parameters

equipment Model/parameter Arbitrary Waveform Generator Tektronix AFG3152C amplifier Mini-Circuits ZHL-6A-S+ Bias SHWBT-006000-SFFF Photodetector PDA10A-EC Mixed Domain Oscilloscope Tektronix MDO3012 led 7.35W

Table 3 Dataset

Experimental results:

The performance of the proposed demodulator based on CNN, DBN and AdaBoost with respect to distance d (N=40) is first investigated. In addition, support vector machine (SVM) based and maximum likelihood (Maximum Likelihood, MLD) based demodulation methods are used as comparison methods.

Figures 8a, 8b, and 8c show the demodulation accuracy of OOK, 32-QAM and 256-QAM modulated signals with respect to distance d, respectively. It can be seen that the demodulation accuracy of all methods decreases as the distance d increases. Specifically, Figure 8a shows that when d≤70cm, the demodulation accuracy of all demodulators for OOK modulated signals is close to 100%; when 70cm<d≤140cm, the proposed AdaBoost-based demodulation method is significantly better than other demodulation methods demodulation method. Figure 8b shows that for d≤40cm, the demodulation accuracy of all demodulators for the 32-QAM modulated signal is close to 100%. For 40cm<d<140cm, the demodulator based on AdaBoost has the highest accuracy among the five demodulation methods, and the accuracy of the demodulator based on DBN and SVM is similar, which is higher than that based on CNN and maximum likelihood. method. This may be because the combined output after the pooling operation of CNN is usually the most active element value (Reference [46]), ignoring the relative position information of each part of the waveform. In addition, the noise of the actual channel does not obey the Gaussian distribution, which may be the reason for the poor performance of the maximum likelihood based demodulation method. Figure 8c shows the demodulation accuracy of all demodulators for a 256-QAM modulated signal, and its performance is similar to that of Figure 8b.

Figure 9 shows the accuracy of the AdaBoost-based demodulator with respect to the distance d, with different sampling points N=10, 20, 40, 80 in one cycle. The signal adopts 32-QAM modulation. As the number of sampling points N increases, the demodulation accuracy increases. In addition, the demodulation accuracy rate when N=40 is better than that when N=80, and the required storage amount when N=40 is only half of that when N=80.

Figure 10 shows the relationship between the demodulation accuracy of the AdaBoost-based demodulator and the number of training cycles K, where the 16-QAM modulated signal is at d=70cm and the 32-QAM modulated signal is at d=60cm. It can be seen that for the 16-QAM modulated signal, the demodulation accuracy increases with the increase of the training period K, and when K≥4000, the demodulation accuracy increases very slowly. Similarly, for the 32-QAM modulated signal, the demodulation accuracy increases with the increase of the training period K, and when K≥8000, the demodulation accuracy remains almost unchanged. Comparing the demodulation accuracy of 16-QAM and 32-QAM modulated signals, more training periods K are required for higher modulation ends.

Figure 11a shows the relationship between the distance d of the demodulation accuracy of the eight modulation signals of OOK, QPSK, 4-PPM, 32-QAM, 64-QAM, 128-QAM and 256-QAM when N=40. It can be seen that the demodulation accuracy of the eight modulation schemes decreases with the increase of the distance d. In addition, for a given distance d, the demodulation accuracy of the eight modulation schemes decreases with the increase of the modulation order, and the higher the modulation order, the faster the decline rate.

Figure 11b shows the effective rate R _eff for the eight modulation schemes. As the distance d increases, the effectiveness of the eight modulation schemes decreases. When d≤30cm, the effective rate of 256-QAM is the highest. When 40cm<d≤50cm, the 128-QAM modulation scheme has the highest effective rate. When the distance d increased from 50 cm to 140 cm, 64-QAM, 32-QAM, and 16-QAM sequentially obtained the highest effective rates. Therefore, for short range or high signal-to-noise ratio situations, higher order modulations are preferred.

In order to overcome the deficiencies of the prior art, the present invention proposes a VLC system design method based on machine learning (ML) (reference [11]). Due to its extensive approximation, learning ability, and adaptive ability, machine learning is able to approximate unknown highly nonlinear and complex functions based on data-driven methods (Reference [12]). Therefore, ML (Machine Learning) has a wide range of applications in many fields, such as computer vision, natural language processing, biomedical engineering, and robotics. This data-driven approach is considered a promising way to think about communication system design from a new perspective in complex communication scenarios.

The invention studies the application of ML in signal demodulation in the physical layer of VLC system. Since the amplitude and phase of the signal represent the information of the signal, feature extraction is crucial for signal demodulation. Therefore, various techniques have been used to extract features of modulated signals.

The present invention provides a physical layer visible light communication method based on machine learning. There are many methods and approaches for implementing the technical solution. The above are only the preferred embodiments of the present invention. In other words, without departing from the principles of the present invention, several improvements and modifications can also be made, and these improvements and modifications should also be regarded as the protection scope of the present invention. All components not specified in this embodiment can be implemented by existing technologies.

Claims (1)

1. a physical layer visible light communication method based on machine learning, is characterized in that, comprises the steps: Step 1, establish a visible light communication VLC system model; Step 2, using any one of the demodulators based on convolutional neural network CNN, deep belief network DBN, adaptive enhancement AdaBoost to demodulate the established VLC system model; Step 1 includes: Step 1-1, build an end-to-end VLC system, which contains a single LED emitter and a single photodetector, the emission signal x(t) is as follows:

Where, t is the time, s(t) is the baseband signal, j is the imaginary unit, fc is the carrier frequency, p( _t ) is the signal pulse, and T is the signal period; Let g denote the channel between the LED and the photodetector, including the direct path and the multi-reflection path, at the receiver, the received signal y(t) is as follows: y(t)=gx(t)+n(t) (2) where n(t) is the received noise, and the received analog signal y(t) is sampled to a digital signal through a digital-to-analog converter; set

represents the vector of signal samples in the ith cycle, where

Represents the nth sampling point:

n ranges from 1 to N, where N is the number of samples in a cycle; Step 1-2, set the training data set to include K received sampling data periods, and i takes a value of 1 to K; before demodulation, the received sampling sequence is

Normalize to the [0,1] interval, as follows:

in,

represents the value of the i-th sampling point after normalization,

represents the minimum value of the sampling sequence,

Represents the maximum value of the sampling sequence; Steps 1-3, for the normalized ith vector

Set its corresponding label z _i , where 1≤i≤K, let

represents the labeled training dataset, let

represents the set of all tags, determined by the modulation used, and

In step 2, the demodulator based on the convolutional neural network includes a visualization module and a CNN, and when the demodulator based on the convolutional neural network is used to demodulate the established VLC system model, it includes the following steps: Step a1, the received data vector

It is converted into a two-dimensional image format by the visualization module and input into a CNN designed for two-dimensional data for image classification. The output image of the visualization module is represented as X, where X is a 28×28 matrix: X∈R ^{28× 28} , R represents the set of real numbers; In step a2, the CNN processing X includes two convolutional layers, two pooling layers and a fully connected layer, using

represents the ith convolution kernel of the first convolutional layer,

Denote by Y _i ¹ by

The resulting feature map is obtained by:

where b _i represents

The bias of , * denotes the convolution operation,

is the activation function,

represents the elements of the matrix obtained after the first convolution,

express

elements of , p=1,2,...,24, q=1,2,...,24; In step a3, the convolutional layer is followed by a pooling layer, which performs a downsampling operation on the feature map output by the previous layer, and uses the maximum pooling method for pooling. The size of the receptive field is 2×2.

Represents the pooling result of the i-th feature map, which is obtained by the following formula:

where pooling( ) represents the downsampling function,

represents the feature map of the ith input; Step a4, use

represents the jth convolution kernel in the second convolutional layer,

set up

is the output of the second convolutional layer,

It is obtained by the following formula:

in

represents the elements of the matrix obtained after the second convolution,

Yes

The elements in , p=1,2,...,10, q=1,2,...,10; Step a5, after the second pooling layer with a receptive field of 2 × 2, its output

After the fully connected layer, it is converted into a one-dimensional label vector y ³ . The number of neurons in this vector is determined by the modulation method. The label output by CNN

Expressed as:

Where [y ³ ] _i represents the value of the i-th dimension element in y ³ , the dimension of y ³ is determined by the modulation method used, and the subscript corresponding to the largest element in y ³ is the label

and can be further mapped to the demodulation result

Step a1 includes: the visualization module performs the following processing:

Each element in is first converted into a point on a two-dimensional plane, and these points are connected with a broken line to obtain the waveform of the received signal. The amplitude information and phase information of the modulating signal are stored in this grayscale image. , use the bicubic interpolation method to reduce the size of the grayscale image, and use the global threshold algorithm to convert the reduced image into a binary image; use X to represent the final output binary image of the visualization module; In step 2, when the demodulator based on the deep belief network is used to demodulate the established VLC system model, the following steps are included: In step b1, a deep belief network with _three restricted Boltzmann machines RBM is established. RBM is an implementation of _{an undirected} graphical model ^. and the hidden layer h=[h ₁ , h ₂ ,..., h _n ] ^T , where vi and h _j represent the value of the i-th unit of the display layer and the value of the _j -th unit of the hidden layer, respectively, i ranges from 1 to m, and j ranges from 1 to n; set W=[w ₁ , w ₂ ,...,w _n ] ^T represents the connection weight matrix between v and h, where w _j =[ w _j1 ,w _j2 ,...,w _jm ] ^T , w _ji represents the connection weight between vi and _{h j ;} a=[a ₁ ,a ₂ ,..., _am ] ^T and b=[ b ₁ ,b ₂ ,...,b _n ] ^T represents the bias of v and the bias of h, respectively, where a _i represents the bias of v _i , and b _j represents the bias of h _j ; Step b2, introduce the energy function to represent the state of the RBM, and use the training data set

normalized signal in

The energy function E(v,h) of the first RBM is as follows: E(v,h)=-a ^T vb ^T hh ^T Wv, (8) in,

The marginal distribution p(v) of the display layer v is expressed as follows:

in,

is the normalization factor; In step b3, the optimal parameters W, a, b are obtained by maximizing the following unconstrained log-likelihood function:

In step b4, the gradient descent method is used to solve the optimization problem in step b3, and the variables W, a, and b are updated as follows:

Among them, ε represents the learning rate, ΔW, Δa and Δb represent the partial derivative of the objective function to W, the partial derivative of a and the partial derivative of b, respectively; In step b5, the partial derivatives of the variables W, a, and b are respectively approximated as:

in,

represents the reconstructed explicit data, where

is the value of the reconstructed i-th explicit layer neuron, i is 1 to m; p(h _j = 1|v) means that for a given explicit layer v, the j-th neuron of the hidden layer is activated The probability;

Represents the explicit layer after a given reconstruction

The probability that the jth neuron of the hidden layer is activated;

Obtained by: Given the explicit layer v, the distribution of the units in the hidden layer h is as follows:

According to the distribution of formula (13), the hidden layer data is generated according to the following formula

in

is the value of the jth neuron of the hidden layer, j=1,2,...,n, then:

Among them, p(h|v) represents the probability of obtaining the hidden layer state h given the explicit layer v; for a given hidden layer

The probability that the i-th unit of the display layer v is activated

is given by:

Step b6, reconstructed display layer data

Produced by the (15) distribution, as follows:

In step b7, after obtaining the optimal parameters W, a, b of the first RBM by using the gradient descent method, the hidden layer h of the first RBM is regarded as the explicit layer of the second RBM, and h ⁽¹⁾ is the first RBM. The hidden layer of two RBMs; after training the weight matrix and bias of the second RBM, consider h ⁽¹⁾ as the explicit layer of the third RBM, and let h ⁽²⁾ be the hidden layer of the third RBM; After training the third RBM, all parameters of the RBM are fine-tuned by a supervised back-propagation algorithm; in the testing phase, DBN is applied to the signal demodulation, and the signal demodulation result is output

In step 2, when adopting the demodulator based on adaptive enhancement AdaBoost to demodulate the established VLC system model, the following steps are included: Step c1, set the strong classifier to be composed of Q k nearest neighbor KNN classifiers, let k=1; for the qth KNN classifier, the training data set

_{The weights of} ^all the _{samples in} the 2,...,Q; when q=1, d _q,i =1/K, i=1,2,...,K; Step c2, according to d _q pair

After resampling, we get

is the training set of the qth KNN classifier; set

where (x _q,i ,z _q,i ) is the ith sample after resampling, x _q,i is the data vector, z _q,i is the corresponding label, there are

is the test set for each KNN classifier; use

Indicates that the training data set is far from the test sample

The most recent sample, namely:

in

is x _q,i and

Euclidean distance between, set

The label is

KNN classifier will

Classify

Step c3, use

Represents the qth KNN classifier, that is, the qth KNN classifier pair sample

The classification result is

The error e _q of the qth KNN classifier is defined as the sum of the weights of the misclassified samples:

where I(a,b) is the indicator function, defined as follows:

Let d _q+1 =[d _q+1,1 ,d _q+1,2 ,...,d _q+1,K ] ^T denotes the training data set corresponding to the q+1th KNN classifier

The weight of the sample in , where d _q+1,i represents the weight of the ith sample, i=1,2,...,K; d _q+1 is obtained by the following formula:

where β _q is determined by the function e _q :

Calculated; β _q < 1 under constraint e _q <0.5; if

is classified correctly, there are

if

misclassified,

To re-evaluate the weights of the samples, redefine d _q+1,i by the following normalization formula:

Step c4, after generating Q KNN classifiers, the strong classifier is defined by the following formula:

Among them, H(y) represents the classification result of the test sample y by the strong classifier.

express;

is the set of all labels, z is the label;

is the coefficient of G _q ; for

I(G _q (y), z) as the vote to label y as z; if I (G _q (y), z) = 1, G _q classifies sample y as label z, otherwise y does not belong to label z ; for all KNN classifiers, with the maximum weighted vote value

The category of is the output of this Adaboost classifier

Further get demodulation results

Images