Sending and detecting device of PAM visible light communication system
By using single-hot encoding, encoding neural network and channel matrix module in the PAM visible light communication system, combined with deep learning technology, an automatic encoder is built, which solves the problem of high bit error rate in optical communication and achieves efficient and versatile Multi-digital modulation communication.
Patent Information
- Application Number
- CN202410955292.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-17
- Publication Date
- 2024-11-26
AI Technical Summary
The existing optical communication channel coding schemes are difficult to meet the bit error rate requirements of high-speed and high-capacity wireless communications, and are especially unable to be suitable for high-order modulation.
A transmission and detection device for PAM visible light communication system was designed, using a one-hot encoding module, a receiving series module, an encoded neural network and a multi-digital and binary operation module, combining a channel matrix module and an additive noise module to utilize depth Learning technology to build an automatic encoder to realize channel encoding processing.
It achieves a lower average bit error rate, improves communication efficiency and versatility, is suitable for multi-digital modulation, and has better performance than traditional solutions.
Smart Images

Figure CN119030621A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of digital communication and digital storage, and in particular relates to a sending and detecting device of a PAM visible light communication system. Background Art
[0002] Visible light communication is a hot research field, and the proposed communication methods must meet lighting-related requirements. Visible light communication is a rapidly developing technology with advantages such as high security, high bandwidth, low power consumption, and no electromagnetic radiation hazards. Compared with traditional electrical signal transmission, optical communication has higher transmission rate, lower transmission loss and wider frequency band. The development of optical communication technology, especially visible light communication technology, has provided new possibilities for high-speed and high-capacity wireless communication. For optical communication, the bit error rate of its channel coding is required to be extremely low, and traditional coding schemes are difficult to meet this requirement.
[0003] As an artificial intelligence algorithm, neural networks have achieved remarkable results in image recognition, speech recognition, natural language processing and other fields in recent years. Neural networks have powerful parallel computing capabilities and can handle complex nonlinear problems, making them very suitable for signal processing and demodulation in optical communications. In recent years, the application of machine learning technology in communication systems has attracted widespread attention. As a powerful machine learning method, deep learning has been widely used in various fields in recent years, including the optimization of physical layer communication systems. In the field of visible light communication, deep learning technology has been widely used to design end-to-end visible light communication systems, optimize the physical layer design of visible light communication systems, and improve the signal processing methods of visible light communication systems.
[0004] Optical communication channel coding technology based on neural networks has the advantages of low bit error rate. In 2018, Karanov B, Chagnon M, Thouin F et al. proposed to implement the optical fiber communication system as an end-to-end deep neural network in the paper "End-to-end deep learning of optical fiber communications" (Journal of Lightwave Technology, 2018, 36 (20): 4843-4855.). This method can produce powerful and flexible transceivers and can achieve reliable transmission in the case of large-scale link dispersion without reconfiguration. In 2019, Lee H, Lee SH, Quek TQ S et al. applied deep learning technology to the design of optical wireless transceivers in their paper "Deep learning framework for wireless systems: Applications to optical wireless communications" (IEEE Communications Magazine, 2019, 57 (3): 35-41.), and proposed a deep learning framework for emerging image sensor communications. Existing schemes are only applicable to binary modulation and cannot be applied to high-order modulation. Summary of the invention
[0005] In order to reduce the average bit error rate in an optical communication transceiver, the present invention provides a sending and detection device for a PAM (Pulse Amplitude Modulation, abbreviated as PAM) visible light communication system. Under the same channel coding processing, the present invention has better performance.
[0006] To achieve the above object, the technical solution of the present invention is:
[0007] A sending and detecting device of a PAM visible light communication system, comprising a transmitter, a channel and a receiver;
[0008] The transmitter includes a one-hot encoding module, a receiving cascade module, an encoding neural network, and a multi-base and binary operation module; the one-hot encoding module is used to map the input information b to obtain a one-hot vector e b The receiving series module is used to b Combined with the dimming target d, the combined vector The combined vector is passed to the encoding neural network, which outputs a continuous value vector u b,d ; The multi-base and binary operation module converts the continuous value vector ub,d Binarization generates a binary output vector Then multiply the matrix H' to get the multi-binary output vector s b,d
[0009] The channel includes a channel matrix module and an additive noise module. The channel matrix module is used to convert s b,d multiplied by the channel matrix H, the additive noise module is used to superimpose additive noise n on the output end of the channel matrix module; the channel output r;
[0010] The receiver includes a receiving series module, a decoding neural network and an argmax function module, wherein the receiving series module is used to combine and connect the channel output r and the dimming target d to obtain a combined vector [r T ,d] T And as the input of the decoding neural network, the decoding neural network outputs [p] b , the code neural network output [p] b After the argmax function module, the information is obtained
[0011] Compared with the prior art, the present invention has the following beneficial effects:
[0012] The present invention proposes a sending and detecting device for a PAM visible light communication system, which has the advantages of high communication efficiency, strong versatility, low bit error rate, etc. BRIEF DESCRIPTION OF THE DRAWINGS
[0013] Figure 1 A schematic diagram of a dimmable VLC (Visible Light Communications, abbreviated as VLC) system of the present invention;
[0014] Figure 2 Schematic diagram of the DNN-based adjustable light PAM modulation VLC system of the present invention;
[0015] Figure 3 The bit error rate of the 4-bit 4-ary dimmable PAM without channel coding of the present invention;
[0016] Figure 4 The bit error rate of the 3-bit 4-ary dimmable PAM without channel coding of the present invention;
[0017] Figure 5 The bit error rate of the 4-bit 4-ary dimmable PAM after channel coding of the present invention;
[0018] Figure 6 The bit error rate of the 3-bit 4-ary adjustable optical PAM after channel coding of the present invention. DETAILED DESCRIPTION
[0019] Example:
[0020] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.
[0021] See also Figure 1-2 As shown, the sending and detecting device of the PAM visible light communication system provided in this embodiment mainly includes a transmitter, a channel and a receiver.
[0022] The transmitter includes a one-hot encoding module, a receiving cascade module, an encoding neural network, and a multi-base and binary operation module; the one-hot encoding module is used to convert the input information Mapping is performed to obtain a unique hot vector e b , receiving series module is used to convert e b Combined with the dimming target d, the combined vector The combined vector is passed to the encoding neural network, which outputs a continuous value vector u b,d , that is, the message b is first mapped to a one-hot vector (One-Hot) representing e b , then the combined vector in Passed to the encoding neural network, the encoding neural network encodes the input (b, d), and the encoding neural network outputs a continuous value vector in, represents the output of the hidden layer l of the encoding neural network, C E,l is the output dimension of the lth hidden layer, where l=1,…,L E , which is a hyperparameter that needs to be optimized through verification. The multi-level and binary operation module converts the continuous value vector u b,d Binarization generates a binary output vector Then multiply the matrix H' to get the multi-binary output vector s b,d , this binarization is a random binarization method. When it is necessary to generate N-bit M-ary, k×N-bit binary can be generated first, where M=2 k .
[0023] The multi-ary output vector Where H' is represented by:
[0024]
[0025] The elements of the noise q are drawn from a Bernoulli distribution:
[0026]
[0027] Among them, h([ub,d ] i ) is the Sigmoid function, that is
[0028] The channel includes a channel matrix module and an additive noise module. The channel matrix module is used to convert the multi-order output vector s b,d Multiplying with the channel matrix H, the additive noise module is used to superimpose additive noise n on the output of the channel matrix module; finally, the channel output r; that is, the transmitter sends a multi-binary output vector s b,d , after passing through the Gaussian white noise channel, it reaches the receiving end and obtains the channel output r.
[0029] The receiver includes a receiving series module, a decoding neural network and an argmax function module. The receiving series module is used to combine and connect the channel output r and the dimming target d to obtain a combined vector [r T ,d] T ∈R (N+1)×1 And as the input of the decoding neural network, the decoding neural network outputs [p] b :
[0030]
[0032] Where p is the probability vector, p∈R z×1 , is the output of the hidden layer of the decoding neural network. Finally, the decoding neural network outputs [p] b After the argmax function module, the information is obtained The estimated value is
[0033] In one embodiment, the encoding neural network is composed of parameters L E hidden layers, where denote the weights and biases of the encoded neural network, C E,l Indicates the dimension of the output, L E is a positive integer. The decoding neural network consists of parameters L D hidden layers, where and Represent the weight and bias of the decoding neural network, C D,l Indicates the dimension of the output, L D is a positive integer. The encoding neural network and the decoding neural network are collectively referred to as an autoencoder. The neural network parameters of the autoencoder are trained using a constrained training method, which specifically includes the following steps:
[0034] Step 1: Initialize t=0, L best =∞, and λ [0] . Where t is the number of training iterations, L best is the smallest Lagrangian, and are the parameters of the initial encoding neural network and the decoding neural network, λ [0] is the initial dual solution.
[0035] Step 2: For each minimum batch, set t = t + 1. Update parameter Z [t] as well as in,
[0036]
[0037] Respectively represent the parameters of the encoding neural network and the decoding neural network of the tth iteration, λ [t] represents the dual solution of the tth iteration. η represents the learning rate,
[0038] Step 3: Calculate Lagrangian on the validation data set
[0039] Step 4: If but And save the deep neural network parameters
[0040] Step 5: Repeat steps 2 to 4 until convergence.
[0041] In a specific embodiment, the noise vector n in the strength vector r received by the receiver has a mean of 0 and a variance of σ 2 The activation function of the encoding neural network is the rectified linear unit, and the output layer uses a linear activation function with dimension KN; the activation of the hidden layer of the decoding neural network is given by the ReLU function, and the output layer is implemented by the Softmax activation function of dimension Z; the cost function is the classification cross entropy between the transmitted message b∈Z and the probability vector p, that is, One-hot vector e b ∈R z×1 The vth element of is equal to 1, and the other elements are set to 0.
[0042] In summary, the present invention provides an input message and dimming target The intensity vector r∈R received by the receiver N×1 It can be expressed as: r = HS b,d +n. Among them, 0≤d1<… <d D≤0.5, N represents the length of the message b encoded, H∈R N×N is an optical communication channel matrix, S b,d is the M-ary vector of length N of the encoded output, and n is the noise vector. In the PAM system, each message b is encoded as an M-ary vector S of length N. b,d , the receiver obtains the estimated message according to the received signal vector r The joint design of the encoding and decoding networks is called an autoencoder. The encoding neural network of the autoencoder is composed of parameters L E hidden layers, where denote the weights and biases of the encoded neural network, C E,l Indicates the dimension of the output, L E is a positive integer. The decoding neural network of the automatic decoder is composed of parameters L D hidden layers, where and Represent the weight and bias of the decoding neural network, C D,l Indicates the dimension of the output, L D is a positive integer. The neural network parameters of the autoencoder are trained using a constraint training algorithm, and the constraint formula is in ρ is the penalty parameter, which can be any decimal number. d (θ E ) is the dimming constraint function on the DNN transmitter, d is the dimming target, Using the Lagrange duality method, the solution of the constraint formula is obtained as L(θ E ,θ D ,λ)=C(θ E ,θ D )+Σ d∈D λ d (F d (θ E )-d)+ρ∑ d∈D (F d (θ E )-d) 2 , where the non-negative parameter λ d Represents the Lagrange multiplier associated with the constraint.
[0043] The technical advantages of the present invention are described below in conjunction with two application examples:
[0044] Application Example 1:
[0045] Setting L E=3 hidden layers to construct the encoding neural network, each hidden layer has dimensions of 8192, 4096, and 2048 respectively. The decoding neural network has an inverse structure to the encoding neural network, so that the L with output dimensions of 2048, 4096, and 8192 D =3 hidden layers. The ADAM algorithm is used to train the DNN, and the number of training and samples is set to 5×10 5 ×64. The selected message b is encoded into a 4-ary vector with a length of 4, that is, M=4, N=4, K=2. The dimming level is set to 0.25. The encoding method of the present invention comprises the following steps:
[0046] Step 1: Message b is first mapped to a one-hot vector representation e b , then the combined vector in Passed to the encoding neural network, the encoding neural network encodes the input (b, d), and the encoding neural network outputs a continuous value vector u b,d .
[0047] Step 2: Use a multi-multiplication binarization method. b,d Binarization generates a binary output vector Multiplying with matrix H' gives the multi-ary output vector s b,d , where H' is expressed as:
[0048]
[0049] The binarization is a random binarization method. When a 4-bit 4-bit system is required, a 2×4-bit binary system can be generated first, where M=2 2 . In the output u of the encoding neural network b,d Add quantization noise q to The elements of the noise q are selected from a Bernoulli distribution, where i = 1, ..., 8:
[0050]
[0051] in
[0052] Step 3: Send multi-binary output vector s b,d , after passing through the Gaussian white noise channel, it reaches the receiving end and obtains the channel output r.
[0053] Step 4: Connect the channel output and dimming target T ,d] T ∈R 5×1 Input the decoding neural network. The output of the decoding neural network p is then expressed as:
[0054]
[0055] in is the output of the hidden layer of the decoding network. Finally, we get the information The estimated value is
[0056] Figure 3 The relationship between the bit error rate and signal-to-noise ratio of the VLC transceiver based on the deep neural network PAM technology and the tunable optical PAM scheme based on the identity code in this example is given, and both methods do not perform channel coding processing. The bit error rate performance of the scheme proposed by the present invention is better than that of the comparative scheme. -6 The performance is 0.4dB better.
[0057] Figure 5 The relationship between the bit error rate and signal-to-noise ratio of the VLC transceiver based on the deep neural network PAM technology and the tunable optical PAM scheme based on the identity code in this example is given, and both methods perform channel coding processing. -6 The performance is 0.65 dB better when . This means that the coding network generates a codebook that is robust to optical noise by learning the random characteristics of the optical channel.
[0058] Application Example 2:
[0059] Setting L E =3 hidden layers to construct the encoding neural network, each hidden layer has dimensions of 8192, 4096, and 2048 respectively. The decoding neural network has an inverse structure to the encoding neural network, so that the L with output dimensions of 2048, 4096, and 8192 D =3 hidden layers. The ADAM algorithm is used to train the DNN, and the number of training and samples is set to 5×10 5 ×64. The selected message b is encoded into a 4-ary vector with a length of 3, that is, M=4, N=3, K=2. The dimming level is set to 0.25. The encoding method of the present invention comprises the following steps:
[0060] Step 1: Message b is first mapped to a one-hot vector representation e b . Then the combined vector in Passed to the encoding neural network, the encoding neural network encodes the input (b, d), and the encoding neural network outputs a continuous value vector u b,d .
[0061] Step 2: Use a multi-multiplication binarization method. b,d Binarization generates a binary output vector Multiplying with matrix H' gives the multi-ary output vector s b,d , where H' is expressed as:
[0062]
[0063] The binarization is a random binarization method. When a 3-bit quaternary system is required, a 2×3-bit binary system can be generated first, where M=2 2 . In the output u of the encoding neural network b,d Add quantization noise q to The elements of the noise q are selected from a Bernoulli distribution, where i = 1, ..., 6:
[0064]
[0065] in
[0066] Step 3: Send multi-binary output vector s b,d , after passing through the Gaussian white noise channel, it reaches the receiving end and obtains the channel output r.
[0067] Step 4: Connect the channel output r and dimming target d [r T ,d] T ∈R 4×1 Input the decoding neural network. The output of the decoding neural network P is then expressed as:
[0068] in is the output of the hidden layer of the decoding network. Finally, we get the information The estimated value is
[0069] Figure 4 The relationship between the bit error rate and signal-to-noise ratio of the VLC transceiver based on the deep neural network PAM technology and the tunable optical PAM scheme based on the identity code in this example is given, and both methods do not perform channel coding processing. The bit error rate performance of the scheme proposed by the present invention is better than that of the comparative scheme. -6 The performance is 0.2dB better.
[0070] Figure 6 The relationship between the bit error rate and signal-to-noise ratio of the VLC transceiver based on the deep neural network PAM technology and the tunable optical PAM scheme based on the identity code in this example is given, and both methods perform channel coding processing. -6 The performance is 0.35dB better.
[0071] Combining two application examples, we can see that:
[0072] 1. The sending and detecting device of a PAM visible light communication system proposed in the present invention can obtain gain in various coding schemes compared with the constant sum code scheme, whether it is channel coded or not, whether it is three-bit quaternary or four-bit quaternary, and can be applied to different types of channel coding schemes, with strong versatility.
[0073] 2. The sending and detection device of a PAM visible light communication system proposed in the present invention has a lower bit error rate and performance compared with existing methods.
[0074] 3. The sending and detection device of a PAM visible light communication system proposed in the present invention is suitable for multi-level modulation.
[0075] In summary, the present invention proposes a sending and detection device for a PAM visible light communication system, which has the advantages of high communication efficiency, strong versatility, and low bit error rate.
[0076] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable ordinary technicians in the field to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent changes or modifications made based on the essence of the content of the present invention should be included in the protection scope of the present invention.
Claims
1. A sending and detecting device for a PAM visible light communication system, characterized in that: Includes transmitter, channel and receiver; The transmitter includes a one-hot encoding module, a receiving cascade module, an encoding neural network, and a multi-base and binary operation module; the one-hot encoding module is used to map the input information b to obtain a one-hot vector e b The receiving series module is used to b Combined with the dimming target d, the combined vector The combined vector is passed to the encoding neural network, which outputs a continuous value vector u b,d ; The multi-base and binary operation module converts the continuous value vector u b,d Binarization generates a binary output vector Then multiply the matrix H' to get the multi-binary output vector s b,d ; The channel includes a channel matrix module and an additive noise module. The channel matrix module is used to convert the multi-ary output vector s b,d multiplied by the channel matrix H, the additive noise module is used to superimpose additive noise n on the output end of the channel matrix module; the channel output r; The receiver includes a receiving series module, a decoding neural network and an argmax function module, wherein the receiving series module is used to combine and connect the channel output r and the dimming target d to obtain a combined vector [r T ,d] T And as the input of the decoding neural network, the decoding neural network outputs [p] b , the code neural network output [p] b After the argmax function module, the information is obtained 2. The sending and detecting device of the PAM visible light communication system according to claim 1, characterized in that: The encoding neural network is parameterized by L E hidden layers, where denote the weights and biases of the encoded neural network, C E,l Indicates the dimension of the output, L E Is a positive integer.
3. The sending and detecting device of the PAM visible light communication system according to claim 1 or 2, characterized in that: The decoding neural network is composed of parameters L D hidden layers, where and Represent the weight and bias of the decoding neural network, C D,l Indicates the dimension of the output, L D Is a positive integer.
4. The sending and detecting device of the PAM visible light communication system according to claim 1, characterized in that: The encoding neural network and the decoding neural network are collectively referred to as an autoencoder, and the neural network parameters of the autoencoder are trained using a constrained training method.
5. The sending and detecting device of the PAM visible light communication system according to claim 4, characterized in that: The constraint training method comprises the following steps: Step 1: Initialize t=0, L best =∞, and λ [0] ; where t is the number of training iterations, L best is the smallest Lagrangian, and are the parameters of the initial encoding neural network and the decoding neural network, λ [0] is the initial dual solution; Step 2: For each minimum batch, set t = t + 1; update parameter Z [t] as well as in, Respectively represent the parameters of the encoding neural network and the decoding neural network of the tth iteration, λ [t] represents the t-th iteration dual solution, η represents the learning rate, Step 3: Calculate Lagrangian on the validation data set Step 4: If but And save the deep neural network parameters Step 5: Repeat steps 2 to 4 until convergence.
6. The sending and detecting device of the PAM visible light communication system according to claim 1 or 2, characterized in that: The activation function of the encoding neural network is a rectified linear unit, and the dimension of the output layer and the linear activation function is kN.
7. The sending and detecting device of the PAM visible light communication system according to claim 1 or 3, characterized in that: The activation of the hidden layer of the decoding neural network is given by the ReLU function, and the output layer is realized by the Softmax activation function of dimension Z.
8. The sending and detecting device of the PAM visible light communication system according to claim 1, characterized in that: The one-hot vector e b ∈R z×1 The b-th element of is equal to 1, and the other elements are set to 0.
9. The sending and detecting device of the PAM visible light communication system according to claim 1, characterized in that: The continuous-valued vector in, represents the output of the hidden layer l of the encoding neural network, C E,l is the output dimension of the lth hidden layer, where l=1,…,L E , is a hyperparameter that needs to be optimized through validation.
10. The sending and detecting device of the PAM visible light communication system according to claim 1, characterized in that: The multi-ary output vector Where H' is represented by: The elements of the noise q are drawn from a Bernoulli distribution: in, is the Sigmoid function, that is
