CN112865866B - Visible light PAM system nonlinear compensation method based on GSN - Google Patents
Visible light PAM system nonlinear compensation method based on GSN Download PDFInfo
- Publication number
- CN112865866B CN112865866B CN202110073541.7A CN202110073541A CN112865866B CN 112865866 B CN112865866 B CN 112865866B CN 202110073541 A CN202110073541 A CN 202110073541A CN 112865866 B CN112865866 B CN 112865866B
- Authority
- CN
- China
- Prior art keywords
- data
- classifier
- training
- pam
- loss
- 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
Images
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
-
- 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
- H04B10/524—Pulse modulation
-
- 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
Landscapes
- Physics & Mathematics (AREA)
- Electromagnetism (AREA)
- Engineering & Computer Science (AREA)
- Computer Networks & Wireless Communication (AREA)
- Signal Processing (AREA)
- Optical Communication System (AREA)
Abstract
The invention belongs to the technical field of visible light communication, and particularly relates to a visible light PAM system nonlinear compensation method based on GSN, which comprises the following steps: inputting the received PAM signal into a GSN-based nonlinear equalization module for nonlinear compensation to obtain a compensated PAM signal, and converting the compensated PAM signal into a binary signal through PAM demapping; the GSN-based nonlinear equalization module consists of an auxiliary classifier network and a classifier network; the auxiliary classifier network mainly performs characteristic mapping on the received data; the classifier network is a multi-classifier, and the compensated PAM signal level is obtained through a classification method; the updating of 2 network parameters in the invention can be influenced by each other, and finally the dynamic balance is achieved, thereby preventing the occurrence of the overfitting phenomenon of the system, reducing the error rate and improving the transmission rate of the system.
Description
Technical Field
The invention belongs to the technical field of visible light communication, and particularly relates to a visible light PAM system nonlinear compensation method based on GSN.
Background
Visible Light Communication (VLC) is an emerging free-space wireless Communication technology that combines LED lighting technology with Communication technology. The traditional wireless communication faces the dilemma of gradually lacking wireless spectrum resources, and the working frequency of the visible light communication system is about 400THz, so that the problem is well made up. At present, visible light communication mainly transfers information by controlling brightness change of LED lamps.
The visible light communication system covers comprehensive technologies such as materials (light source, receiver material), optical design (optical antenna), electrical design (driving circuit), modulation and demodulation technology (high-order modulation and demodulation), digital signal processing (signal compensation technology), system integration (visible light communication system) and the like. In VLC Digital Signal Processing (DSP), conventional equalization algorithms include LMS (least mean square algorithm), RLS (recursive least square algorithm), CMA (constant modulus blind equalization algorithm), and the like, and there are many algorithms that are improved based on the above algorithms, such as DD-LMS (direct decision-minimum mean square error algorithm), CMMA (cascaded multi-mode algorithm), MCMMA (improved cascaded multi-mode algorithm), and the like. In recent years, machine learning algorithms have also been increasingly applied to digital signal processing, including: the method comprises the following steps of judging a boundary based on a K-means post-equalization algorithm perception boundary (CAPD) aiming at a standard constellation point, judging a boundary based on a K-means pre-equalization algorithm, aiming at PAM signal jitter, aiming at a DBSAN algorithm, aiming at phase deviation, and the like. In addition to the conventional machine learning algorithm, the deep learning method is also gradually applied to VLC, such as LSTM (long short term memory) algorithm that solves the problem of long short term memory. In terms of modulation and demodulation, visible light mainstream modulation includes Pulse Amplitude Modulation (PAM), Quadrature Amplitude Modulation (QAM), Orthogonal Frequency Division Multiplexing (OFDM), and the like. Compared with QAM and OFDM systems, the PAM system has lower cost and has important cost performance characteristics of a terminal communication system, so that the PAM system is widely applied to visible light communication systems, nonlinear distortion generated by PAM signals in the transmission process is always an important influence on the transmission rate of high-speed visible light, and the problem of nonlinearity in practical application is particularly important to solve. The traditional least Mean square algorithm LMS (least Mean Square) can solve the linear problem of the system, such as intersymbol interference (ISI), but can not solve the nonlinear problem well. An improved LMS algorithm is proposed in the patent "a visible light communication system based on sinusoidal function variable step length LMS equalization (application number CN 202010276743.7)", which considers the problem of tap coefficient convergence speed and has faster tap coefficient convergence speed than the conventional LMS algorithm. The system does not take into account non-linear effects.
In addition, a method for performing system nonlinear modeling by using a multilayer neural network is proposed in the patent "nonlinear channel modeling method for visible light communication system based on neural network (application number CN 202010044920.9)" to solve the nonlinear problem generated in visible light communication. The neural network is trained to express the nonlinear characteristic of a channel, so that an input n-dimensional vector can be converted into an estimated result to be output. However, the design of the network structure is realized by adopting a multi-layer full connection mode, the problems of over-fitting and under-fitting are not considered, and the model of the method has the memory problem in the training process.
Disclosure of Invention
In order to solve the problems in the prior art, the invention provides a visible light PAM system nonlinear compensation method based on GSN, which comprises the following steps: the physical layer receives signals point to point, the received signals are input into a GSN-based nonlinear equalization module for nonlinear compensation, compensated PAM signals are obtained, and the compensated PAM signals are converted into binary signals through PAM demapping; outputting the binary signal to obtain an output signal; wherein, GSN represents a generation auxiliary network (generic support Networks), PAM represents Pulse Amplitude Modulation (Pulse Amplitude Modulation);
the GSN-based nonlinear equalization module consists of an auxiliary classifier network and a classifier network; the auxiliary classifier network mainly performs characteristic mapping on the received data; the classifier network is a multi-classifier, and the compensated PAM signal level is obtained through a classification method.
Preferably, the auxiliary classifier network comprises an input layer, a hidden layer and an output layer; the input layer is a 19-dimensional vector; the hidden layer comprises two layers of structures which are respectively formed by fully connecting 64 neurons and 32 neurons; the output layer is a 9-dimensional vector.
Further, the hidden layer of the auxiliary classifier network selects a ReLU function as an activation function, the output layer adopts a tanh function as the activation function, and the formula of the ReLU function is as follows:
the formula for the tanh function is:
preferably, the classifier network comprises 2 one-dimensional convolutional layers and a full-link layer; the size of the filter of the first one-dimensional convolutional layer is 5, and the size of the filter of the second one-dimensional convolutional layer is 3; both one-dimensional convolutional layers use LeakyReLU as the activation function.
Preferably, the training of the GSN based non-linear equalization module comprises:
step 1: acquiring an original signal data set, preprocessing the original signal data set, and dividing the preprocessed data to obtain a training data set;
step 2: selecting a training data from the training data set, and inputting a training sequence of an auxiliary classifier corresponding to the selected training data into an auxiliary classifier network to obtain 9-dimensional vector output; taking the output as a negative sample;
and step 3: inputting the negative sample into a classifier network to obtain the signal level classification loss s _ c _ loss of the negative sample;
and 4, step 4: inputting a corresponding classifier training sequence in the training data into a classifier network to obtain the signal level classification loss c _ loss of the positive sample;
and 5: calculating the total loss according to the signal level classification loss s _ c _ loss of the negative sample and the signal level classification loss c _ loss of the positive sample, and updating the parameters of the classifier network according to the total loss;
step 6: randomly selecting a training data in a training set, and inputting a training sequence of a corresponding auxiliary classifier in the training data into an auxiliary classifier network;
and 7: inputting the output of the auxiliary classifier network into the classifier network, and outputting a classification result to obtain the loss s _ c __ loss; updating network parameters of the auxiliary classifier according to the loss;
and 8: and obtaining a trained model after traversing all the training data in the training set.
Further, the process of obtaining the training data set includes:
step 11: acquiring an original data set, wherein data in the original data set comprises a sending sequence x and a receiving sequence y;
step 12: obtaining the data length L of the transmission sequence xx(ii) a Initializing segmentation times i;
step 13: acquiring the size of a sliding window of received data and the size of a sliding window of sent data, wherein the size of the sliding window of the received data is LgThe size of the sliding window of the sending data is LdAnd calculating the translation length of the initial segmentation subscript of the sending data relative to the initial segmentation subscript of the receiving data according to the sizes of the sliding window of the receiving data and the sliding window of the sending data, wherein the calculation formula is as follows: t ═ Lg-Ld) /2, e.g. starting slicing index of received sequence is yiThen the initial slicing index of the transmitted data is xi+t；
Step 14: judging the data length i + Lg-1 and the data length L of the transmission sequence xxIf i + Lg-1 is less than LxIf not, executing step 15, otherwise, executing step 16;
step 15: dividing a transmitting sequence and a receiving sequence by adopting a sliding window, and subscript y of the receiving sequencei,yi+1,…,yi+Lg-1Sending sequence subscript x as auxiliary classifier training datai+t,xi+t+1,…,xi+t+Ld-1As classifier training sequence, xi+(Lg-1)/2As the labels of the sending sequence and the receiving sequence, storing the division result into a training set, adding 1 to the division times, and returning to the step 14;
step 16: and outputting the training set.
Further, a loss function of the multi-classification cross entropy calculation system is adopted, and the expression of the loss function is as follows:
the invention is suitable for the visible light communication system with nonlinear phenomenon, the invention changes the received distortion signal into the normal signal before sending through the nonlinear compensation of the distortion signal in the visible light PAM system, and solves the nonlinear problem; the updating of 2 network parameters in the invention can be influenced by each other, and finally the dynamic balance is achieved, thereby preventing the occurrence of the overfitting phenomenon of the system, reducing the error rate and improving the transmission rate of the system.
Drawings
FIG. 1 is a PAM-visible light communication system topology diagram using GSN-based nonlinear compensation method according to the present invention;
FIG. 2 is a network structure of the visible light PAM system nonlinear compensation method based on GSN of the present invention;
FIG. 3 is a flow chart of the process of segmenting data of the present invention;
FIG. 4 is a flow chart of the inventive network training process.
Detailed Description
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings, and the described embodiments are only a part of the embodiments of the present invention, but not all of the embodiments. All other embodiments, which can be obtained by a person skilled in the art without inventive effort based on the embodiments of the present invention, are within the scope of the present invention.
The PAM-visible light communication system topology diagram is shown in FIG. 1, and the system comprises a waveform generator, a hardware pre-equalizer, two amplifiers, a bias device, an LED, a diaphragm, a filter, a PIN receiver and a signal compensation module; the waveform generator generates signals, pre-equalization processing is carried out through a hardware pre-equalizer, and the processed signals are transmitted to the biaser after passing through the first amplifier; the biaser processes the amplified signal to enable the voltage intensity of the incidental signal to reach a starting voltage threshold of the LED; the LED sends out optical signals after receiving the signals, and the optical signals are filtered through the diaphragm and the filter; and a PIN receiver is adopted to receive the signal, the optical signal is converted into an electric signal, and the electric signal passes through a second amplifier and then is input into a signal compensation module for signal compensation to obtain a compensated signal.
The signal compensation module comprises a signal receiving module, a GSN-based nonlinear equalization module, a PAM demapping module and a signal output module, wherein the GSN-based nonlinear equalization module corrects a received distorted signal.
The system has a logic structure of a PAM system, an obtained original binary bit stream is converted into a corresponding PAM signal through PAM mapping, intensity modulation is carried out by an LED after upsampling and upconversion processing, an electric signal is converted into an optical signal for transmission, a photodiode PIN is used for receiving, the data which is subjected to synchronization, downsampling and downconversion is input into a visible light PAM system nonlinear compensation module based on a GSN for signal equalization, finally, the output result of the module is converted into the binary bit stream through PAM demapping, and data transmission is completed.
A visible light PAM system nonlinearity compensation method based on GSN, as shown in fig. 1, the method comprising: a physical layer of the system receives signals point to point, and inputs the received signals into a GSN-based nonlinear equalization module for nonlinear compensation to obtain a compensated PAM signal; converting the compensated signal into a binary signal through PAM demapping; outputting the binary signal to obtain an output signal; where GSN denotes the generation assist network and PAM denotes the pulse amplitude modulation.
A nonlinear equalization module based on GSN is a trained neural network; the neural network consists of an auxiliary classifier network and a classifier network; the auxiliary classifier network mainly performs characteristic mapping on the received data; the classifier network is a multi-classifier, and the compensated PAM signal level is obtained through a classification method.
The generation countermeasure network (GAN) is a deep learning model proposed according to the game theory, and consists of a generation network and a discrimination network. The purpose of training the network is achieved through mutual game between the two networks, and finally the generated network can simulate the sample distribution which is the same as the training set distribution. Based on the mutual countermeasure thought, the invention provides a mutually-assisted neural network GSN.
As shown in fig. 2, the GSN based non-linear equalization module consists of 2 networks, one is an auxiliary classifier network and one is a classifier. The classifier network is a multi-classifier in order to enable the classifier to handle the classification task of multiple signals. For the non-linear compensation problem of the PAM8 signal, the multi-classifier is an 8-classifier, and the classification result is 8 different signal levels, "-7, -5, -3, -1, 1, 3, 5, 7". The auxiliary classifier is realized in a multi-layer perceptron mode, the output dimension is consistent with the input dimension of the classifier network, and the purpose is to enable the classifier network to correctly classify PAM8 signal levels after the distortion vector of the input auxiliary classifier network is mapped through the features of the auxiliary classifier network, so that the nonlinear compensation of signals is completed.
The auxiliary classifier network structure is shown as the auxiliary classifier network portion in fig. 3. Wherein, the input layer is a 19-dimensional vector; the hidden layer has 2 layers and is formed by fully connecting 64 neurons and 32 neurons respectively; the output layer is a 9-dimensional vector.
In the hidden layer, a ReLU function is selected as an activation function, and the activation function formula is as follows:
in the output layer, tanh is used as an activation function, and the activation function formula is as follows:
where x represents the output of the hidden layer.
The output layer may map the output value between-1 and 1 using tanh as an activation function.
The input of the classifier network is a 9-dimensional time series vector, and the classifier network is composed of 2 one-dimensional convolutions and a full-link layer. Wherein, the size of the filter of the first one-dimensional convolution is set to 5, the size of the filter of the second one-dimensional convolution is set to 3, LeakyReLU is used as an activation function, and the activation function formula is as follows:
preferably, a is 0.2.
The input vector is subjected to one-dimensional convolution to obtain 64 characteristic graphs (C1), the size of the characteristic graphs is 5 x 1, the input vector is subjected to one-dimensional convolution to obtain 16 characteristic graphs (C2), the size of the characteristic graphs is 3 x 1, and the output is subjected to full connection and then is subjected to a softmax function to obtain corresponding PAM8 level classification. The number of profiles in this process is chosen as an empirical value.
The calculation formula of the softmax function is as follows:
wherein alpha isiThe value of j is 1 to the total number of output layer neurons as the input of the ith neuron of the output layer.
The loss is calculated by adopting multi-class cross entropy cross-entropy loss, and the calculation formula is as follows:
wherein k represents the PAM signal level, here PAM8 system, so the value of k is "-7, -5, -3, -1, 1, 3, 5, 7", tkIndicating a target class level, t of the target class levelkT equal to 1, other class levelkEqual to 0, P (y ═ k) denotes the probability value obtained by the softmax function in the classifier network, and y denotes the predicted output level.
As shown in fig. 4, the process of training the GSN-based nonlinear equalization module includes: and acquiring a receiving sequence corresponding to the sending sequence, preprocessing the sending sequence and the receiving sequence to obtain a training data set t, wherein the length of the obtained training set data is sample _ length. Firstly, training a classifier network: selecting training data from a training set t, inputting a training sequence of a corresponding auxiliary classifier into an auxiliary classifier network to obtain 9-dimensional vector output as a negative sample, inputting the output into the classifier network, and outputting a result to obtain the signal level classification loss s _ c _ loss of the negative sample; and inputting a classifier training sequence (positive sample) into a classifier network, outputting a result to obtain the signal level classification loss c _ loss of the positive sample, wherein the total loss is c _ loss + s _ c __ loss, and updating the classifier network parameters. Then training the auxiliary classifier network, randomly selecting a training data from the training set, inputting the training sequence of the auxiliary classifier into the auxiliary classifier network, inputting the obtained output into the classifier network to output the classification result, obtaining the loss s _ c __ loss, and updating the network parameters of the auxiliary classifier. And repeating the process until the training set with the length of sample _ length is traversed, and storing the training model.
The fitting process of the model parameters can be completed by multiple times of training, and when the prediction accuracy rate is almost not changed, the highest stored model is taken as the model of the final balance part.
As shown in fig. 3, the process of preprocessing the training data set includes: the transmission sequence is x, and the data length is LxThe corresponding receive sequence is y. L isgTo receive the size of the sliding window of data, in the present invention, L is the length of the input vector into the network of auxiliary classifiersg19, the receive sequence y is of length LgEqual length segmentation is used for training the auxiliary classifier network, and the segmentation result of the ith time is { yi，yi+1，…，yi+Lg-1}；LdTo send the size of the data sliding window, in the present invention, i.e., the length of the input vector, L, into the classifier networkdTransmit sequence x by length L, 9dEqual length segmentation is used for training a classifier network, and the ith segmentation result is { xi+t，xi+t+1，…，xi+t+Ld-1Where t ═ Lg-Ld) 2; intermediate number x of ith segmentation data of transmission sequence xi+(Lg-1)/2As a training label for the set of sliced data.
The above-mentioned embodiments, which further illustrate the objects, technical solutions and advantages of the present invention, should be understood that the above-mentioned embodiments are only preferred embodiments of the present invention, and should not be construed as limiting the present invention, and any modifications, equivalents, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims (5)
1. A visible light PAM system nonlinearity compensation method based on GSN is characterized by comprising the following steps: the physical layer receives signals point to point, and inputs the received signals into a GSN-based nonlinear equalization module for nonlinear compensation to obtain compensated PAM signals; converting the compensated PAM signal into a binary signal through PAM demapping; outputting the binary signal to obtain an output signal; wherein GSN represents generation auxiliary network, PAM represents pulse amplitude modulation;
the GSN-based nonlinear equalization module consists of an auxiliary classifier network and a classifier network; the auxiliary classifier network mainly performs characteristic mapping on the received data; the classifier network is a multi-classifier, and the compensated PAM signal level is obtained through a classification method; the auxiliary classifier network comprises an input layer, a hidden layer and an output layer; the input layer is a 19-dimensional vector; the hidden layer comprises two layers of structures which are respectively formed by fully connecting 64 neurons and 32 neurons; the output layer is a 9-dimensional vector;
the process of training the GSN-based nonlinear equalization module comprises the following steps:
step 1: acquiring an original signal data set, preprocessing the original signal data set, and dividing the preprocessed data to obtain a training data set;
step 2: selecting a training data from the training data set, and inputting a training sequence of an auxiliary classifier corresponding to the selected training data into an auxiliary classifier network to obtain 9-dimensional vector output; taking the output as a negative sample;
and step 3: inputting the negative sample into a classifier network to obtain the signal level classification loss s _ c _ loss of the negative sample;
and 4, step 4: inputting a corresponding classifier training sequence in the training data into a classifier network to obtain the signal level classification loss c _ loss of the positive sample;
and 5: calculating the total loss according to the signal level classification loss s _ c _ loss of the negative sample and the signal level classification loss c _ loss of the positive sample, and updating the parameters of the classifier network according to the total loss;
step 6: randomly selecting a training data in a training set, and inputting a training sequence of a corresponding auxiliary classifier in the training data into an auxiliary classifier network;
and 7: inputting the output of the auxiliary classifier network into the classifier network, and outputting a classification result to obtain the loss s _ c __ loss; updating network parameters of the auxiliary classifier according to the loss;
and 8: and obtaining a trained model after traversing all the training data in the training set.
2. The visible PAM system nonlinearity compensation method of claim 1, wherein the hidden layer of the assisted classifier network selects a ReLU function as an activation function, the output layer uses a tanh function as an activation function, and the ReLU function has the following formula:
the formula for the tanh function is:
where x represents the output of the hidden layer.
3. The visible PAM system nonlinearity compensation method of claim 1, wherein the classifier network comprises 2 one-dimensional convolutional layers and a full link layer; the size of the filter of the first one-dimensional convolutional layer is 5, and the size of the filter of the second one-dimensional convolutional layer is 3; both one-dimensional convolutional layers use LeakyReLU as the activation function.
4. The visible PAM system nonlinearity compensation method based on GSN of claim 1, wherein the obtaining of the training data set comprises:
step 11: acquiring an original data set, wherein data in the original data set comprises a sending sequence x and a receiving sequence y;
step 12: obtaining the data length L of the transmission sequence xx(ii) a Initializing segmentation times i;
step 13: acquiring the size of a sliding window of received data and the size of a sliding window of sent data, wherein the size of the sliding window of the received data is LgThe size of the sliding window of the sending data is LdAnd calculating the translation length of the initial segmentation subscript of the sending data relative to the initial segmentation subscript of the receiving data according to the sizes of the sliding window of the receiving data and the sliding window of the sending data, wherein the calculation formula is as follows: t = (L)g-Ld) If the starting slicing index of the received sequence is yiThen the initial slicing index of the transmitted data is xi+t；
Step 14: judging the data length i + Lg-1 and the data length L of the transmission sequence xxIf i + Lg-1 is less than LxIf not, executing step 15, otherwise, executing step 16;
step 15: dividing a transmitting sequence and a receiving sequence by adopting a sliding window, and subscript y of the receiving sequencei,yi+1,…,yi+Lg-1Sending sequence subscript x as auxiliary classifier training datai+t,xi+t+1,…,xi+t+Ld-1As classifier training sequence, xi+(Lg-1)/2As the labels of the sending sequence and the receiving sequence, storing the division result into a training set, adding 1 to the division times, and returning to the step 14;
step 16: and outputting the training set.
5. The visible light PAM system nonlinearity compensation method of claim 1, wherein the multi-class cross entropy is used to calculate the system loss, and the loss function is expressed as:
where k represents the PAM signal level,
a target class level is indicated which is,
representing the probability value obtained by the softmax function in the classifier network, and y representing the predicted output level.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110073541.7A CN112865866B (en) | 2021-01-20 | 2021-01-20 | Visible light PAM system nonlinear compensation method based on GSN |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110073541.7A CN112865866B (en) | 2021-01-20 | 2021-01-20 | Visible light PAM system nonlinear compensation method based on GSN |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112865866A CN112865866A (en) | 2021-05-28 |
| CN112865866B true CN112865866B (en) | 2022-04-05 |
Family
ID=76007495
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110073541.7A Active CN112865866B (en) | 2021-01-20 | 2021-01-20 | Visible light PAM system nonlinear compensation method based on GSN |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112865866B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114500189B (en) * | 2022-01-24 | 2023-06-20 | 华南理工大学 | Direct pre-equalization method, system, device and medium for visible light communication |
| CN114978313B (en) * | 2022-05-18 | 2023-10-24 | 重庆邮电大学 | Compensation method of visible light CAP system based on Bayesian neurons |
| CN115865199A (en) * | 2022-11-10 | 2023-03-28 | 北京理工大学 | Nonlinear compensation method for optical fiber communication based on residual error neural network |
| CN116346217B (en) * | 2023-05-25 | 2023-08-08 | 北京理工大学 | Deep learning-based optical communication system channel construction method |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109347555A (en) * | 2018-09-19 | 2019-02-15 | 北京邮电大学 | A kind of visible light communication equalization methods based on radial basis function neural network |
| CN110097103A (en) * | 2019-04-22 | 2019-08-06 | 西安电子科技大学 | Based on the semi-supervision image classification method for generating confrontation network |
| WO2019191099A1 (en) * | 2018-03-26 | 2019-10-03 | Zte Corporation | Non-linear adaptive neural network equalizer in optical communication |
Family Cites Families (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11741693B2 (en) * | 2017-11-15 | 2023-08-29 | Palo Alto Research Center Incorporated | System and method for semi-supervised conditional generative modeling using adversarial networks |
| CN109905170B (en) * | 2019-01-17 | 2021-09-17 | 复旦大学 | K-DNN-based nonlinear distortion compensation algorithm and visible light communication device |
| CN109948660A (en) * | 2019-02-26 | 2019-06-28 | 长沙理工大学 | A kind of image classification method improving subsidiary classification device GAN |
| CN110598530A (en) * | 2019-07-30 | 2019-12-20 | 浙江工业大学 | Small sample radio signal enhanced identification method based on ACGAN |
| CN111178260B (en) * | 2019-12-30 | 2023-04-07 | 山东大学 | Modulation signal time-frequency diagram classification system based on generation countermeasure network and operation method thereof |
| CN112036543B (en) * | 2020-07-16 | 2022-05-03 | 北京大学 | Time domain equalizer combining neural network equalization and linear equalization and equalization method |
-
2021
- 2021-01-20 CN CN202110073541.7A patent/CN112865866B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2019191099A1 (en) * | 2018-03-26 | 2019-10-03 | Zte Corporation | Non-linear adaptive neural network equalizer in optical communication |
| CN109347555A (en) * | 2018-09-19 | 2019-02-15 | 北京邮电大学 | A kind of visible light communication equalization methods based on radial basis function neural network |
| CN110097103A (en) * | 2019-04-22 | 2019-08-06 | 西安电子科技大学 | Based on the semi-supervision image classification method for generating confrontation network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112865866A (en) | 2021-05-28 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112865866B (en) | Visible light PAM system nonlinear compensation method based on GSN | |
| CN109067688B (en) | Dual-drive OFDM (orthogonal frequency division multiplexing) receiving method of data model | |
| Lee et al. | Feature image-based automatic modulation classification method using CNN algorithm | |
| Chuang et al. | Employing deep neural network for high speed 4-PAM optical interconnect | |
| CN108667523B (en) | Optical fiber nonlinear equalization method based on data-aided KNN algorithm | |
| CN109905170B (en) | K-DNN-based nonlinear distortion compensation algorithm and visible light communication device | |
| CN112600618B (en) | Attention mechanism-based visible light signal equalization system and method | |
| CN107566039A (en) | A kind of VISIBLE LIGHT SYSTEM non-linear compensation method based on cluster judgement | |
| CN109818889B (en) | Equalization algorithm for SVM classifier optimization in high-order PAM optical transmission system | |
| CN111313971B (en) | Lightgbm equalization system and method for improving IMDD short-distance optical communication system | |
| CN111988249B (en) | Receiving end equalization method based on adaptive neural network and receiving end | |
| Ney et al. | Unsupervised ANN-based equalizer and its trainable FPGA implementation | |
| He et al. | Design and implementation of adaptive filtering algorithm for vlc based on convolutional neural network | |
| CN114978313B (en) | Compensation method of visible light CAP system based on Bayesian neurons | |
| CN111327558B (en) | Method and system for GMM non-uniform quantization for filter multi-carrier modulation optical communication | |
| Mathews et al. | A Non-Linear Improved CNN Equalizer with Batch Gradient Decent in 5G Wireless Optical Communication | |
| Sahu et al. | Neural network training using FFA and its variants for channel equalization | |
| Kumar et al. | Channel Estimation and BER Reduction Using Artificial Neural Network | |
| Solazzi et al. | Complex discriminative learning Bayesian neural equalizer | |
| CN112737694B (en) | Non-uniform quantization system of filter multi-carrier modulation optical communication system based on SOM | |
| CN114204993B (en) | Nonlinear equalization method and system based on polynomial mapping feature construction | |
| CN115733548A (en) | Nonlinear damage compensation system and method based on neural network equalizer | |
| CN115276818B (en) | Deep learning-based optical-load wireless transmission link demodulation method | |
| CN114500189B (en) | Direct pre-equalization method, system, device and medium for visible light communication | |
| Elbibas et al. | Neuro-Fuzzy Network for Equalization of Different Channel Models |
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 |