Visible light communication system modeling method and system adopting neural network equation discovery
Through the neural network equation discovery method, combined with neural network and differential equations, the nonlinear characteristic problem in LED-VLC system modeling is solved, and high-precision, interpretable and applicable modeling is achieved. It is suitable for different hardware platforms and communication environments, improving the applicability and optimization capabilities of the system.
Patent Information
- Application Number
- CN202510610354.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-13
- Publication Date
- 2025-07-04
AI Technical Summary
The existing LED-VLC system modeling technology has problems such as difficult to accurately describe nonlinear characteristics, lack of interpretability of empirical models, insufficient computational complexity and generalization capabilities, and strict requirements for system assumptions by theoretical models, resulting in limited modeling accuracy and applicability.
The neural network equation discovery method is used, combined with neural networks and differential equations, and the dynamic characteristics of the system are automatically extracted from experimental data, and an accurate and concise transfer function is constructed. Through the neural network, a complex mapping between input and output signals is learned to generate interpretable differential equations.
It realizes high-precision nonlinear modeling, enhances the interpretability and generalization capabilities of the model, reduces the computational complexity, is suitable for different hardware platforms and communication environments, and improves the applicability and optimization capabilities of the system.
Smart Images
Figure CN120263324A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of visible light communication systems, and particularly to a modeling method and system for visible light communication systems discovered by using neural network equations. Background Art
[0002] Visible Light Communication (VLC) has the characteristics of strong anti-interference, good security and confidentiality, and rich spectrum resources, and is considered to be one of the key technologies for the next generation of communication. Compared with traditional radio frequency communication, VLC has advantages such as rich spectrum resources, strong electromagnetic compatibility, and high security, and is widely used in indoor wireless communication, vehicle networking, intelligent lighting and other fields. A visible light communication system that uses a light-emitting diode (LED) for data transmission is also called an LED visible light communication (LED-VLC) system. It uses an LED as a light source, transmits data by modulating the brightness change of the LED, and uses a photodetector (such as a photodiode) for signal reception.
[0003] In the current LED-VLC system, accurately identifying the transfer function of the system is crucial for system optimization and performance improvement. In communication system modeling, system identification mainly includes two types of methods: empirical models and theoretical models. Empirical models rely on experimental data and use technologies such as machine learning to construct input-output relationships. Although they can adapt to complex systems, they lack interpretability and have limited generalization ability. Theoretical models are based on the mathematical principles of the system, such as differential equations or Fourier analysis, and have strong interpretability, but it is difficult to accurately describe the nonlinear characteristics of complex systems. In the LED-VLC system, due to the nonlinear response of the LED and the channel complexity, traditional methods have limitations in modeling accuracy and applicability, and there is an urgent need for an efficient identification method that combines data-driven and theoretical modeling.
[0004] In summary, the current LED-VLC system modeling technology has the following problems:
[0005] (1) It is difficult to accurately model the nonlinear characteristics: Traditional theoretical models are mainly based on the assumption of linear systems, and it is difficult to effectively describe the nonlinear distortion and inter-symbol interference (ISI) in the LED-VLC system, resulting in limited model accuracy.
[0006] (2) The empirical model lacks interpretability: Although the data-driven empirical modeling method can improve the fitting accuracy to a certain extent, it often lacks clear meaning, making it difficult to promote and optimize the model.
[0007] (3) Trade-off between computational complexity and generalization ability: High-order theoretical models or complex machine learning models may perform well on specific datasets, but they have high computational costs and insufficient generalization ability, making it difficult to be efficiently applied in different VLC systems or hardware environments.
[0008] (4) Strict requirements of theoretical models for system assumptions: Many theoretical modeling methods rely on idealized assumptions, such as linear time-invariance (LTI) or specific channel distributions. However, actual LED-VLC systems are restricted by device characteristics and environmental factors and are difficult to fully meet these assumptions, resulting in limited applicability of the models in practical applications. Summary of the Invention
[0009] To solve the above problems, the present invention provides a method and system for modeling a visible light communication system using equation discovery based on neural network. The present invention aims to solve the limitations of existing system modeling techniques based on empirical models and theoretical models, especially the challenges faced when describing the nonlinear distortion and inter-symbol interference (ISI) of the system. By combining neural networks and differential equations, the present invention can automatically extract the dynamic characteristics of the system from experimental data and construct an accurate and concise transfer function.
[0010] In a first aspect, a method for modeling a visible light communication system using equation discovery based on neural network designed by the present invention includes:
[0011] S1. Obtain transmitted symbols, and obtain the true received symbols through the visible light communication model.
[0012] S2. Input the transmitted symbols into the neural network to obtain predicted received symbols, and calculate the data error according to the predicted received symbols and the true received symbols.
[0013] S3. Construct a candidate symbol library according to the transmitted symbols and the predicted received symbols.
[0014] S4. Construct a differential transfer function including weighting coefficients according to the predicted received symbols and the candidate symbol library, and calculate the transfer loss.
[0015] S5. Introduce a regularization constraint loss based on the differential transfer function.
[0016] S6. Calculate the total loss according to the data error, the transfer loss, and the regularization constraint loss.
[0017] S7. Repeat steps S1 - S6 to optimize the neural network parameters and the weighting coefficients until the total loss converges.
[0018] Furthermore, the visible light communication model includes a signal transmission module, an intermediate processing module, and a signal reception module. The signal transmission module includes an upsampling module, a pulse shaping module, and a carrier modulation module. The intermediate processing module includes an arbitrary waveform generator, a pre-equalizer, an LED, a visible light channel, a photodetector, and an oscilloscope. The signal reception module includes a downsampling module, a synchronization module, and a demodulation module.
[0019] Furthermore, step S2 inputs the transmitted symbol into the neural network to obtain the predicted received symbol, denoted as
[0020]
[0021] where Tx i represents the i-th transmitted symbol, represents the i-th predicted received symbol, NN() represents the neural network, and Θ represents the trainable parameters of the neural network;
[0022] The data error is denoted as
[0023]
[0024] where Loss data represents the data error, MSE() represents the mean square error, and Rx i represents the i-th received symbol.
[0025] Furthermore, step S3 constructs a candidate symbol library based on the transmitted symbol and the predicted received symbol, including:
[0026] Calculate the partial derivative according to the transmitted symbol Tx i and the predicted received symbol
[0027] Obtain the basic term set of the transmitted symbol Tx i where n represents the quantity factor;
[0028] Obtain the cross-multiplication term set of the transmitted symbol Tx i
[0029] Combine the basic term set and the cross-multiplication term set of the transmitted symbol Tx i to form the candidate symbol library of the transmitted symbol Tx i Each item in the candidate symbol library is a basis function.
[0030] Furthermore, construct a differential transfer function including weighted coefficients based on the predicted received symbol and the candidate symbol library, denoted as
[0031]
[0032] Among them, Tx i represents the i-th transmitted symbol, represents the i-th predicted received symbol, Φ(Tx i ) represents the candidate symbol library of the transmitted symbol Tx i , and Λ represents the weighting coefficient;
[0033] The transmission loss is expressed as
[0034]
[0035] Among them, Loss transfer represents the transmission loss, and MSE() represents the mean square error.
[0036] Furthermore, the regularization constraint loss is expressed as
[0037] Loss reg = ||Λ||1
[0038] Among them, Loss reg represents the regularization constraint loss, ||·||1 represents the first norm, and Λ represents the weighting coefficient.
[0039] In a second aspect, based on the method proposed in the first aspect, the present invention further provides a visible light communication system modeling system discovered by using a neural network equation, including:
[0040] A communication system data acquisition and preprocessing module, configured to obtain transmitted symbols and obtain true received symbols through a visible light communication model for the transmitted symbols;
[0041] A neural network training and candidate symbol library construction module, configured to input the transmitted symbols into a neural network to obtain predicted received symbols, and construct a candidate symbol library according to the transmitted symbols and the predicted received symbols;
[0042] A differential equation discovery and simplification module, configured to construct a differential transfer function including a weighting coefficient according to the predicted received symbols and the candidate symbol library;
[0043] In the visible light communication system modeling system, the total loss is calculated according to the data error, the transmission loss, and the regularization constraint loss, and the neural network parameters and the weighting coefficient are optimized through the total loss.
[0044] Advantages of the present invention:
[0045] High-precision non-linear modeling: The present invention combines neural network and differential equation modeling methods, can accurately describe complex factors such as non-linear distortion, inter-symbol interference (ISI), and channel noise in the LED-VLC system, improve the fitting accuracy of the model, and make it more in line with the characteristics of the actual visible light communication system.
[0046] Enhanced model interpretability: Compared with pure data-driven deep learning methods, the present invention extracts mathematical expressions through symbolic regression to generate differential equations with clear meanings, making the model no longer a "black box", facilitating users to understand system characteristics and optimizing the design of communication systems.
[0047] Improved generalization ability: The method of the present invention is not only applicable to specific VLC systems, but also can be extended to different hardware platforms and communication environments, helping to improve the applicability and scalability of the system and reducing the dependence on the scale of the dataset.
[0048] Optimized computational complexity: Compared with traditional high-order theoretical models or deep learning models, the present invention uses differential equations to describe system characteristics, reduces the consumption of computing resources, achieves a balance between computational complexity and modeling accuracy, and makes it applicable to resource-constrained devices or embedded systems.
[0049] Enhanced communication system optimization ability: The present invention can not only be used for system identification, but also guide the design and optimization of LED-VLC systems, such as improving modulation methods and optimizing signal equalization strategies, providing new modeling methods and theoretical support for efficient and low-power optical wireless communication. Description of the Drawings
[0050] Figure 1 It is a flowchart of the method of the present invention;
[0051] Figure 2 It is a structural diagram of the visible light communication model of the present invention;
[0052] Figure 3 It is a diagram of the construction process of the candidate symbol library of the present invention;
[0053] Figure 4 It is a flowchart of the algorithm of the embodiment of the present invention;
[0054] Figure 5 It is a structural diagram of the system of the present invention;
[0055] Figure 6 It is a diagram of the simulation results of the embodiment of the present invention;
[0056] Figure 7 It is a schematic diagram of the weighted coefficient structure of the present invention;
[0057] Figure 8 It is a heat map of the weighted coefficients of different systems in the embodiment of the present invention. Detailed Implementation Manner
[0058] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.
[0059] The present invention designs a modeling method for a visible light communication system discovered by using a neural network equation. As Figure 1 shown, it includes the following steps:
[0060] S1. Obtain the transmitted symbol, and obtain the true received symbol through the visible light communication model.
[0061] Specifically, as Figure 2 shown, the visible light communication model includes a signal transmission module, an intermediate processing module, and a signal reception module. The transmitted signal first passes through the signal transmission module, and after upsampling, pulse shaping, and carrier modulation, a modulated signal is obtained; then the modulated signal is input into the intermediate processing module. In the intermediate processing module, it first passes through an arbitrary waveform generator, and drives the LED transmitter to change the brightness by direct detection / intensity modulation to transmit the modulated signal. After the modulated signal (at this time, it is an optical signal) passes through the visible light channel, it is detected by a photodetector and converted into an electrical signal, and then captured by an oscilloscope; the captured electrical signal is sent into the signal reception module, and after downsampling, synchronization, and demodulation, a received signal is obtained.
[0062] S2. Input the transmitted symbol into the neural network to obtain the predicted received symbol, and calculate the data error according to the predicted received symbol and the true received symbol.
[0063] Specifically, for a visible light communication system, during the process of the transmitter sending a signal to the receiver receiving the signal, the devices, digital-to-analog conversion process, analog-to-digital conversion process, and free space visible light channel therein will have linear and non-linear effects on the signal. Assume that there is the following relationship between the transmitted symbol at the transmitter and the received symbol at the receiver:
[0064] Rx i =KH(Tx i )
[0065] where, Rx i represents the i-th received symbol, and Tx i represents the transmitted symbol corresponding to the i-th received symbol. H(Tx i ) is a candidate library of transfer function terms including linear terms and non-linear terms related to Tx i . K is the coefficient of each term in the transfer function.
[0066] Based on the transmitted and received data of the visible light communication system, this invention utilizes the powerful fitting ability of neural networks and the automatic differentiation ability of deep learning frameworks to construct H(Tx i ) and find appropriate coefficients for H(Tx i ) to find the end-to-end transfer function of the visible light communication system.
[0067] Specifically, this invention uses the neural network NN() to construct the continuous representation of the transmitted symbol and the received symbol to overcome the discontinuity problem of the transmitted and received symbols as digital signals:
[0068]
[0069] where Θ represents the trainable parameters of the neural network; the goal of this neural network is to learn the implicit mapping relationship from the transmitted symbol to the received symbol so that it can accurately predict the change trend of the received symbol. To ensure that the output of the neural network is consistent with the real received symbol data, the data loss function Loss data is defined as the error metric:
[0070]
[0071] where MSE() represents the mean square error. This loss term is used to constrain the fitting ability of the neural network so that its output is as close as possible to the real change of the received signal.
[0072] S3. Construct a candidate symbol library based on the transmitted symbol and the predicted received symbol.
[0073] Specifically, according to the above operations, the neural network can obtain the corresponding predicted received symbol for each input transmitted symbol Tx i With the help of the deep learning framework, it is also possible to obtain the partial derivative of with respect to Tx i . As shown in Figure 3 , this invention uses the linear term, non-linear term of Tx i , and the cross-multiplication terms of these terms as basis functions to construct the candidate symbol library Φ(Tx i ) for modeling the channel transmission characteristics.
[0074] Specifically, step S3 of constructing the candidate symbol library based on the transmitted symbol and the predicted received symbol includes:
[0075] Calculate the partial derivative i based on the transmitted symbol Tx and the predicted received symbol
[0076] Obtain the transmitted symbol Tx iBase item set where n represents the quantity factor, and n > 1;
[0077] Obtain the transmitted symbol Tx i Cross - multiplication item set
[0078] The transmitted symbol Tx i The base item set and the cross - multiplication item set of the transmitted symbol Tx are combined to form the transmitted symbol Tx i The candidate symbol library of the transmitted symbol Tx, and each item in the candidate symbol library is a basis function.
[0079] S4. Construct a differential transfer function including weighted coefficients based on the predicted received symbol and the candidate symbol library, and calculate the transfer loss.
[0080] Specifically, through the weighted coefficient Λ and the candidate symbol library Φ(Tx i ), a weighted combination is performed, so as to describe the transfer function using an implicit differential equation, expressed as
[0081]
[0082] where Λ represents the optimal weighted coefficient to be solved, which is used to screen basis functions from the candidate symbol library to construct the optimal transfer equation;
[0083] During the training process, the transfer loss is defined as
[0084]
[0085] where Loss transfer represents the transfer loss, and MSE() represents the mean square error. The transfer loss function is used to constrain the output of the neural network so that its output can meet the characteristics of the transfer function, and construct and optimize the transfer equation of the channel. During the neural network training process, by minimizing this loss, the residual between Rx i and the candidate symbol library Φ(Tx i ) is minimized, so as to automatically discover the implicit differential equation in the transmitted and received data.
[0086] S5. Introduce a regularization constraint loss based on the differential transfer function.
[0087] Specifically, in order to further optimize the discovered differential equation to make its form more compact and interpretable, L1 - regularization is introduced as a regularization constraint to simplify the sparse constraint on Λ:
[0088] The regularization constraint loss is expressed as
[0089] Loss reg = ||Λ||1
[0090] Among them, Loss reg represents the regularization constraint loss, and ||·||1 represents the first norm. This loss can promote the form of Λ to be sparse, so as to screen out the most important basis functions for the channel transmission characteristics from the candidate symbol library Φ(Tx i ).
[0091] S6. Calculate the total loss according to the data error, transmission loss, and regularization constraint loss.
[0092] Specifically, with the help of modern deep learning frameworks and automatic differentiation, the optimal weighted coefficient vector Λ and the weight parameters Θ of the neural network can be jointly optimized and solved:
[0093] Loss total = Loss data + Loss transferfunction + Loss reg
[0094] S7. Repeat steps S1 - S6 to optimize the neural network parameters and weighted coefficients until the total loss converges.
[0095] Specifically, during the training process, after Loss total converges, through the weighted combination of the weighted coefficient vector Λ and the candidate symbol library Φ(Tx i ), a sparse and concise differential equation form can be obtained to describe the transfer function of the system:
[0096] Rx i = -Φ(Tx i )Λ
[0097] This differential equation can not only accurately describe the dynamic characteristics of the system, but also serve as an analytical expression of the transmission channel, providing theoretical support for subsequent signal equalization, bit error rate optimization, and system performance evaluation. In the present invention, a neural network is used as a tool for constructing a candidate symbol library. After iteration, the result of combining the candidate symbol library with the weighted coefficients is used as the model constructed for the system. During implementation, first, a data-driven method is used to determine the optimal coefficient vector and neural network weights, and then they are transformed into an analytical form of the transfer equation to ensure its interpretability. As Figure 4 shown, this method can be used for system modeling, signal prediction, and error compensation. Further combined with numerical simulation and experimental verification, its applicability and generalization ability are optimized. Based on this, the proposed method can be widely applied to visible light communication (VLC), wireless transmission systems, and other complex channel environments, providing reliable support for the design and optimization of intelligent communication systems. In addition, by integrating an online learning mechanism, this method can adapt to dynamically changing channel conditions, achieve real-time adjustment and optimization, thereby improving the overall communication performance and system robustness.
[0098] Preferably, the present invention also provides a visible light communication system modeling system discovered by using a neural network equation, as Figure 5 shown, including:
[0099] A communication system data acquisition and preprocessing module, which is used to obtain the input and output data of the VLC system and perform standardization processing on the data to ensure the accuracy and stability of subsequent modeling. This module first collects signals from the VLC system, including the transmitted signal and its corresponding actual received signal.
[0100] A neural network training and candidate symbol library construction module, which is used to learn the dynamic characteristics of the VLC system and construct a candidate symbol library of differential equations based on the neural network learning results.
[0101] A differential equation discovery and simplification module, which is used to automatically generate the differential equation of the VLC system according to the candidate symbol library and optimize it to ensure that the model has both high precision and controllable computational complexity.
[0102] The present invention realizes high-precision modeling of the VLC system by innovatively adopting a method combining neural network and differential equation. Specifically:
[0103] The present invention breaks through the limitation that traditional theoretical models are difficult to accurately describe the non-linear characteristics of the VLC system under linear assumptions. The present invention uses a neural network to learn the complex mapping between input and output signals, and can characterize the non-linear dynamic characteristics of the VLC system without linear assumptions. In addition, combining symbolic regression, equation discovery and neural network methods, a simple and interpretable mathematical expression is extracted from the high-dimensional features learned by the neural network to construct a high-precision differential equation to achieve accurate modeling of the VLC system.
[0104] In an embodiment, the simulation results of the method (ATF-NN) proposed by the present invention are as Figure 6 shown, where the black line represents the transmitted signal, the red line represents the actual received signal corresponding to the transmitted signal, and the blue color is the transfer function obtained by the present invention.
[0105] Particularly, the weighted coefficient structure in the transfer function of the present invention is as Figure 7 shown, where purple and green are linear regions, and orange and blue are non-linear regions. The method of the present invention is used for modeling on two different systems (one with weaker non-linear effects and almost only linear effects; the other with stronger non-linear effects), and the weighted coefficients are plotted as a heat map. As Figure 8 shown, the former has larger weights only in the linear region, while the latter has larger weights in both the linear region and the non-linear region.
[0106] Experimental verification shows that this method can effectively capture the system's non - linear distortion and inter - symbol interference, significantly improving the modeling accuracy of the VLC system.
[0107] The present invention overcomes the disadvantage of poor interpretability of data - driven empirical models, and has both high fitting accuracy and good generalizability. Although traditional deep - learning models have excellent fitting capabilities, due to their "black - box" characteristics, it is difficult to generalize them to different scenarios. The present invention converts the learning results of neural networks into analytic differential equations, which not only retains the high - dimensional feature - learning advantages of neural networks but also enhances the interpretability of the model, making the dynamic characteristics of the VLC system intuitively understandable. In addition, this method can adapt to different VLC systems. By retraining the neural network and performing symbolic regression under specific experimental conditions, an applicable mathematical model can be generated to improve the generalization ability. At the same time, an adjustable modeling framework is provided, enabling users to optimize the balance between modeling accuracy and computational efficiency according to their needs.
[0108] The present invention effectively solves the trade - off between computational complexity and generalization ability and can be efficiently applied in different VLC systems and hardware environments. By combining neural networks and symbolic regression, a compact and efficient differential - equation model is generated, avoiding the need for a large amount of computational resources required by traditional deep - learning models and enabling real - time solution on an embedded platform without the need for a high - performance GPU. In addition, the method does not rely on strict system assumptions and automatically learns the system's dynamic characteristics from experimental data, ensuring strong applicability to different VLC systems, especially being able to effectively model in complex environments. By adaptively selecting the optimal symbolic - regression expression, the computational efficiency is improved, ensuring that the system - modeling process is accurate and efficient. At the same time, the computational - resource allocation is optimized, enhancing the real - time modeling ability.
[0109] The present invention overcomes the problem of strict requirements for system assumptions in theoretical models, making the constructed differential - equation model more applicable in practical applications. Traditional theoretical - modeling methods usually rely on specific mathematical assumptions, such as LTI - system assumptions or specific channel - distribution assumptions. However, actual VLC systems are limited by factors such as device non - linearity and ambient - light interference and are difficult to fully meet these assumptions, resulting in limited applicability of theoretical models. The present invention automatically mines the dynamic characteristics of the system based on experimental data and the system response learned by neural networks and generates a differential - equation model. Therefore, there is no need to pre - assume specific system characteristics, ensuring that the model can be applied to different types of VLC systems. In addition, since the method of the present invention can automatically adapt to different channel conditions, it can effectively model even in complex indoor VLC environments and provide theoretical support for subsequent signal equalization and system optimization.
[0110] In the present invention, unless otherwise clearly specified and defined, terms such as "installation", "setting", "connection", "fixation", "rotation" and the like shall be understood in a broad sense. For example, it may be a fixed connection, a detachable connection, or integrated; it may be a mechanical connection or an electrical connection; it may be directly connected or indirectly connected through an intermediate medium, and it may be the communication inside two components or the interaction relationship between two components. Unless otherwise clearly defined, for those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.
[0111] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A modeling method for visible light communication systems discovered using neural network equations, characterized in that, It includes the following steps: S1. Obtain the transmitted symbol, and obtain the true received symbol through the visible light communication model; S2. Input the transmitted symbol into the neural network to obtain the predicted received symbol, and calculate the data error according to the predicted received symbol and the true received symbol; S3. Construct a candidate symbol library according to the transmitted symbol and the predicted received symbol; S4. Construct a differential transfer function including the weighting coefficient according to the predicted received symbol and the candidate symbol library, and calculate the transfer loss; S5. Introduce the regularization constraint loss based on the differential transfer function; S6. Calculate the total loss according to the data error, the transfer loss and the regularization constraint loss; S7. Repeat steps S1 - S6 to optimize the neural network parameters and the weighting coefficient until the total loss converges.
2. A method for modeling a visible light communication system discovered by using a neural network equation according to claim 1, characterized in that The visible light communication model includes a signal transmission module, an intermediate processing module and a signal reception module. The signal transmission module includes an upsampling module, a pulse shaping module and a carrier modulation module; the intermediate processing module includes an arbitrary waveform generator, a pre - equalizer, an LED, a visible light channel, a photodetector and an oscilloscope; the signal reception module includes a downsampling module, a synchronization module and a demodulation module.
3. A method for modeling a visible light communication system discovered by using a neural network equation according to claim 1, characterized in that Step S2 inputs the transmitted symbol into the neural network to obtain the predicted received symbol, which is expressed as where Tx i represents the i-th transmitted symbol, represents the i-th predicted received symbol, NN() represents a neural network, and Θ represents the trainable parameters of the neural network; The data error is expressed as Among them, Loss data represents the data error, MSE() represents the mean square error, and Rx i represents the i-th received symbol.
4. A method for modeling a visible light communication system discovered by using a neural network equation according to claim 1, characterized in that Step S3 constructs a candidate symbol library according to the transmitted symbol and the predicted received symbol, including: Based on the transmitted symbol Tx i and the predicted received symbol calculate the partial derivative Obtain the transmitted symbol Tx i of the basic item set where n represents the quantity factor; Obtain the transmitted symbol Tx i Cross - multiplication term set The transmitted symbol Tx i is formed by combining the base item set and the cross-multiplied item set of i to form a candidate symbol library of the transmitted symbol Tx. Each item in the candidate symbol library is a basis function.
5. A method for modeling a visible light communication system discovered using a neural network equation according to claim 1, characterized in that, Construct a differential transfer function including the weighting coefficient according to the predicted received symbol and the candidate symbol library, which is expressed as Among them, Tx i represents the i-th transmitted symbol, represents the i-th predicted received symbol, Φ(Tx i ) represents the candidate symbol library of the transmitted symbol Tx i , and Λ represents the weighting coefficient; The transfer loss is expressed as Among them, Loss transfer represents the transmission loss, and MSE() represents the mean squared error.
6. A method for modeling a visible light communication system discovered by using a neural network equation according to claim 1, characterized in that, The regularization constraint loss is expressed as Loss reg = ||Λ||1 Among them, Loss reg represents the regularization constraint loss, ||·||1 represents the first norm, and Λ represents the weighting coefficient.
7. A visible light communication system modeling system discovered using neural network equations, characterized in that, Adopt a visible light communication system modeling method discovered by using a neural network equation as described in any one of claims 1 - 6. The visible light communication system modeling system includes: A communication system data acquisition and pre - processing module, which is used to obtain the transmitted symbol and obtain the true received symbol through the visible light communication model; A neural network training and candidate symbol library construction module, which is used to input the transmitted symbol into the neural network to obtain the predicted received symbol and construct a candidate symbol library according to the transmitted symbol and the predicted received symbol; A differential equation discovery and simplification module, which is used to construct a differential transfer function including the weighting coefficient according to the predicted received symbol and the candidate symbol library; In the visible light communication system modeling system, calculate the total loss according to the data error, the transfer loss and the regularization constraint loss, and optimize the neural network parameters and the weighting coefficient through the total loss.