A nonlinear compensation system for optical communication based on photon convolution processor

Abstract

The present invention belongs to the field of optical computing and optical communication technology, and is specifically an optical communication nonlinear compensation system based on a photon convolution processor. The system is used to perform nonlinear post-equalization at the waveform level on the optical communication signal of an optical wireless communication system based on a light emitting diode (LED). The photon convolution processor includes a tunable light source unit, a weight imprinting unit, a signal modulation unit, a dispersion delay unit, a summing unit, and a data acquisition unit. The convolution layer that bears the heaviest computing task in the convolutional neural network is implemented in a photon manner; the system of the present invention can realize optical computing with large bandwidth and low latency. The present invention uses optical devices to construct a large-scale photon convolutional neural network, which can compensate for nonlinear distortion in high-speed optical communication systems, reduce system energy consumption, and ensure system performance.

Description

Technical Field

[0001] The present invention relates to the fields of optical computing and optical communication technology, and in particular to an optical communication nonlinear compensation system based on a photon convolution processor. Background Art

[0002] In an era of explosive growth in information, communication networks are facing extremely high capacity demands. Optical communication technology (including fiber-optic communication and wireless optical communication) can provide ultra-high transmission rates and is expected to become the ultimate solution for future networks and data centers. However, nonlinear distortion in optical communication systems can seriously damage the performance of the system and increase the bit error rate, especially in the case of high-speed signal transmission. In recent years, with the rise of machine learning and deep learning, neural networks have been proven to be able to successfully recover signals with linear and high-order nonlinear distortion.

[0003] In addition, as computing demands grow, the development of efficient neuromorphic computing architectures has received widespread attention. Optical neural networks (ONNs) are a promising approach given their ultrafast computing speed, energy efficiency, and parallel computing capabilities. Various optical network architectures based on wavelength division multiplexing and spatial multiplexing have been proposed. In the field of optical communications, the use of optical neural networks for signal processing has also received increasing attention. The optical implementation of advanced neural networks in optical communication signal processing provides a promising alternative to traditional DSP circuits, which can speed up processing speed and efficiency. Summary of the invention

[0004] In view of the deficiencies of the above-mentioned prior art, the purpose of the present invention is to provide an optical communication nonlinear compensation system based on a photon convolution processor; the present invention adopts a photon convolution neural network (PCNN) and a minimum mean square error (LMS) equalizer to perform linear and nonlinear equalization on the optical communication CAP (carrier-free amplitude and phase modulation)-QAM (quadrature amplitude modulation) signal. The photon convolution processor of the present invention can realize large-bandwidth, low-latency optical computing. The present invention uses optical devices to construct a large-scale photon convolution neural network, which can realize nonlinear distortion compensation in high-speed optical communication systems, while reducing system energy consumption and ensuring system performance.

[0005] The technical solution of the present invention is specifically described as follows.

[0006] The present invention provides an optical communication nonlinear compensation system based on a photon convolution processor, which can be applied to a visible light CAP communication system. The system is composed of a tunable light source unit, a weight imprinting unit, a signal modulation unit, a dispersion delay unit, a summing unit and a data acquisition unit, which are sequentially connected to form a photon convolution processor for executing convolution layer operations in a convolutional neural network (CNN); wherein: the tunable light source unit is used to provide a multi-channel laser with the same wavelength spacing that carries signals and neural network weights;

[0007] The weight imprinting unit is used to imprint the CNN weights w=(w1,…,w i ,…,w n ) is converted from electrical information into laser intensity information, that is, different powers are assigned to the lasers generated by the tunable light source unit. The n-channel laser power corresponds to the weight w of a convolution kernel of CNN = (w1,…,w i ,…,w n ) It uses a wavelength selective switch WSS to control the power of the laser by controlling the attenuation coefficient at different frequencies;

[0008] The signal modulation unit is used to convert the optical communication signal x of the optical wireless communication system into j ,…,x m ) is loaded onto each laser generated by the tunable light source unit, and the modulated output is the product of the laser power w and x:

[0009]

[0010] The dispersion delay unit is used to introduce an arithmetic delay to each laser carrying the optical communication signal. The delay time is equal to the length of one symbol of the optical communication signal. Its laser output is expressed as:

[0011]

[0012] The summing unit is used to sum the laser signals with equal delays in time and convert the optical signals into electrical signals. Multiple wavelengths are superimposed on the summing unit. These wavelengths carry signals x with different delays and different intensities. The incoherent intensity superposition is X FOUT Adding them column by column is equivalent to the convolution result of w and x. Finally, the convolution output signal is trimmed to the same length as the input signal.

[0013] A data acquisition unit, used for converting the analog signal output by the summing unit into a digital signal;

[0014] During operation, after multiple lasers are generated from the tunable light source, they obtain different powers in the weight imprinting unit. The optical signal that needs to be compensated is loaded on each laser through the signal modulation unit, and the signal delay alignment is achieved through the dispersion delay unit; finally, the summing unit performs the summing and collects the digital signal through the data acquisition unit to complete a convolution operation; the weights of the convolutional neural network are trained using the original data of the optical communication transmitter and the distorted data of the receiver, and the weights are loaded into the weight imprinting unit, and the convolution calculation and electrical nonlinear activation function are performed multiple times to realize the complete CNN, that is, to achieve nonlinear compensation.

[0015] In the present invention, the photon convolution processor has different responsiveness to lasers of different frequencies. Since the laser power detected at the summing unit represents the convolution kernel weight, before using the weight imprinting unit, the transfer function of each laser channel, that is, the precise correspondence between the CNN electrical weight and the optical weight, must be measured and saved separately; in order to more conveniently adjust the laser power, a waveform selection switch WSS is used to set different attenuation coefficients; the specific steps are:

[0016] Step 1: Load a randomly generated waveform into the signal modulation unit and turn on the tunable light source with a frequency of f i The laser on the other roads is turned off;

[0017] Step 2: Control WSS at the laser f i and the attenuation coefficient α within the bandwidth including the left and right Δf;

[0018] Step 3: Traverse α from 0 with a certain step length; each time α is set, a piece of data y is sampled at the receiving end α ;

[0019] Step 4: Calculate the average absolute value of this sampling data

[0020] Step 5: Turn off the current laser of the tunable laser, turn on the next laser, and repeat the above steps until each laser obtains the corresponding relationship between the attenuation coefficient of WSS and the received power, and uses Fit the transmission data of each channel, where x represents the receiving strength and y represents the attenuation coefficient.

[0021] Step 6: In the stage of converting electrical weights into optical weights, the weight w i Substituting the value into the transfer function, the attenuation coefficient of WSS is obtained, and the light weight can be changed in real time.

[0022] In the present invention, the dispersion delay unit adopts a standard single-mode optical fiber SSMF; the laser wavelength interval Δλ of the tunable laser unit, the input signal baud rate B of the signal modulation unit, the length L of the dispersion delay unit and the dispersion coefficient D satisfy the following relationship:

[0023] τ=1 / B=D·L·Δλ.

[0024] In the present invention, CNN is used to perform nonlinear post-equalization on waveform-level signals. CNN includes multiple convolutional layers with ReLU activation functions and a fully connected layer. In order to reduce information loss, no pooling layer is used. The training data of CNN is the receiving end data of visible light communication, and the label set is the corresponding transmitting end data. The training process is completed on an electronic computer. Through training, the complex mapping function h between the receiving data and the transmitting data is found. -1 (·):

[0025]

[0026] The training process of convolutional neural network is expressed as the following optimization process:

[0027]

[0028] After training, the weights of the convolutional neural network are loaded into the weight imprinting unit, converted into optical weights, and the operations of all convolution parts of the CNN are performed on the photonic convolution processor.

[0029] The present invention also includes an LMS equalizer, which is used to solve the partial linear distortion of the symbol level after the signal is downsampled; the received serial signal is divided into N blocks of size in a sliding window manner. LMS subsets of them are sent in parallel to the LMS equalizer, in parallel signals and tap vectors Perform a product between:

[0030]

[0031] In order to optimize the equalization performance, by calculating y LMS (n) and the distance between the transmitted signal of the optical communication system to update the tap vector in the gradient descent method and feed back the error.

[0032] The present invention also provides an optical wireless communication system based on a light emitting diode (LED), which utilizes the above optical communication nonlinear compensation system to perform nonlinear post-equalization on an optical communication signal, and comprises a transmitting end and a receiving end; the transmitting end comprises an arbitrary waveform generator, a hardware pre-equalizer, a circuit amplifier, a bias three-way combination, an LED transmitting chip and a transmitting collimating convex lens; the receiving end comprises a receiving focusing convex lens, a photodetector and a digital storage oscilloscope; wherein:

[0033] Arbitrary waveform generator: used to convert the input pre-processed digital signal into an electrical signal;

[0034] Hardware pre-equalizer: used to input electrical signals and compensate for high-frequency attenuation introduced by LED devices;

[0035] Circuit amplifier: amplifies input electrical signals;

[0036] Bias tee combination: used to add a DC bias to the electrical signal and apply it to the LED;

[0037] LED transmitting chip: converts electrical signals into optical signals and transmits them into free space;

[0038] Emission collimating convex lens: used to collimate the light emitted by the LED;

[0039] Receiving focusing convex lens: used to focus the optical signal in free space onto the photodetector;

[0040] Photodetector: converts the received light signal into an electrical signal;

[0041] Digital storage oscilloscope: used to sample the received electrical signal for post-processing of the signal.

[0042] In the present invention, the preprocessed digital signal input into the arbitrary waveform generator is obtained by the following method:

[0043] The original bit sequence is sequentially subjected to orthogonal amplitude modulation (QAM) mapping and up-sampling. The I and Q paths of the QAM complex signal are filtered using a pair of shaping filters that satisfy the Hilbert transform pair characteristics. The filtered orthogonal signals are subtracted to obtain the CAP signal.

[0044] In the present invention, the digital storage oscilloscope samples the received electrical signal to perform nonlinear and linear equalization post-processing on the signal; the post-processing method includes:

[0045] First, the signals are synchronized, the power is normalized, and nonlinear post-equalization is performed at the waveform level using a PCNN based on a photonic convolution processor;

[0046] Then, the equalized signal is processed by a pair of orthogonal matched filters to obtain the I and Q signals of QAM respectively;

[0047] Then downsample the data and use the LMS equalizer to perform LMS adaptive post-equalization at the symbol level;

[0048] Finally, the QAM mapping is deconstructed to obtain the final bit sequence.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] The present invention discloses an optical communication nonlinear compensation system based on a photon convolution processor, which can solve the problem of linear and nonlinear distortion in the optical communication system and improve the transmission rate of the system. The photon convolution processor in the present invention adopts a time-space multiplexing architecture. The weight of the convolution calculation can be changed by adjusting the wavelength selection switch, and the scale of the convolution operation can be increased by increasing the number of wavelengths. Therefore, it is a reconfigurable and scalable architecture. Each unit of the convolution optoelectronic processor is an existing telecommunication component, which can realize the large-scale deployment of optical communication networks, thereby reducing the energy consumption of optical networks and improving efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 .Implementation block diagram of the photon convolution processor described in the embodiment.

[0052] Figure 2 .The optical communication transmission system described in the embodiment, and its digital signal processing block diagram.

[0053] Figure 3 In the optical communication system: (a) the constellation diagram after using only LMS equalization, (b) the constellation diagram after using PCNN+LMS equalization. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical scheme and advantages of the present invention clearer, the following will be described in detail with reference to the accompanying drawings. However, it will be appreciated by those skilled in the art that in the various embodiments of the present invention, many technical details are provided in order to enable the reader to better understand the present application. However, even without these technical details and various changes and modifications based on the following embodiments, the technical scheme claimed for protection in the claims of the present application can be implemented.

[0055] The purpose of the present invention is to provide an optical communication nonlinear compensation system based on a photon convolution processor, which can realize optical computing with large bandwidth and low latency. The present invention uses optical devices to construct a large-scale photon convolution neural network, which can solve the nonlinear distortion problem in high-speed optical communication systems and reduce system energy consumption while ensuring system performance.

[0056] Figure 1 A system block diagram of a time-wavelength multiplexed convolution optoelectronic processor is shown.

[0057] The photonic convolutional neural network PCNN is based on a photonic convolution processor, such as Figure 1As shown in the figure, it consists of a tunable light source (TLS), a wavelength selective switch (WSS), a Mach-Zehnder modulator (MZM), an erbium-doped fiber amplifier (EDFA), a photodetector (PD), a standard single-mode fiber (SSMF), and a digital storage oscilloscope (OSC); after multiple lasers are generated from the tunable light source (TLS), the WSS is used as a weight imprinting unit to obtain different powers, and the optical signal to be compensated is loaded on each laser through the MZM, i.e., the signal modulation unit, and the signal delay alignment is achieved through the SSMF as a dispersion delay unit; finally, the PD is used as a summing unit to sum and the digital signal is collected through the data acquisition unit OSC to complete a convolution operation. It uses the original data of the transmitting end of the optical communication and the distorted data of the receiving end to train the weights of the convolutional neural network, and loads the weights to the weight imprinting unit, and performs convolution calculations multiple times to achieve CNN nonlinear compensation.

[0058] Among them, TLS is used to provide multiple lasers that carry signals and neural network weights for the entire system, and the frequency intervals of adjacent lasers are equal; WSS is used to give different powers to the lasers generated by the tunable light source unit, corresponding to the weights of the convolutional neural network w=(w1,…,w i ,…,w n ); MZM is used to convert the received signal x obtained from the optical communication system into (x1,…,x j ,…,x m ) is loaded onto each laser generated by the tunable light source unit, and the SSMF is used to introduce an arithmetic delay to each laser carrying an optical communication signal, and the delay time is equal to the length of one symbol of the optical communication signal;

[0059] PD is used to sum the lasers that carry weights and optical communication signals and are delayed arithmetic difference in time, and convert the optical signals into electrical signals;

[0060] The OSC is used to convert the analog signal output by the summing unit into a digital signal to facilitate subsequent processing.

[0061] The LMS equalizer uses the classic LMS algorithm for adaptive equalization;

[0062] The convolution layer in CNN undertakes the heaviest computing tasks in the network, which is implemented by the photonic convolution processor in a photon manner, while the non-activation function ReLU is executed by an electronic computer. The fully connected layer is a special convolution operation, which can also be implemented by the photonic convolution processor.

[0063] When the optical communication nonlinear compensation system of the present invention is working, the tunable laser source (TLS) generates multi-wavelength lasers with the same wavelength spacing and is coupled to the waveform selection switch (WSS). The WSS controls the optical power of n wavelengths to correspond to the trained convolution kernel w=(w1,…,wi ,…,w n ). The electrical input signal x of the convolution optoelectronic processor is x = (x1,…,x j ,…,x m ) represents the receiving signal of the optical communication system. The input signal x is generated by a high-bandwidth arbitrary waveform generator (AWG), and the laser power is modulated by the Mach-Zehnder modulator MZM driven by the signal x, and multicast to each laser channel. In this way, each wavelength of laser carries a signal x with a different weight amplitude. The output laser power after electro-optical modulation is the product of w and x:

[0064]

[0065] The multi-channel light source propagates through the SSMF of length L, resulting in different time delays due to dispersion. The time delay between adjacent wavelengths is exactly the signal x j The duration of one symbol of . Therefore, the following formula needs to be matched

[0066]

[0067] Here, τ represents the time delay between adjacent wavelengths, B represents the baud rate of signal x, D represents the dispersion index of the optical fiber, and Δλ represents the laser wavelength interval. After the signal is dispersed and delayed by SSMF, it can be expressed as

[0068]

[0069] Finally, multiple wavelengths are superimposed on the PD, which carry signals x with different delays and different intensities. The incoherent intensity superposition is X FoUT Adding them column by column is equivalent to the convolution result of w and x. Finally, we trim the convolution output signal to the same length as the input signal.

[0070] In order to prove the effect of the present invention in the communication system, the verification steps and verification effect of the PCNN combined with LMS post-equalization method proposed by the present invention in the optical wireless communication system are introduced below.

[0071] Optical wireless communication system and its digital signal processing block diagram Figure 2 shown.

[0072] The optical wireless communication system includes an arbitrary waveform generator (AWG), a hardware pre-equalizer (eq), a circuit amplifier (EA), a bias tee combination (Bias Tee), an LED transmitting chip, a transmitting collimating convex lens, a receiving focusing convex lens, a photodetector (PD), a transimpedance amplifier (TIA) and a digital storage oscilloscope (OSC).

[0073] The specific working steps of the optical wireless communication system for processing digital signals are as follows:

[0074] First, at the transmitter, during digital preprocessing, the original bit sequence undergoes quadrature amplitude modulation (QAM) mapping, upsampling, and carrier-free amplitude modulation (CAP). The normalized sequence is then loaded into an arbitrary waveform generator (AWG) for digital-to-analog conversion, and the electrical signal passes through a pre-equalization circuit and an electrical amplifier (EA), and is finally coupled with direct current (DC) to drive the LED to emit light. The light carrying the information is focused by a lens and transmitted 1.2 meters in free space.

[0075] At the receiver, a focusing lens focuses the light onto a photodetector (PD), which converts the optical signal into an electrical signal. A transimpedance amplifier (TIA) amplifies the received electrical signal, and a digital storage oscilloscope (OSC) samples the signal and sends it for processing. In the digital post-processing section, the signal undergoes first-stage equalization, matched filter, downsampling, second-stage equalization, and QAM demapping. The distortion in the system mainly comes from the nonlinear response of the LED, photodetector (PD), transmitter amplifier circuit, and receiver amplifier circuit. To solve this problem, a two-stage signal recovery algorithm is used: the CNN equalizer compensates for nonlinearity with the received waveform signal as the target, which is computationally intensive; while the LSM equalizer handles the residual linear distortion after downsampling by applying it to the QAM signal.

[0076] The transfer function of the optical wireless communication system is h(.), the transmitted data sequence is s, the received data is represented by x, and the additive noise of the signal is represented by N0. The system can be briefly represented by the following equation:

[0077] x=h(s)+N0

[0078] CNN is used to perform nonlinear post-equalization on waveform-level signals. CNN includes multiple convolutional layers with ReLU activation function and a fully connected layer. To reduce information loss, no pooling layer is used. The training data of CNN is the receiving end data of visible light communication, and the label set is the corresponding transmitting end data. The training process is completed on an electronic computer. Through training, the complex mapping function h between the receiving data and the transmitting data is found. -1 (·):

[0079]

[0080] The training process of CNN is expressed as the following optimization process:

[0081]

[0082] After training, the weights of the CNN are loaded into the weight imprinting unit, converted into optical weights, and the operations of all convolution parts of the CNN are performed on the photonic convolution processor.

[0083] Figure 3 The figure shows the effect of post-equalization of the received signal at the receiving end of the optical wireless communication system of the present invention. Figure 3 (a) is the effect of using only LMS post-equalization. Figure 3 (b) shows the effect of PCNN and LMS post-equalization. The original constellation points are 16 clustered points. Figure 3 From the results, we can see that when only LMS post-equalization is used, all constellation points are very scattered and the boundary of each constellation point is basically indistinguishable. When PCNN and LMS post-equalization are used, the noise points in the constellation points are greatly reduced, so that each constellation point can be basically separated.

[0084] Those skilled in the art will appreciate that the above-mentioned embodiments are specific examples for implementing the present invention, and in actual applications, various changes may be made thereto in form and detail without departing from the spirit and scope of the present invention.

[0085] Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An optical communication nonlinear compensation system based on a photon convolution processor, which is applied in a visible light CAP communication system, characterized in that: A photon convolution processor is formed by sequentially connecting a tunable light source unit, a weight imprinting unit, a signal modulation unit, a dispersion delay unit, a summing unit and a data acquisition unit to perform convolution layer operations in a convolutional neural network (CNN); wherein: A tunable light source unit for providing multiple lasers with the same wavelength spacing that carry signals and neural network weights; The weight imprinting unit is used to imprint the CNN weights w=(w1,…,w i ,…,w n ) is converted from electrical information into laser intensity information, that is, different powers are assigned to the lasers generated by the tunable light source unit. The n-channel laser power corresponds to the weight w of a convolution kernel of CNN = (w1,…,w i ,…,w n ) It uses a wavelength selective switch WSS to control the power of the laser by controlling the attenuation coefficient at different frequencies; The signal modulation unit is used to convert the optical communication signal x of the optical wireless communication system into j ,…,x m ) is loaded onto each laser generated by the tunable light source unit, and the modulated output is the product of the laser power w and x: The dispersion delay unit is used to introduce an arithmetic delay to each laser carrying the optical communication signal. The delay time is equal to the length of one symbol of the optical communication signal. Its laser output is expressed as: The summing unit is used to sum the laser signals with equal delays in time and convert the optical signals into electrical signals. Multiple wavelengths are superimposed on the summing unit. These wavelengths carry signals x with different delays and different intensities. The incoherent intensity superposition is X FOUT Adding them column by column is equivalent to the convolution result of w and x. Finally, the convolution output signal is trimmed to the same length as the input signal. A data acquisition unit, used for converting the analog signal output by the summing unit into a digital signal; During operation, after multiple lasers are generated from the tunable light source unit, they obtain different powers in the weight imprinting unit. The optical communication signal that needs to be compensated is loaded on each laser through the signal modulation unit, and the signal delay alignment is achieved through the dispersion delay unit; finally, the summation unit performs the summation and collects the digital signal through the data acquisition unit to complete a convolution operation; the weights of the convolutional neural network are trained using the original data from the optical communication transmitter and the distorted data from the receiver, and the weights are loaded into the weight imprinting unit, and the optical convolution calculation and electrical nonlinear activation function are performed multiple times to realize the complete CNN, that is, to achieve nonlinear compensation.

2. The optical communication nonlinear compensation system according to claim 1, characterized in that: The photon convolution processor has different responsiveness to lasers of different frequencies. Since the laser power entering the summing unit represents the convolution kernel weight, before using the weight imprinting unit, the transfer function of each laser channel, that is, the precise correspondence between the CNN electrical weight and the optical weight, must be measured and saved separately; in order to more conveniently adjust the laser power and keep the laser power of the tunable light source unit unchanged, a waveform selection switch WSS is used to set different attenuation coefficients; the specific steps are: Step 1: Load a randomly generated waveform into the signal modulation unit and turn on the tunable light source with a frequency of f i The laser on the other roads is turned off; Step 2: Control WSS at the laser f i and the attenuation coefficient α within the bandwidth including the left and right Δf; Step 3: traverse α from 0 with a certain step length; each time α is set, the summing unit at the receiving end performs intensity detection and samples a segment of data y α ; Step 4: Calculate the average absolute value of this sampling data Step 5: Turn off the current laser of the tunable laser, turn on the next laser, and repeat the above steps until each laser obtains the corresponding relationship between the attenuation coefficient of WSS and the received power, and uses Fit the transmission data of each channel, where x represents the receiving strength and y represents the attenuation coefficient; Step 6: In the stage of converting electrical weights into optical weights, the weight w i The value is substituted into the transmission function to obtain the attenuation coefficient of WSS to achieve real-time change of light weight.

3. The optical communication nonlinear compensation system according to claim 1, characterized in that: The dispersion delay unit adopts a standard single-mode optical fiber SSMF; the laser wavelength interval Δλ of the tunable laser unit, the input signal baud rate B of the signal modulation unit, the length L of the dispersion delay unit and the dispersion coefficient D satisfy the following relationship: 1 / B=D·L·Δλ.

4. The optical communication nonlinear compensation system according to claim 1, characterized in that: CNN is used to perform nonlinear post-equalization on waveform-level signals. CNN includes multiple convolutional layers with ReLU activation functions and a fully connected layer. To reduce information loss, no pooling layer is used. The training data of CNN is the receiving end data of visible light communication, and the label set is the corresponding transmitting end data. The training process is completed on an electronic computer. Through training, find the complex mapping function h between the received data x and the transmitted data s -1 (·): The training process of convolutional neural network is expressed as the following optimization process: After training, the weights of the convolutional neural network are loaded into the weight imprinting unit, converted into optical weights, and the operations of all convolution parts of the CNN are performed on the photonic convolution processor.

5. The optical communication nonlinear compensation system according to claim 1, characterized in that: It also includes an LMS equalizer, which is used to solve the partial linear distortion of the symbol level after the signal is downsampled; the received serial signal is divided into N blocks of size in a sliding window manner. LMS subsets of them are sent in parallel to the LMS equalizer, and the parallel signals x LMS and tap vector Perform a product between: In order to optimize the equalization performance, by calculating y LMS The distance between (p) and the transmitted signal of the optical communication system is used to update the tap vector in the gradient descent method and feed back the error.

6. An optical wireless communication system based on light emitting diodes (LEDs), which uses the optical communication nonlinear compensation system of claim 1 to perform nonlinear post-equalization on optical communication signals, characterized in that: It includes a transmitting end and a receiving end; the transmitting end includes an arbitrary waveform generator, a hardware pre-equalizer, a circuit amplifier, a bias three-way combination, an LED transmitting chip and a transmitting collimating convex lens; the receiving end includes a receiving focusing convex lens, a photodetector and a digital storage oscilloscope; wherein: Arbitrary waveform generator: used to convert the input pre-processed digital signal into an electrical signal; Hardware pre-equalizer: used to input electrical signals and compensate for high-frequency attenuation introduced by LED devices; Circuit amplifier: amplifies input electrical signals; Bias tee combination: used to add a DC bias to the electrical signal and apply it to the LED; LED transmitting chip: converts electrical signals into optical signals and transmits them into free space; Emission collimating convex lens: used to collimate the light emitted by the LED; Receiving focusing convex lens: used to focus the optical signal in free space onto the photodetector; Photodetector: converts the received light signal into an electrical signal; Digital storage oscilloscope: used to sample the received electrical signal for post-processing of the signal.

7. The optical wireless communication system based on light emitting diodes (LEDs) according to claim 6, characterized in that: The pre-processed digital signal input to the arbitrary waveform generator is obtained by the following method: The original bit sequence is sequentially subjected to orthogonal amplitude modulation (QAM) mapping and up-sampling. The I and Q paths of the QAM complex signal are filtered using a pair of shaping filters that satisfy the Hilbert transform pair characteristics. The filtered orthogonal signals are subtracted to obtain the CAP signal.

8. The optical wireless communication system based on light emitting diodes (LEDs) according to claim 6, characterized in that: The digital storage oscilloscope samples the received electrical signal to perform nonlinear and linear equalization post-processing on the signal; the post-processing method includes: First, the signal is synchronized, the power is normalized, and nonlinear post-equalization is performed at the waveform level using a photon convolution neural network (PCNN) based on a photon convolution processor. Then, the equalized signal is processed by a pair of orthogonal matched filters to obtain the I and Q signals of QAM respectively; Then downsample the data and use the LMS equalizer to perform LMS adaptive post-equalization at the symbol level; Finally, the QAM mapping is deconstructed to obtain the final bit sequence.

Images