Channel construction method for passive optical network communication system

Abstract

The invention discloses a channel construction method for a passive optical network communication system, and belongs to the field of optical fiber communication. The implementation method of the invention is as follows: based on an M-QAM signal sequence after synchronization processing, a current M-QAM signal is regarded as real data. And the condition vector and the real data are combined to construct a training data set, and a TFPNet network model for channel construction of the passive optical network system is constructed. The TFPNet network model takes the combination of real data and conditional vectors as an input feature sequence for training, signal features are fully processed, when signal data at the current moment are processed, preorder signal data information in training data can be combined and utilized, serialized feature fusion can be performed on the signal data sequence, and the signal data sequence is subjected to feature fusion. And the nonlinear interference relationship between the current signal and the preorder signal can be better represented. According to the method, the calculation complexity is lower, data symbols transmitted in a mode division multiplexing system can be efficiently recovered, and linear damage and non-linear damage in a passive optical network optical communication system are compensated.

Description

Technical Field

[0001] The invention relates to a channel construction method of a passive optical network communication system, and belongs to the field of optical fiber communication. Background Art

[0002] With the rise of big data, cloud computing, and the era of the metaverse, the demand for capacity and bandwidth in fiber-optic communications is rapidly increasing. Some important physical properties of light waves, such as amplitude, phase, polarization, and frequency, have been fully developed and applied. Passive optical networks are an efficient fiber-optic communication technology that supports multiple transmission modes through schemes such as multimode fiber (MMF) or few-mode fiber (FMF), thereby significantly improving the transmission capacity of fiber links. However, in actual communication channels, with the increase in multiplexing modes and transmission distances, as well as the nonlinear effects of optoelectronic devices inherent in the system, serious nonlinear damage is inevitable. In passive optical network fiber-optic communication systems, multiple optoelectronic devices are required to modulate signals. The nonlinear effects of these devices are their inherent characteristics, but these characteristics can cause serious damage to the signal. At present, accurate modeling of the impact of the nonlinear effects of these devices on the signal is still limited, so it is difficult to determine an accurate theoretical model. Moreover, due to the mutual influence of the nonlinear effects of multiple devices, the nonlinear model of the signal becomes more complicated. There is currently no accurate theoretical model to describe the passive optical network system channel. Therefore, deep learning has become a powerful tool to solve the problem of unknown channel modeling. In this work, a new data-driven approach is proposed to estimate the distribution of accurate channel transfer functions using a time-frequency joint prediction network (TFPNet). This approach lays the foundation for building digital networks for passive optical network systems.

[0003] At present, most of the channel construction methods based on deep learning are to model the channels of existing theoretical models. Such channels have a clear theoretical basis and neural networks can effectively model them. However, so far, there has not been a modeling method specifically used to deal with channels such as actual passive optical network systems that do not have a clear theoretical model. In passive optical network transmission, due to the influence of the nonlinear effects of various devices, the nonlinear model of the signal becomes extremely complex. Traditional channel estimation methods based on generative adversarial neural networks (GANs) cannot accurately estimate the channels of passive optical network systems because the discriminator loss function is difficult to converge accurately, leading to problems such as system instability and mode collapse. Summary of the invention

[0004] In order to cope with the challenge of lack of accurate channel model for signals in optical fiber transmission in passive optical network communication systems, the purpose of the present invention is to provide a channel construction method for passive optical network communication systems, generate conditional vectors according to the transmitted M-QAM signal sequence, generate real data according to the received M-QAM signal sequence, and construct a training data set; adopt a time domain and frequency domain joint prediction method to construct the channel of the passive optical network communication system, use the gradient back propagation algorithm and the Adam parameter optimization algorithm to update the TFPNet network model parameters, fully characterize the optical fiber nonlinear effect on the signal during the transmission process of the passive optical network system, and improve the accuracy and stability of the channel construction of the communication system.

[0005] The objective of the present invention is achieved through the following technical solutions:

[0006] The present invention discloses a channel construction method for a passive optical network communication system. At the transmitting end of the passive optical network communication system, a transmitted data sequence is mapped into a constellation symbol. Through upsampling and matched filtering processing, the current M-QAM signal and the n M-QAM signals before and after it are combined into a conditional vector. At the receiving end of the passive optical network communication system, an M-QAM signal sequence after hollow-core optical fiber transmission is received. The received M-QAM signal needs to be synchronized. Based on the M-QAM signal sequence after synchronization processing, the current M-QAM signal is regarded as real data. The conditional vector and the real data are combined to construct a training data set, and a TFPNet network model for channel construction of a passive optical network system is constructed. The TFPNet network model is trained with a combination of real data and conditional vectors as an input feature sequence. The TFPNet network model is trained using the training data set. The feature sequence of the newly generated M-QAM signal is input into the trained TFPNet network model. The TFPNet network model outputs a predicted signal for each M-QAM signal. After clock recovery of the predicted signal and the signal transmitted by the corresponding passive optical network system channel, each symbol is compared one by one, and the normalized mean square error of all signals is calculated to obtain the channel construction accuracy of the TFPNet network model. Based on the channel construction results, the complex nonlinear effects in the passive optical network communication system are characterized to improve the accuracy of channel construction. The TFPNet network model adopts a joint prediction method in the time domain and frequency domain, and constructs different joint prediction networks in the time domain and frequency domain according to the characteristics of different damages to predict nonlinear damage. It also uses the gradient back propagation algorithm and the Adam parameter optimization algorithm to update the TFPNet network model parameters, reduce nonlinear damage, improve the stability of channel construction, and achieve high-precision channel construction in the passive optical network communication system.

[0007] The present invention discloses a channel construction method for a passive optical network communication system, comprising the following steps:

[0008] Step 1: At the transmitting end of the coherent optical fiber communication system, the transmitted binary data sequence is subjected to M-QAM constellation symbol mapping processing. The constellation diagram of the M-QAM modulation format signal has M standard constellation points. Therefore, the M-QAM signal is divided into M different categories. Each standard constellation point corresponds to a category and is numbered from 1 to M as a label of the category. At the receiving end of the coherent optical fiber communication system, the M-QAM signal sequence after optical fiber transmission is received. For each M-QAM signal, the real data and imaginary data of the current M-QAM signal and the n M-QAM signals before and after it are combined as the conditional vector of the M-QAM signal.

[0009] At the transmitting end of the passive optical network system, the transmitted binary data sequence is subjected to M-QAM symbol mapping, the M-QAM symbol sequence is subjected to up-sampling and pulse shaping, the current M-QAM signal is combined with the n M-QAM signals before and after it, and a conditional vector for characterizing inter-symbol interference is constructed; the electrical signal is modulated onto an optical carrier by a Mach-Zehnder modulator, and the center frequencies of the optical carrier are set to 191.6, 191.7, 191.8, 191.9, 192.0, 192.1, 192.2, 192.3, 192.4, 192.5, 192.6, 192.7, 192.8, 192.9, 193.0, and 193.1 THz, respectively, and an erbium-doped fiber amplifier is used to amplify the signal. These 16 signals are modulated into one optical signal by a wavelength division multiplexer and transmitted through a hollow core optical fiber. At the receiving end of the passive optical network system, a wavelength selective switch is used to demultiplex the received signal to obtain an M-QAM signal sequence after transmission through a specified band, and clock recovery and synchronization processing are performed on the M-QAM signal. Based on the synchronized M-QAM signal sequence, the current M-QAM signal is used as the real data of the M-QAM signal sequence, and a training data set is constructed based on the conditional vector and the real data.

[0010] At the transmitting end of the passive optical network system, the transmitted binary data sequence is subjected to M-QAM symbol mapping, and then the M-QAM symbol sequence is subjected to upsampling and pulse shaping. According to the inter-symbol crosstalk caused by dispersion, the current M-QAM signal is combined with the n M-QAM signals before and after it to construct a conditional vector [x t-n ,…,x t ,…,x t+n ], i = 1, 2, ..., N, where x t Indicates the sample currently sent, [x t-n ,…,x t-1 ] indicates the sample that has been sent before the current sample, [x t+1 ,…,xt+n ] represents the sample to be transmitted, and n represents the number of symbols before and after, which is related to the strength of inter-symbol crosstalk. The M-QAM signal sequence is modulated onto the optical carrier by a Mach-Zehnder modulator, and the signal is amplified by an erbium-doped fiber amplifier. At the receiving end of the passive optical network system, the M-QAM signal sequence after hollow-core optical fiber transmission is received, and the M-QAM signal is clock recovered and synchronized to obtain a PAM-8 signal sequence s=[s1,s2,…,s N ], i = 1, 2, ..., N, where the vector s i =[s 1 ,s 2 ,…,s 6 ] represents the i-th M-QAM signal in the M-QAM signal sequence. For the PAM-8 signal sequence s i (i=1,2,…,N), for each M-QAM signal s i , will s i As the real data y of the M-QAM signal (i) =[s i ], i = 1, 2, ..., N, N represents the size of the training data set, construct the real data feature sequence corresponding to each M-QAM signal, and construct the training data set {x (i) ,y (i)}.

[0011] Step 2: Construct a TFPNet network model for channel construction of a passive optical network system. The TFPNet network model is a time-frequency joint prediction network, which consists of three subnetworks, namely a time domain prediction network, a high-frequency prediction network, and a nonlinear prediction network. The output signals of the three subnetworks are added to obtain the output signal y of the time-frequency joint prediction network. The data in the conditional vector feature sequence of the M-QAM signal is fully characterized by serialization and fusion processing through the fully connected layer in the TFPNet network model. When processing the M-QAM signal data at the current moment, the pre-order M-QAM signal data information and the post-order M-QAM signal data information in the conditional vector are combined and utilized, that is, the sequence M-QAM signal data in the conditional vector is subjected to serialization feature fusion, so as to better characterize the nonlinear interference relationship between the current M-QAM signal and the pre-order M-QAM signal and the post-order M-QAM signal, and enhance the nonlinear construction capability of the TFPNet network model for the M-QAM signal, and output a signal sequence representing the real channel effect. Through the training method of the time-frequency joint prediction, a TFPNet network model for characterizing the channel of a real passive optical network system is obtained.

[0012] The TFPNet network model and loss function for channel construction of passive optical network system are constructed. The TFPNet network model is a time-frequency joint prediction network, which consists of three sub-networks, namely the time domain prediction network, the high frequency prediction network and the nonlinear prediction network. The training data set constructed by the conditional vector of the M-QAM signal and the real data, and the training set and the noise vector with Gaussian distribution are simultaneously input into the time domain prediction network and the high frequency prediction network for prediction. After the original signal x is input into the time domain prediction network, it passes through the fully connected layer to obtain the output signal y1, as shown in formula (1),

[0013] y1=w*x+b (1)

[0014] Where W and b represent the weight and bias in the fully connected layer, respectively. In the fully connected layer, the linear features of the original signal are captured and fitted to predict the linear damage of the original signal. The original signal x is input into the frequency domain prediction network and then passes through the time-frequency converter to convert the time domain signal x into the frequency domain signal X[k]. The process is shown in formula (2):

[0015]

[0016] Where x[n] represents the nth symbol, and then the frequency domain signal X[k] is input into the fully connected layer and activated by the activation function ReLu to obtain the frequency domain output signal Y[k], as shown in formula (3),

[0017] Y[k]=ReLu(w*X[k]+b) (3)

[0018] Where W and b represent the weight and bias in the fully connected layer respectively. The output signal y2 is obtained by inverse Fourier transform through the frequency-time converter. In the fully connected layer, the high-frequency features of the original signal are captured and fitted to predict the nonlinear damage to the original signal x caused by high-frequency fading caused by nonlinear devices. The original input signal x is subtracted from y1 and y2 to obtain x3 containing only nonlinear damage, and x3 is input into the nonlinear prediction network. After x3 is input into the time domain prediction network, it passes through the fully connected layer to obtain the output signal y3, as shown in formula (4).

[0019] y3=w*x3+b (4)

[0020] In the fully connected layer, the nonlinear characteristics of x3 are captured and fitted to predict the nonlinear damage caused to the original signal by the passive optical network system; the output signals of the three sub-networks are added to obtain the output signal y of the time-frequency joint prediction network. The data in the conditional vector feature sequence of the M-QAM signal is fully processed by the fully connected layer in the TFPNet network model. When processing the M-QAM signal data at the current moment, the preceding M-QAM signal data information and the subsequent M-QAM signal data information in the conditional vector are combined and utilized, that is, the sequence M-QAM signal data in the conditional vector is subjected to serialized feature fusion, so as to better characterize the nonlinear interference relationship between the current M-QAM signal and the preceding M-QAM signal and the subsequent M-QAM signal, and enhance the nonlinear construction capability of the TFPNet network model for the M-QAM signal, and output a signal sequence representing the real channel effect. Through the training method of the time-frequency joint prediction, a TFPNet network model for characterizing the channel of a real passive optical network system is obtained.

[0021] Step 3: For the TFPNet network model constructed in step 2 for passive optical network system channel construction, configure the parameters required for model training, set the learning rate, batch size, weight initialization method, optimization method, and number of iterations; use the training data set constructed in step 1 to train the TFPNet network model for passive optical network system channel construction, and fully characterize the real passive optical network channel response through the trained TFPNet network model.

[0022] For the TFPNet network model constructed in step 2 for passive optical network system channel construction, configure the parameters required for model training, set the learning rate, batch size, weight initialization method, optimization method, and number of iterations.

[0023] Using the training data set {x (i) ,y (i)}, train the TFPNet network model constructed in step 2, use the gradient back propagation algorithm and the Adam parameter optimization algorithm to determine the optimal model parameters, and obtain a trained TFPNet network model. The conditional vector x of the current M-QAM signal is constructed through the trained TFPNet network model (i) The corresponding real data y (i) The TFPNet network model is used to fully characterize the influence of the M-QAM signal on the hollow-core optical fiber transmission process caused by the optical fiber nonlinear effect.

[0024] Step 4: Input the feature sequence of the newly generated M-QAM signal into the trained TFPNet network model, and output the predicted signal of each M-QAM signal. Calculate the normalized mean square error between the output predicted signal result and the signal transmitted by the corresponding passive optical network system channel to obtain the channel construction result of the TFPNet network model, characterize the complex nonlinear effects of the passive optical network communication system, reduce nonlinear damage according to the channel construction result, improve the stability of channel construction, and achieve high-precision channel construction in the passive optical network communication system.

[0025] The conditional vector feature sequence x of the newly generated M-QAM signal * Input into the trained TFPNet network model and output the predicted data Y′ of the M-QAM signal.

[0026] The normalized mean square error NMSE is calculated between the output prediction signal result Y′ and the signal Y transmitted through the corresponding passive optical network system channel.

[0027]

[0028] Among them, S is the total number of predicted symbols, Y′ is the predicted data, and Y is the true data.

[0029] TFPNet achieves high-precision channel construction, fully characterizes the fiber nonlinear effects on signals during transmission in passive optical network systems, reduces nonlinear damage based on channel construction results, improves the stability of channel construction, and can achieve high-precision channel construction in passive optical network communication systems.

[0030] In order to accurately simulate the complex characteristics of channel impairments in communications, it is preferred that the transmitted data sequence is mapped to 64-QAM constellation symbols.

[0031] Beneficial effects:

[0032] 1. The present invention discloses a channel construction method for a passive optical network communication system. The method adopts a time-frequency joint prediction method to construct different networks according to the characteristics of different damages to predict the damages, and uses a gradient back-propagation algorithm and an Adam parameter optimization algorithm to update the TFPNet network model parameters. Compared with the CGAN channel construction method, the method fully characterizes the optical fiber nonlinear effects on the signal during the transmission of the passive optical network system, and improves the accuracy and stability of the communication system channel construction.

[0033] 2. The present invention discloses a channel construction method for a passive optical network communication system, which uses a time-frequency joint prediction network TFPNet as a bridge layer in the autoencoder to provide accurate gradient information for end-to-end optimization based on the autoencoder, thereby achieving more accurate end-to-end optimization of the passive optical network system.

[0034] 3. The present invention discloses a channel construction method for a passive optical network communication system. According to the inter-symbol crosstalk caused by dispersion and a spatial light modulator SLM, a current M-QAM signal is combined with the n M-QAM signals before and after it to construct a conditional vector for characterizing the inter-symbol crosstalk; an M-QAM signal sequence after hollow-core optical fiber transmission is received, clock recovery and synchronization processing are performed on the M-QAM signal, and based on the synchronized M-QAM signal sequence, the current M-QAM signal is used as the real data of the M-QAM signal; a training data set is constructed based on the conditional vector and the real data, and a TFPNet network model is constructed using the data set data-driven method. Compared with previous data-driven modeling strategies, a time-frequency joint prediction network is proposed to achieve higher accuracy channel construction and reduce the complexity of training the TFPNet network model in view of the nonlinearity unique to passive optical network systems.

[0035] 4. The present invention discloses a channel construction method for a passive optical network communication system, which implements the channel construction of the passive optical network system based on the TFPNet network and uses a GPU to train the TFPNet network, thereby increasing the channel construction speed of the passive optical network system and improving the efficiency of data generation.

[0036] 5. The present invention discloses a channel construction method for a passive optical network communication system. By learning and processing the transmitted and received M-QAM signal data, the trained TFPNet network model constructs a nonlinear relationship between the characteristic sequence of the current M-QAM signal and its corresponding received data, and the nonlinear relationship can fully characterize the influence of the M-QAM signal caused by the nonlinear effect of the optical fiber during the optical fiber transmission process; accurate signal generation is achieved according to the nonlinear relationship, and the signal recovery is achieved based on the trained TFPNet network model without relying on the accurate parameter information of the optical fiber transmission link, thereby improving the generalization of the present invention and being able to be universally applied to the channel construction of all optical fiber communication links. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 The present invention discloses a flow chart of a channel construction method of a passive optical network communication system.

[0038] Figure 2 This is the network architecture diagram of TFPNet;

[0039] Figure 3 It is the structural diagram of TFPNet conditional vector;

[0040] Figure 4 Waveform comparison chart of TFPNet generated data and real channel output data;

[0041] Figure 5 Comparison chart of normalized mean square error of TFPNet and GAN generated data at different received optical powers at a transmission frequency of 191.6THz;

[0042] Figure 6 For a transmission frequency of 192.6 THz, the normalized mean square error comparison chart of TFPNet and GAN generated data at different received optical powers. DETAILED DESCRIPTION

[0043] The present invention will be described in detail below with reference to the accompanying drawings and embodiments. The technical problems solved by the technical solution of the present invention and the beneficial effects are also described. It should be noted that the described embodiments are only intended to facilitate the understanding of the present invention and do not have any limiting effect on it.

[0044] Embodiment 1:

[0045] In the transmission of passive optical network systems, due to the complex nonlinearity of devices in passive optical network systems, the nonlinear damage of optical fibers is intricate, so in the process of channel construction, traditional GAN ​​methods are difficult to achieve accurate convergence. In contrast, TFPNet adopts a dual residual structure and updates model parameters through forward propagation, which significantly improves the accuracy and stability of channel construction. In addition, TFPNet is based on a time-frequency joint prediction network to handle the complex nonlinear effects of passive optical network systems.

[0046] This embodiment discloses a channel construction method based on a time-frequency joint prediction network, which is applied to a passive optical network transmission system. Different networks are constructed according to the characteristics of different impairments to predict them, and the dual residual structure is used to update the TFPNet network model parameters through forward propagation to fully characterize the optical fiber nonlinear effects on the signal during the transmission process of the passive optical network system.

[0047] like Figure 1 As shown, the channel construction method of a passive optical network communication system disclosed in this embodiment is specifically implemented in the following steps:

[0048] Step 1: At the transmitting end of the passive optical network system, the transmitted binary data sequence is subjected to 64-QAM symbol mapping processing, and then the 64-QAM symbol sequence is subjected to up-sampling and pulse shaping processing. According to the inter-code interference caused by dispersion, the current 64-QAM signal is combined with the five 64-QAM signals before and after it to construct a conditional vector for characterizing the inter-code interference; the electrical signal is modulated onto the optical carrier by a Mach-Zehnder modulator, and the signal is amplified by an erbium-doped fiber amplifier. At the receiving end of the passive optical network system, the 64-QAM signal sequence after hollow-core optical fiber transmission is received, the 64-QAM signal is subjected to clock recovery and synchronization processing, and based on the synchronized 64-QAM signal sequence, the current 64-QAM signal is used as the real data of the 64-QAM signal, and a training data set is constructed based on the conditional vector and the real data.

[0049] At the transmitting end of the passive optical network system, the transmitted binary data sequence is processed by PAM-8 symbol mapping, and then the 64-QAM symbol sequence is upsampled and pulse-shaped. According to the inter-symbol crosstalk caused by dispersion and spatial light modulator, the current 64-QAM signal is combined with the five previous and next 64-QAM signals to construct a conditional vector x for characterizing the inter-symbol crosstalk. (i) =[x t-5 ,…,x t ,…,x t+5 ],i=1,2,…,128000, where x t Indicates the sample currently sent, [x t-5 ,…,x t-1 ] indicates the sample that has been sent before the current sample, [x t+1 ,…,x t+5 ] represents the sample to be transmitted, and the conditional vector structure is as follows Figure 2 As shown. The 64-QAM signal sequence is modulated onto an optical carrier by a Mach-Zehnder modulator, and an erbium-doped fiber amplifier is used to amplify the signal. The signal is coupled from the optical fiber to space through a polarization controller, a collimator, and linear polarization. At the receiving end of the passive optical network system, the 64-QAM signal sequence after hollow-core fiber transmission is received, and the 64-QAM signal is clock recovered and synchronized to obtain a 64-QAM signal sequence s = [s1, s2, …, s with a sequence length of 128,000 after synchronization. N ],i=1,2,…,N,where s i =[s 1 ,s 2 ,…,s 6 ] vector represents the i-th 64-QAM signal in the 64-QAM signal sequence. i(i=1,2,…,N), for each 64-QAM signal s i , which is taken as the real data y of the 64-QAM signal (i) =[s i ], i = 1, 2, ..., N, construct the real data feature sequence corresponding to each 64-QAM signal, and construct the training data set {x (i) ,y (i)}, where the first 80% of the training set is used as the training set, and the last 20% is used as the test set.

[0050] Step 2: Construct a TFPNet network model and loss function for passive optical network system channel construction. The TFPNet network model is a time-frequency joint prediction network, which consists of three sub-networks, namely, a time domain prediction network, a high frequency prediction network, and a nonlinear prediction network. The conditional vector of the 64-QAM signal and the training data set constructed by the real data and the noise vector with Gaussian distribution are respectively input into the time domain prediction network and the high frequency prediction network for prediction. After the original signal x is input into the time domain prediction network, it passes through the fully connected layer to obtain the output signal y1. In the fully connected layer, the linear features of the original signal are captured and fitted to predict the linear damage of the original signal. After the original signal x is input into the frequency domain prediction network, it passes through the time-frequency converter to convert the time domain signal x into a frequency domain signal fx. Then the frequency domain signal fx is input into the fully connected layer and activated by the activation function ReLu to obtain the frequency domain output signal fy2. Finally, the output signal y2 is obtained by passing through the frequency-time converter. In this fully connected layer, the high-frequency features of the original signal are captured and fitted to predict the nonlinear damage to the original signal x caused by the high-frequency fading caused by the nonlinear device. The original input signal x is subtracted from y1 and y2 to obtain x3 containing only nonlinear damage, and x3 is input into the nonlinear prediction network. After x3 is input into the time domain prediction network, it passes through the fully connected layer to obtain the output signal y3. In the fully connected layer, the nonlinear characteristics of x3 are captured and fitted to predict the nonlinear damage caused to the original signal by the passive optical network. Finally, the output signals of the three sub-networks are added to obtain the output signal y of the time-frequency joint prediction network. The data in the conditional vector feature sequence of the 64-QAM signal is fully processed by feature serialization fusion through the fully connected layer in the TFPNet network model. When processing the 64-QAM signal data at the current moment, the preceding 64-QAM signal data information and the subsequent 64-QAM signal data information in the conditional vector can be combined and utilized, that is, the sequence 64-QAM signal data in the conditional vector is subjected to serialization feature fusion, so as to better characterize the nonlinear interference relationship between the current 64-QAM signal and the preceding 64-QAM signal and the subsequent 64-QAM signal, and enhance the nonlinear construction capability of the TFPNet network model for the 64-QAM signal, and output a signal sequence representing the real channel effect. Through the training method of the joint prediction of time and frequency, a TFPNet network model capable of representing the channel of a real passive optical network system is obtained.

[0051] Step 3: For the TFPNet network model constructed in step 2 for passive optical network system channel construction, configure the parameters required for model training, set the learning rate, batch size, weight initialization method, optimization method, and number of iterations; use the training data set constructed in step 1 to train the TFPNet network model for passive optical network system channel construction, and construct the nonlinear relationship between the current 64-QAM signal feature sequence and its corresponding real channel signal through the trained TFPNet network model, which fully characterizes the network model of the real passive optical network channel response.

[0052] For the TFPNet network model constructed in step 2, the proposed TFPNet network model was trained and evaluated in Pytorch 1.6.0, with the learning rate set to 0.0005, the batch size set to 500, the pooling kernels k of the three blocks set to 6, 3, and 1 respectively, the random weight initialization method was used, the gradient backpropagation algorithm and the Adam optimization algorithm were used, and the maximum number of training steps was set to 500.

[0053] Using the training data set {x (i) ,y (i)},i=1,2,…,128000,The TFPNet network model constructed in step 2 is trained, and the gradient back propagation algorithm and Adam parameter optimization algorithm are used to determine the optimal model parameters to obtain a trained TFPNet network model. The conditional vector x of the current 64-QAM signal is constructed by the trained TFPNet network model (i) The corresponding real data y (i) The TFPNet network model can fully characterize the influence of the 64-QAM signal on the optical fiber transmission process caused by the optical fiber nonlinear effect.

[0054] Step 4: Input the newly generated 64-QAM signal at the transmitter as a conditional vector into the trained TFPNet network model, and output the predicted signal Y′ corresponding to each 64-QAM signal; compare the output predicted signal result with the corresponding actual channel data. Figure 3 The waveform diagram of the channel output data and the data generated based on TFPNet at 20GBaud, 64-QAM in 5 km hollow core fiber and 0dBm received optical power is shown. From the overall waveform diagram, the waveform of the real channel output and the waveform generated by TFPNet have consistent power. Zooming in on some waveforms, it is found that the waveforms of the TFPNet-generated data and the real channel data overlap to a large extent. The high consistency of the signal waveforms shows that TFPNet effectively characterizes the characteristics of the signal passing through the passive optical network channel.

[0055] In order to quantitatively express the gap between the modeling effect based on GAN and TFPNet and the output based on the real passive optical network channel, the normalized mean square error method is used, as shown in the following formula:

[0056]

[0057] Wherein S is the data length, Y is the data after clock synchronization and correlation at the receiving end of the passive optical network system, and Y′ is the data generated by the TFPNet.

[0058] like Figure 4 and Figure 5 As shown in the figure, using the TFPNet-based passive optical network channel construction method, 20GBaud, 64-QAM achieves a normalized mean square error of 1.2×10- -2 Compared with the GAN-based channel construction method, the modeling accuracy is improved by up to 34.5%, which effectively characterizes the complex channel effects of the communication system, reduces the impact of complex nonlinearities caused by multi-devices on channel construction, and improves the construction accuracy of the communication system channel.

[0059] The specific description above further illustrates the purpose, technical solutions and beneficial effects of the invention in detail. It should be understood that the above is only a specific embodiment of the present invention and does not limit the scope of protection of the present invention. Any modifications, replacements, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A channel construction method for a passive optical network communication system, characterized in that: The following steps are included: Step 1: At the transmitting end of the coherent optical fiber communication system, the transmitted binary data sequence is subjected to M-QAM constellation symbol mapping processing. The constellation diagram of the M-QAM modulation format signal has M standard constellation points. Therefore, the M-QAM signal is divided into M different categories. Each standard constellation point corresponds to a category and is numbered from 1 to M as a category label. At the receiving end of the coherent optical fiber communication system, an M-QAM signal sequence after optical fiber transmission is received, and for each M-QAM signal, the real data and imaginary data of the current M-QAM signal and the n M-QAM signals before and after it are combined as a conditional vector of the M-QAM signal; Step 2: Construct a TFPNet network model for passive optical network system channel construction; the TFPNet network model is a time-frequency joint prediction network, which consists of three sub-networks, namely a time domain prediction network, a high-frequency prediction network and a nonlinear prediction network; the output signals of the three sub-networks are added to obtain the output signal y of the time-frequency joint prediction network; the data in the conditional vector feature sequence of the M-QAM signal is fully feature serialized and fused through the fully connected layer in the TFPNet network model, and when processing the M-QAM signal data at the current moment, the preceding M-QAM signal data information and the subsequent M-QAM signal data information in the conditional vector are combined and utilized, that is, the sequence M-QAM signal data in the conditional vector is subjected to serialized feature fusion, so as to better characterize the nonlinear interference relationship between the current M-QAM signal and the preceding M-QAM signal and the subsequent M-QAM signal, and enhance the nonlinear construction capability of the TFPNet network model for the M-QAM signal, and output a signal sequence characterizing the real channel effect; through the training method of the time-frequency joint prediction, a TFPNet network model for characterizing the real passive optical network system channel is obtained; Step 3: For the TFPNet network model for passive optical network system channel construction constructed in step 2, configure the parameters required for model training, set the learning rate, batch size, weight initialization method, optimization method, and number of iterations; use the training data set constructed in step 1 to train the TFPNet network model for passive optical network system channel construction, and fully characterize the real passive optical network channel response through the trained TFPNet network model; Step 4: Input the feature sequence of the newly generated M-QAM signal into the trained TFPNet network model, and output the predicted signal of each M-QAM signal. Calculate the normalized mean square error between the output predicted signal result and the signal transmitted by the corresponding passive optical network system channel to obtain the channel construction result of the TFPNet network model, characterize the complex nonlinear effects of the passive optical network communication system, reduce nonlinear damage according to the channel construction result, improve the stability of channel construction, and achieve high-precision channel construction in the passive optical network communication system.

2. A channel construction method for a passive optical network communication system as claimed in claim 1, characterized in that: The implementation method of step one is: At the transmitting end of the passive optical network system, the transmitted binary data sequence is subjected to M-QAM symbol mapping, the M-QAM symbol sequence is subjected to up-sampling and pulse shaping, the current M-QAM signal is combined with the n M-QAM signals before and after it, and a conditional vector for characterizing inter-symbol interference is constructed; the electrical signal is modulated onto an optical carrier by a Mach-Zehnder modulator, and the center frequencies of the optical carrier are set to 191.6, 191.7, 191.8, 191.9, 192.0, 192.1, 192.2, 192.3, 192.4, 192.5, 192.6, 192.7, 192.

8. 192.9, 193.0, 193.1 THz, use an erbium-doped fiber amplifier to amplify the signal; modulate these 16 signals into one optical signal through a wavelength division multiplexer and transmit it through a hollow core optical fiber; at the receiving end of the passive optical network system, use a wavelength selective switch to demultiplex the received signal to obtain an M-QAM signal sequence after transmission through a specified band, perform clock recovery and synchronization processing on the M-QAM signal, and based on the synchronized M-QAM signal sequence, use the current M-QAM signal as the real data of the M-QAM signal sequence, and construct a training data set based on the conditional vector and the real data; At the transmitting end of the passive optical network system, the transmitted binary data sequence is subjected to M-QAM symbol mapping, and then the M-QAM symbol sequence is subjected to upsampling and pulse shaping. According to the inter-symbol crosstalk caused by dispersion, the current M-QAM signal is combined with the n M-QAM signals before and after it to construct a conditional vector [x t-n ,…,x t ,…,x t+n ], i = 1, 2, ..., N, where x t Indicates the sample currently sent, [x t-n ,…,x t-1 ] indicates the sample that has been sent before the current sample, [x t+1 ,…,x t+n ] represents the sample to be transmitted, n represents the number of symbols before and after, which is related to the strength of inter-symbol crosstalk; the M-QAM signal sequence is modulated onto the optical carrier by a Mach-Zehnder modulator, and an erbium-doped fiber amplifier is used to amplify the signal; At the receiving end of the passive optical network system, the M-QAM signal sequence after hollow-core optical fiber transmission is received, and the clock recovery and synchronization processing of the M-QAM signal are performed to obtain a PAM-8 signal sequence s=[s1,s2,…,s N ], i = 1, 2, ..., N, where the vector s i =[s 1 ,s 2 ,…,s 6 ] represents the i-th M-QAM signal in the M-QAM signal sequence; for the PAM-8 signal sequence s i (i=1,2,…,N), for each M-QAM signal s i , will s i As the real data y of the M-QAM signal (i) =[s i ], i = 1, 2, ..., N, N represents the size of the training data set, construct the real data feature sequence corresponding to each M-QAM signal, and construct the training data set {x (i) ,y (i) }.

3. A channel construction method for a passive optical network communication system as claimed in claim 2, characterized in that: The implementation method of step 2 is: Construct a TFPNet network model and loss function for passive optical network system channel construction. The TFPNet network model is a time-frequency joint prediction network, which consists of three sub-networks, namely a time domain prediction network, a high frequency prediction network and a nonlinear prediction network; The training data set is constructed by using the conditional vector of the M-QAM signal and the real data, and the training data set and the noise vector with Gaussian distribution are simultaneously input into the time domain prediction network and the high frequency prediction network for prediction; The original signal x is input into the time domain prediction network and then passes through the fully connected layer to obtain the output signal y1, as shown in formula (1). y1=w*x+b (1) Where W and b represent the weight and bias in the fully connected layer, respectively. In the fully connected layer, the linear features of the original signal are captured and fitted to predict the linear damage of the original signal. The original signal x is input into the frequency domain prediction network and then passes through the time-frequency converter to convert the time domain signal x into the frequency domain signal X[k]. The process is shown in formula (2): Where x[n] represents the nth symbol, and then the frequency domain signal X[k] is input into the fully connected layer and activated by the activation function ReLu to obtain the frequency domain output signal Y[k], as shown in formula (3), Y[k]=ReLu(w*X[k]+b) (3) Where W and b represent the weight and bias in the fully connected layer respectively. The output signal y2 is obtained by inverse Fourier transform through the frequency-time converter. In the fully connected layer, the high-frequency features of the original signal are captured and fitted to predict the nonlinear damage to the original signal x caused by high-frequency fading caused by nonlinear devices. The original input signal x is subtracted from y1 and y2 to obtain x3 containing only nonlinear damage, and x3 is input into the nonlinear prediction network. After x3 is input into the time domain prediction network, it passes through the fully connected layer to obtain the output signal y3, as shown in formula (4). y3=w*x3+b (4) In the fully connected layer, the nonlinear characteristics of x3 are captured and fitted to predict the nonlinear damage caused to the original signal by the passive optical network system; the output signals of the three sub-networks are added to obtain the output signal y of the time-frequency joint prediction network; the data in the conditional vector feature sequence of the M-QAM signal is fully feature-serialized and fused through the fully connected layer in the TFPNet network model; when processing the M-QAM signal data at the current moment, the preceding M-QAM signal data information and the subsequent M-QAM signal data information in the conditional vector are combined and utilized, that is, the sequence M-QAM signal data in the conditional vector is subjected to serialized feature fusion to better characterize the nonlinear interference relationship between the current M-QAM signal and the preceding M-QAM signal and the subsequent M-QAM signal, thereby improving the nonlinear construction capability of the TFPNet network model for the M-QAM signal, and outputting a signal sequence characterizing the real channel effect; through the training method of the time-frequency joint prediction, a TFPNet network model for characterizing the channel of the real passive optical network system is obtained.

4. A channel construction method for a passive optical network communication system as claimed in claim 3, characterized in that: The implementation method of step three is: For the TFPNet network model constructed in step 2 for passive optical network system channel construction, configure the parameters required for model training, set the learning rate, batch size, weight initialization method, optimization method, and number of iterations; Using the training data set {x (i) ,y (i) }, train the TFPNet network model constructed in step 2, use the gradient back propagation algorithm and the Adam parameter optimization algorithm to determine the optimal model parameters, and obtain a trained TFPNet network model; construct the conditional vector x of the current M-QAM signal through the trained TFPNet network model (i) The corresponding real data y (i) The TFPNet network model is used to fully characterize the influence of the M-QAM signal on the hollow-core optical fiber transmission process caused by the optical fiber nonlinear effect.

5. A channel construction method for a passive optical network communication system as claimed in claim 4, characterized in that: The conditional vector feature sequence x of the newly generated M-QAM signal * Input into the trained TFPNet network model, and output the predicted data Y′ of the M-QAM signal; Calculate the normalized mean square error NMSE between the output prediction signal result Y′ and the signal Y transmitted by the corresponding passive optical network system channel; Among them, S is the total number of predicted symbols, Y′ is the predicted data, and Y is the real data; The TFPNet model achieves high-accuracy channel construction, fully characterizes the fiber nonlinear effects on signals during transmission in passive optical network systems, reduces nonlinear damage based on the channel construction results, improves the stability of channel construction, and can achieve high-precision channel construction in passive optical network communication systems.

6. A channel construction method for a passive optical network communication system as claimed in claim 1, 2, 3, 4 or 5, characterized in that: The transmitted data sequence is mapped into 64-QAM constellation symbols.

Images