A PAM4 Equalization Transmission System and Method under Bandwidth Limitation

Abstract

The present invention relates to a PAM4 equalization transmission system under bandwidth limitation, which includes: at the transmitting end, the PRBS sequence signal to be transmitted is mapped into an initial PAM4 signal, and after digital signal processing, a PAM4 pulse signal is generated and sent; the transmission link performs optoelectronic conversion on the PAM4 pulse signal to obtain a transmitted optical signal for transmission, and at the end of the transmission, it is converted into an adjusted electrical signal and sent to the receiving end; the receiving end demodulates the received adjusted electrical signal into the PRBS signal to be transmitted, completing the signal transmission; at the receiving end, the feed-forward equalizer performs equalization processing on the adjusted electrical signal that has been signal synchronized and matched filtered, generates a serial signal, then converts it into a multi-dimensional parallel signal, inputs it into a pre-trained LSTM network for equalization processing, and outputs a multi-dimensional PAM4 signal; the PAM demapping unit is used for demapping to obtain the PRBS signal to be transmitted, completing the PAM4 equalization transmission.

Description

Technical Field

[0001] The present invention relates to the technical field of PAM4 transmission systems, and in particular to a PAM4 equalization transmission system and method under bandwidth limitation. Background Art

[0002] With the rapid development of mobile devices, cloud computing and big data, the demand for high-speed and large-capacity communication in data centers remains high. Inside the data center, the four-level pulse amplitude modulation format PAM4 (4 Pulse Amplitude Modulation) has been widely used, and the intensity modulation direct detection (IM / DD) system can ensure the low cost and low complexity of the transmission system. However, the inter-symbol interference (ISI) caused by insufficient system bandwidth is one of the main channel distortions of low-cost, low-bandwidth optical high-speed transmission systems. In addition, the non-linear response caused by the response of devices (drivers, modulators, etc.) and the interaction between dispersion and direct detection will also affect the performance of the system. Therefore, when better-performance devices cannot be replaced, many advanced DSP algorithms have been developed to overcome problems such as bandwidth limitation brought by the devices.

[0003] In low-cost bandwidth-limited systems, super-Nyquist (FTN) technology or coding technology can be used to mitigate the impact of bandwidth limitation and significantly improve spectral efficiency, but this will introduce greater system impairments. And in large-capacity FTN transmission systems, traditional linear equalization methods can no longer cope with severe non-linear inter-symbol interference ISI. Therefore, many non-linear equalization schemes have been introduced to improve the performance of received signals, such as combining the use of a Volterra equalizer with the maximum likelihood sequence estimation (MLSE) scheme, or using the look-up table method (LUT), etc. However, in severely bandwidth-limited transmission systems, these methods cannot achieve good performance. To further improve system performance, neural networks have now become an effective solution for non-linear equalization tasks. Many neural network structures such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), etc. have been applied to the field of optical fiber communication. CNNs are particularly effective in examining data images and are one of the most useful image processing methods in deep learning. In optical communication, most data is represented in the format of digital signals. Different from CNNs designed for image data, RNNs are specifically proposed for sequential data. Among them, a variant of RNN, the LSTM network, is very suitable for compensating for the ISI problem in FTN systems and the non-linearity problem in IM / DD links because it can fully exploit the temporal dependence of time series.

[0004] However, existing LSTM networks mostly use one-dimensional sliding window serial input. In this way, there will be a large amount of similar data in adjacent windows, and calculating similar input data may bring unnecessary complexity to the network. Moreover, a large amount of similar data will prevent the LSTM network from obtaining signal features well, resulting in insufficient equalization. And the equalizer of the one-dimensional serial input LSTM network has little improvement in the transmission bit error rate performance of the severely bandwidth-limited FTN-PAM4 system.

[0005] In summary, due to the limited bandwidth of the devices in the optical transmission system and the severe inter-symbol interference (ISI) problem introduced by the FTN technology used, the transmitted PAM4 signal is severely distorted. When using the existing technology to equalize the signal distortion by using the LSTM network, the problem of inaccurate signal feature extraction caused by the one-dimensional serial input of the LSTM network is not considered, resulting in insufficient equalization of the transmitted signal in the optical transmission system, inability to effectively reduce the transmission bit error rate, and low signal transmission accuracy. Summary of the Invention

[0006] Therefore, the technical problem to be solved by the present invention is to overcome the problems in the prior art that the inaccurate signal feature extraction and insufficient signal equalization caused by the one-dimensional serial input of the LSTM network cannot effectively reduce the transmission bit error rate and the signal transmission accuracy is low.

[0007] To solve the above technical problem, the present invention provides a PAM4 equalization transmission system under bandwidth limitation, including:

[0008] A transmitting end, which maps the PRBS sequence signal to be transmitted into an initial PAM4 signal, and after digital signal processing, generates and transmits a PAM4 pulse signal;

[0009] A transmission link, which performs optoelectronic conversion on the PAM4 pulse signal to obtain a transmitted optical signal for transmission, and at the end of the transmission, converts it into an adjusted electrical signal and sends it to the receiving end;

[0010] A receiving end, which demodulates the received adjusted electrical signal into the PRBS signal to be transmitted, completing the signal transmission, including:

[0011] A feed-forward equalizer, which performs equalization processing on the received adjusted electrical signal after signal synchronization and matching filtering, and generates a serial signal for output;

[0012] A pre-trained LSTM network, which converts the serial signal into a multi-dimensional parallel signal, and

[0013] performs equalization processing after parallel input and outputs a multi-dimensional PAM4 signal;

[0014] The PAM demapping unit demaps the multi-dimensional PAM4 signal to obtain the PRBS signal to be transmitted.

[0015] Preferably, the transmitting end sequentially includes, along the signal transmission direction:

[0016] The sequence mapping unit maps the PRBS sequence signal to be transmitted into an initial PAM4 signal and outputs it;

[0017] The FTN acceleration unit improves the spectral efficiency of the initial PAM4 signal and outputs an optimized PAM4 signal;

[0018] The upsampling unit increases the sampling rate of the optimized PAM4 signal and outputs a smoothed PAM4 signal;

[0019] The pulse shaping unit performs pulse shaping on the smoothed PAM4 signal and outputs a PAM4 pulse signal.

[0020] Preferably, the pulse shaping unit is an RRC roll-off filter.

[0021] Preferably, the transmission link includes, sequentially connected along the signal propagation direction through an optical fiber:

[0022] The waveform generator converts the PAM4 pulse signal into an initial PAM4 electrical signal and outputs it;

[0023] The MZM modulator modulates the initial PAM4 electrical signal into a PAM4 optical signal and outputs it;

[0024] The first variable optical attenuator adjusts the power of the PAM4 optical signal and outputs an adjusted optical signal;

[0025] The photodiode converts the adjusted optical signal into an adjusted electrical signal and outputs it;

[0026] The real-time oscilloscope captures and analyzes the output of the adjusted electrical signal.

[0027] Preferably, the transmission link further includes a noise control module, including:

[0028] The second variable optical attenuator, whose input end is connected to the output end of the MZM modulator through an optical fiber, adjusts the power of the PAM4 optical signal;

[0029] The erbium-doped fiber amplifier, whose input end is connected to the output end of the second variable optical attenuator, amplifies the PAM4 optical signal whose power has been adjusted by the second variable optical attenuator and outputs it to the first variable optical attenuator.

[0030] Preferably, the feed-forward equalizer performs equalization processing on the input electrical signal, including:

[0031] The input electrical signal is passed through multiple delay units to obtain the delayed electrical signals output by each delay unit;

[0032] Multiply each delayed electrical signal by its corresponding tap coefficient to obtain multiple multiplication results;

[0033] Sum the multiple multiplication results after linear weighting to generate a serial signal output, expressed as:

[0034]

[0035] where y(t) represents the serial signal output by the feed-forward equalizer at time t, x(t) represents the electrical signal input to the feed-forward equalizer at time t, and x(t - n) represents the delayed electrical signal output by the nth delay unit at time t; W n represents the nth tap coefficient in the feed-forward equalizer, which is iteratively updated using the recursive least squares method, 1 ≤ n ≤ N, and N represents the total number of tap coefficients in the feed-forward equalizer.

[0036] Preferably, an LSTM network is used to perform equalization processing on the multi-dimensional parallel signal, and a multi-dimensional PAM4 signal is output, including:

[0037] Input the multi-dimensional parallel signal at the current moment into the multi-dimensional input layer of the LSTM network and process it into a feature vector;

[0038] Input the feature vector into the LSTM layer of the LSTM network, and calculate the input gate, forget gate, output gate, and candidate state at the current moment with the external state and candidate state at the previous moment, respectively expressed as:

[0039] Input gate: i t = σ(W i z t + U i h t-1 + b i );

[0040] Forget gate: f t = σ(W f z t + U f h t-1 + b f );

[0041] Output gate: o t = σ(W o z t + U o h t-1 + b o );

[0042] Candidate state:

[0043] External state: h t = o t ⊙tanh(C t );

[0044] Update the candidate state C based on the forget gate and the input gate t ; Update the external state h based on the output gate t , and convert it into a multi-dimensional PAM4 signal output in the multi-dimensional output layer of the input LSTM network;

[0045] Among them, σ represents the activation function; W i , W f , W o and W C respectively represent the input weights of the input gate, the forget gate, the output gate and the candidate state; U i , U f , U o and U C respectively represent the recurrent weights of the input gate, the forget gate, the output gate and the candidate state; b i , b f , b o and b C respectively represent the bias parameters of the input gate, the forget gate, the output gate and the candidate state; z t represents the multi-dimensional parallel signal; h t-1 represents the external state at the previous moment.

[0046] Preferably, the training process of the LSTM network includes:

[0047] Map the sequence generated based on the PRBS sequence seed number 1 to obtain multiple initial PAM4 signals;

[0048] For each initial PAM4 signal, obtain the output PAM4 signal obtained by demapping at the transmitter, the transmission link and the output end; Multiple groups of initial PAM4 signals and their corresponding output PAM4 signals form a training set;

[0049] Preset the number of layers and the number of neurons in the input layer and the output layer of the initial LSTM network, preset the learning rate, the learning rate decay period, the learning rate decay factor and the maximum number of learning times, and use the training set to train the initial LSTM network;

[0050] Based on the PAM4 signal output by the output end after each initial PAM4 signal in the training set passes through the LSTM network and its corresponding initial PAM4 signal, construct a loss function; Use the Adam optimizer to optimize the LSTM network parameters until the loss function converges to obtain a pre-trained LSTM network.

[0051] Preferably, the receiving end further includes an error rate calculation unit, which compares the initial PAM4 signal and the multi-dimensional PAM4 signal, and obtains the error rate based on the quotient of the number of bits different between the multi-dimensional PAM4 signal and the initial PAM4 signal and the total number of bits of the initial PAM4 signal.

[0052] An embodiment of the present invention also provides an equalization transmission method applied to the PAM4 equalization transmission system under bandwidth limitation as described above, which is applied to the receiving end and includes:

[0053] After synchronizing and matching filtering the received adjusted electrical signal by using a signal synchronization unit and a matching filter, perform equalization processing through a feed-forward equalizer to output a serial signal;

[0054] Convert the serial signal into a multi-dimensional parallel signal and input it into a pre-trained LSTM network for equalization to obtain a multi-dimensional PAM4 signal;

[0055] The multi-dimensional PAM4 signal is demapped by a PAM demapping unit to obtain the PRBS signal to be transmitted.

[0056] The above technical solution of the present invention has the following beneficial effects compared with the prior art:

[0057] In the PAM4 equalization transmission system under bandwidth limitation described in the present invention, by adding a feed-forward equalizer FFE and an LSTM network at the receiving end of the system, the serious signal distortion problem in the bandwidth-limited PAM4 transmission system is solved; the feed-forward equalizer is used to adjust the amplitude and phase of the signal, pre-compensate the linear damage suffered by the received adjusted electrical signal of the receiver, and generate a one-dimensional serial signal; convert the one-dimensional serial signal into a multi-dimensional parallel signal and input it into the LSTM network for equalization processing; the multi-dimensional parallel signal increases the diversity of data samples, avoids the problem of similar data in adjacent windows, and can make full use of the characteristics of the multi-dimensional parallel signal, predicts and corrects the current distortion situation through the output of the previous stage, so that the signal equalization is sufficient and the system error rate is effectively reduced. And the FFE and the LSTM network are combined and used. The FFE is responsible for initially equalizing and correcting the signal at the physical layer, and the LSTM network is used to deeply process the signal at a higher level to further reduce the distortion, so that under the condition of bandwidth limitation, the signal distortion problem of the PAM4 transmission system is effectively solved, the signal distortion caused by interference is effectively corrected, the quality of the received signal is significantly improved, and the error rate is reduced.

[0058] In the transmission link of the present invention, a noise control module composed of a second variable optical attenuator and an erbium-doped fiber amplifier is provided; the second variable optical attenuator can effectively control the intensity of the optical signal, ensure that the signal power input to the receiver is within the dynamic range of the receiving end, contribute to maintaining the stability of signal transmission, and reduce the noise and distortion caused by over-strong signals; the erbium-doped fiber amplifier directly amplifies the PAM4 optical signal, compensates for the signal attenuation caused by optical devices and optical fibers, enhances the signal transmission ability, extends the signal transmission distance, improves the signal-to-noise ratio of the signal, and further reduces the impact of noise on the signal quality. By using the noise control module, precise control of the intensity of the PAM4 optical signal and improvement of the signal quality are achieved, effectively reducing the noise level of the system and improving the communication quality. Brief Description of the Drawings

[0059] In order to make the content of the present invention easier to be clearly understood, the following further details the present invention according to the specific embodiments of the present invention in conjunction with the drawings, where

[0060] Figure 1 is the structural diagram of the PAM4 equalization transmission system under bandwidth limitation provided by the present invention;

[0061] Figure 2 is the schematic structural diagram of the feedforward equalizer and the LSTM network provided by the present invention;

[0062] Figure 3 is the schematic structural diagram of the LSTM network provided by the present invention;

[0063] Figure 4 is the end-to-end frequency domain response curve diagram of the PAM4 equalization transmission system under bandwidth limitation provided by the present invention;

[0064] Figure 5 is the schematic structural diagram of the LSTM layer in the LSTM network provided by the present invention;

[0065] Figure 6 is the comparison diagram of the measured BER curves of the 30GBd PAM4 signal in back-to-back, 10-kilometer single-mode fiber, 10-kilometer 7-core fiber transmission and full-core transmission provided by the present invention;

[0066] Figure 7 is the comparison diagram of the measured BER curves of applying three-dimensional input LSTM equalization in 30G-PAM4 back-to-back, 10km single-mode fiber and 10km seven-core fiber transmission provided by the present invention;

[0067] Figure 8 is the comparison diagram of the measured BER curves of the LSTM network under different input dimensions provided by the present invention. Detailed Embodiment

[0068] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, so that those skilled in the art can better understand the present invention and be able to implement it. However, the exemplified embodiments are not intended to limit the present invention.

[0069] The problem solved by the present invention is that in a low-cost bandwidth-limited transmission system, when using super-Nyquist FTN technology or coding technology to improve spectral efficiency, a large system impairment will be introduced. Therefore, in the present invention, a feed-forward equalizer FFE and an LSTM network equalizer with multi-dimensional parallel input are combined to solve the serious signal distortion problem in a severely bandwidth-limited FTN-PAM4 optical transmission system.

[0070] Referring to Figure 1 As shown, the structural diagram of a PAM4 equalization transmission system under bandwidth limitation provided by the present invention specifically includes:

[0071] A transmitting end, which maps the PRBS sequence signal to be transmitted into an initial PAM4 signal, and after digital signal processing, generates a PAM4 pulse signal for transmission;

[0072] A transmission link, which performs optoelectronic conversion on the PAM4 pulse signal to obtain a transmitted optical signal for transmission, and at the end of the transmission, converts it into an adjusted electrical signal and sends it to the receiving end;

[0073] A receiving end, which demodulates the received adjusted electrical signal into the PRBS signal to be transmitted, completing the signal transmission, including:

[0074] A feed-forward equalizer, which performs equalization processing on the received adjusted electrical signal that has been signal synchronized and matched filtered, and generates a serial signal for output;

[0075] A pre-trained LSTM network, which converts the serial signal into a multi-dimensional parallel signal, and

[0076] After parallel input, performs equalization processing and outputs a multi-dimensional PAM4 signal;

[0077] A PAM demapping unit, which demaps the multi-dimensional PAM4 signal to obtain the PRBS signal to be transmitted.

[0078] In the embodiments of the present invention, a feed - forward equalizer (FFE) and an LSTM network equalizer with multi - dimensional parallel input are combined to solve the serious signal distortion problem in a severely bandwidth - limited FTN - PAM4 optical transmission system. The feed - forward equalizer (FFE) can pre - compensate for the linear impairments suffered by the transmitted signal, so as to facilitate the subsequent LSTM network to better learn the sample features. And using a multi - dimensional parallel data input LSTM network can increase the diversity of data samples, avoiding the problem of similar data in adjacent windows. Moreover, multi - dimensional parallel input can increase the training labels of the signal, enabling the network to better learn the band - limited channel configuration to compensate for signal impairments.

[0079] Specifically, in the embodiments of the present invention, the transmitting end, in the signal transmission direction, sequentially includes:

[0080] A sequence mapping unit that maps the PRBS sequence signal to be transmitted into an initial PAM4 signal for output;

[0081] An FTN acceleration unit that improves the spectral efficiency of the initial PAM4 signal and outputs an optimized PAM4 signal;

[0082] An up - sampling unit that increases the sampling rate of the optimized PAM4 signal and outputs a smoothed PAM4 signal;

[0083] A pulse shaping unit that performs pulse shaping on the smoothed PAM4 signal and outputs a PAM4 pulse signal.

[0084] Among them, the pulse shaping unit is an RRC roll - off filter.

[0085] Specifically, in the embodiments of the present invention, the transmission link includes, in the signal propagation direction along the optical fiber, sequentially connected:

[0086] A waveform generator that converts the PAM4 pulse signal into an initial PAM4 electrical signal for output;

[0087] An MZM modulator that modulates the initial PAM4 electrical signal into a PAM4 optical signal for output;

[0088] A first variable optical attenuator that adjusts the power of the PAM4 optical signal and outputs an adjusted optical signal;

[0089] A photodiode that converts the adjusted optical signal into an adjusted electrical signal for output;

[0090] A real - time oscilloscope that captures and analyzes the output of the adjusted electrical signal.

[0091] In another embodiment of the present invention, the transmission link further includes a noise control module, which includes:

[0092] A second variable optical attenuator, whose input end is connected to the output end of the MZM modulator through an optical fiber to adjust the power of the PAM4 optical signal;

[0093] An erbium-doped fiber amplifier, whose input end is connected to the output end of a second variable optical attenuator, amplifies the PAM4 optical signal whose power is adjusted by the second variable optical attenuator, and outputs it to the first variable optical attenuator.

[0094] In the present invention, a noise control module composed of a second variable optical attenuator and an erbium-doped fiber amplifier is provided in the transmission link; the second variable optical attenuator can effectively control the intensity of the optical signal, ensure that the signal power input to the receiver is within the dynamic range of the receiving end, contribute to maintaining the stability of signal transmission, and reduce the noise and distortion caused by the over-strong signal; the erbium-doped fiber amplifier directly amplifies the PAM4 optical signal, compensates for the signal attenuation caused by optical devices and optical fibers, enhances the signal transmission ability, extends the signal transmission distance, improves the signal-to-noise ratio of the signal, and further reduces the influence of noise on the signal quality. The precise control of the intensity of the PAM4 optical signal and the improvement of the signal quality are realized by using the noise control module, effectively reducing the noise level of the system and improving the communication quality.

[0095] Refer to Figure 2 As shown, it is a schematic structural diagram of the feed-forward equalizer and the LSTM network provided by the present invention; the present invention uses a method combining a feed-forward equalizer FFE and an LSTM network equalizer with multi-dimensional parallel inputs to overcome the limitations of traditional equalization methods, enabling the network to better learn the band-limited channel configuration and better compensating for signal impairments.

[0096] Specifically, the feed-forward equalizer performs equalization processing on the input electrical signal, including:

[0097] Pass the input electrical signal through multiple delay units to obtain the delayed electrical signals output by each delay unit;

[0098] Multiply each delayed electrical signal by its corresponding tap coefficient to obtain multiple multiplication results;

[0099] Linearly weight and sum the multiple multiplication results to generate a serial signal output, expressed as:

[0100]

[0101] Among them, y(t) represents the serial signal output by the feed-forward equalizer at time t, x(t) represents the electrical signal input to the feed-forward equalizer at time t, and x(t - n) represents the delayed electrical signal output by the nth delay unit at time t; W n represents the nth tap coefficient in the feed-forward equalizer, which is iteratively updated using the recursive least squares method, 1 ≤ n ≤ N, and N represents the total number of tap coefficients in the feed-forward equalizer.

[0102] Specifically, the training process of the LSTM network includes:

[0103] Map the sequence generated based on the PRBS sequence seed number 1 to obtain multiple initial PAM4 signals;

[0104] For each initial PAM4 signal, obtain the output PAM4 signal obtained by demapping through the transmitter, transmission link, and output end; multiple groups of initial PAM4 signals and their corresponding output PAM4 signals form a training set;

[0105] Preset the number of layers and the number of neurons in the input layer and output layer of the initial LSTM network, preset the learning rate, learning rate decay period, learning rate decay factor, and maximum number of learning times, and use the training set to train the initial LSTM network;

[0106] Based on the PAM4 signal output by the output end after each initial PAM4 signal in the training set passes through the LSTM network and its corresponding initial PAM4 signal, construct a loss function; use the Adam optimizer to optimize the LSTM network parameters until the loss function converges to obtain a pre-trained LSTM network.

[0107] Specifically, refer to Figure 3 As shown, it is a schematic structural diagram of the LSTM network provided by the present invention; based on the above pre-trained LSTM network, perform equalization processing on the multi-dimensional parallel signal and output a multi-dimensional PAM4 signal, including:

[0108] Input the multi-dimensional parallel signal at the current moment into the multi-dimensional input layer of the LSTM network and process it into a feature vector;

[0109] Input the feature vector into the LSTM layer of the LSTM network, and calculate the input gate, forget gate, output gate, and candidate state at the current moment with the external state and candidate state at the previous moment, which are respectively expressed as:

[0110] Input gate: i t = σ(W i z t + U i h t-1 + b i );

[0111] Forget gate: f t = σ(W f z t + U f h t-1 + b f );

[0112] Output gate: o t = σ(W o z t + U o ht-1 +b o );

[0113] Candidate status:

[0114] External status: h t = o t ⊙tanh(C t );

[0115] Update the candidate status C based on the forget gate and the input gate t ; Update the external status h based on the output gate t , in the multi-dimensional output layer of the input LSTM network, convert it into a multi-dimensional PAM4 signal for output;

[0116] Among them, σ represents the activation function; W i , W f , W o and W C respectively represent the input weights of the input gate, forget gate, output gate and candidate status; U i , U f , U o and U C respectively represent the recurrent weights of the input gate, forget gate, output gate and candidate status; b i , b f , b o and b C respectively represent the bias parameters of the input gate, forget gate, output gate and candidate status; z t represents the multi-dimensional parallel signal; h t-1 represents the external status at the previous moment.

[0117] In the embodiment of the present invention, the receiving end further includes an error rate calculation unit, which compares the initial PAM4 signal and the multi-dimensional PAM4 signal, and obtains the error rate based on the quotient of the number of bits different between the multi-dimensional PAM4 signal and the initial PAM4 signal and the total number of bits of the initial PAM4 signal.

[0118] The PAM4 equalization transmission system under bandwidth limitation described in the present invention solves the serious signal distortion problem in the PAM4 transmission system with limited bandwidth by adding a feed-forward equalizer FFE and an LSTM network at the receiving end of the system; uses the feed-forward equalizer to adjust the amplitude and phase of the signal, pre-compensate the linear damage suffered by the regulated electrical signal received by the receiver, and generate a one-dimensional serial signal; converts the one-dimensional serial signal into a multi-dimensional parallel signal and inputs it into the LSTM network for equalization processing; the multi-dimensional parallel signal increases the diversity of data samples, avoids the problem of similar data in adjacent windows, and can make full use of the characteristics of the multi-dimensional parallel signal, predicts and corrects the current distortion situation through the output of the previous stage, so that the signal equalization is sufficient and the system bit error rate is effectively reduced. And the FFE and the LSTM network are combined and used. The FFE is responsible for initially equalizing and correcting the signal at the physical layer, and the LSTM network deeply processes the signal at a higher level to further reduce distortion, so that under the condition of limited bandwidth, the signal distortion problem of the PAM4 transmission system is effectively solved, the signal distortion caused by interference is effectively corrected, the quality of the received signal is significantly improved, and the bit error rate is reduced.

[0119] Specifically, based on the above embodiments, in the embodiments of the present invention, to solve the serious distortion problem of FTN-PAM4 signals, the PAM4 equalization transmission system provided by the present invention is used to transmit PAM4 signals. The specific steps include:

[0120] S201: At the sending end, map the PRBS sequence into a uniform PAM4 signal, and after being processed by the FTN acceleration module, perform pulse shaping on the input signal through an RRC roll-off filter.

[0121] Generally speaking, the lower the roll-off coefficient of the RRC filter, the higher the spectral efficiency of the output signal. However, too small a roll-off coefficient is difficult to achieve in practical applications and has no practical significance. Therefore, in this embodiment, the roll-off coefficient of the RRC filter is set to 0.35.

[0122] S202: After the DSP at the sending end is completed, load the generated signal onto an arbitrary waveform generator (AWG), thereby generating a 30G-FTN-PAM4 electrical signal. The PAM4 electrical signal is modulated into an optical signal by a Mach-Zehnder modulator (MZM), where the wavelength of the optical signal is 1550 nm. The optical signal can be transmitted to a 10-kilometer seven-core optical fiber without any optical amplifier, simulating the scenario of short-distance optical interconnection in a data center.

[0123] In this embodiment, to test the performance of the algorithm, a variable optical attenuator (VOA1) and an erbium-doped fiber amplifier (EDFA) are added to control the noise level of the transmission system, thereby testing the bit error rate. Finally, after adjusting the optical power through VOA2, the transmitted optical signal is received by a 10 GHz PIN photodiode (PD) and data sampling is performed by a real-time oscilloscope (RTO). In this embodiment, the end-to-end frequency-domain response curve of the PAM4 equalization transmission system under bandwidth limitation is as shown in Figure 4 shown, and the -3 dB and -10 dB bandwidths of the experimental system are 5.47 GHz and 10.16 GHz respectively.

[0124] S203: At the receiving end, the received data is sent to the equalizer for data processing after synchronization and downsampling; in the embodiment of the present invention, the equalizer includes a feed-forward equalizer FFE and an LSTM network;

[0125] S203-1: The distorted PAM4 electrical signal received at the receiving end is input into the feed-forward equalizer FFE for equalization after synchronization and matched filtering. The FFE is implemented as a finite impulse response filter, which delays and weights the received data bits and sums them up. The signal received at the receiving end is multiplied by N + 1 corresponding tap coefficients Wn after passing through N delay units, and then the linearly weighted results are accumulated, and finally sent to the decision maker for decision. Among them, the tap coefficients Wn are iteratively updated using the recursive least squares algorithm (RLS). The expression of the signal after passing through the FFE is:

[0126] where N is the number of tap coefficients of the FFE, x(t) represents the input signal of the FFE at time t, and y(t) represents the output signal of the FFE at time t.

[0127] S203-2: Then the signal is input into the LSTM network in a multi-dimensional parallel manner. The LSTM network mainly controls the information transfer path through the input gate i t , the forget gate f t and the output gate o t .

[0128] The structural diagram of the LSTM network equalizer with multi-dimensional input and output is as shown in Figure 3 shown, and the schematic diagram of the structure of the LSTM layer is as shown in Figure 5 shown. The multi-dimensional input layer inputs a multi-dimensional PAM signal, and the output layer is also a multi-dimensional signal. In this embodiment, the selected input and output dimensions are the same, both are three-dimensional signals. The number of neurons in the LSTM layer is 40. At each time t, the internal state of the LSTM network records the historical information of the current time and passes the information to the next time. From Figure 5It can be seen that through the LSTM recurrent unit, the entire network can establish long-distance temporal dependencies. Specifically, the calculation formulas for the three gates of the LSTM network are expressed as:

[0129] Input gate: i t = σ(W i z t + U i h t-1 + b i );

[0130] Forget gate: f t = σ(W f z t + U f h t-1 + b f );

[0131] Output gate: o t = σ(W o z t + U o h t-1 + b o );

[0132] Candidate state:

[0133] External state: h t = o t ⊙ tanh(C t ));

[0134] Where σ represents the activation function, which is the logistic function in this embodiment, z t is the input at the current moment, and h t-1 is the external state at the previous moment.

[0135] Specifically, the acquisition of the pre-trained LSTM network used in the embodiments of the present invention includes:

[0136] First, use the sequence generated by the PRBS sequence seed number 1 for PAM4 mapping. After digital signal processing and transmission through the PAM4 equalization transmission system provided in this embodiment, send multiple groups of received signals to the equalizer for training, use this as the training set, and save the trained LSTM network;

[0137] Second, after using the sequence generated by the PRBS sequence seed number 2 for PAM4 mapping and then through the PAM4 equalization transmission system provided in this embodiment for digital signal processing and transmission, use the multiple groups of received signals as the test set of the LSTM network;

[0138] In this embodiment, the number of input layers and output layers of the LSTM network are both set to 3. If higher-dimensional signals are used for training, the number of input and output layers will be increased as the input dimension of the signal increases. The number of neurons in the input layer and output layer of the LSTM network are both configured to be 40. The learning rate of the LSTM network is set to 0.005, the maximum number of times for the entire data in the dataset to fully train the model is set to 350, the learning rate decay period is set to 300, and the learning rate decay factor is set to 0.2. After training, the trained LSTM network is saved and then applied to the test set for testing until the test results meet the expectations, and a pre-trained LSTM network is obtained.

[0139] In this embodiment, three parallel input signals are selected; these gating mechanisms can be regarded as a layer of sub-neural network and are trained to be in a gated state to respectively control the network to process the main data, forget, and selectively remember new units. In this embodiment, the received multi-frame FTN-PAM4 signals are used as the multi-dimensional parallel input of the LSTM network and are converted into the form of an input matrix and sent into the LSTM network unit to multiply with the weighted matrix of the gating mechanism. After being processed by the LSTM unit, the data is flattened and fed into a fully-connected neural network. Since the network input is a multi-dimensional PAM4 signal, the output is also set to be a multi-dimensional PAM4 signal. This enables multiple frames of signals to share the same training network, which can increase the training labels and thus enable the network to fully balance the signals. The data is divided into a training set and a test set. First, the training set is used to train the network. After the offline training is completed, it is then applied to the test set data for balancing, so as to analyze the improvement of its performance.

[0140] To verify the effectiveness of the PAM equalization transmission system provided by the present invention, the embodiments of the present invention use the equalization transmission method provided by the present invention to perform equalization transmission with the LSTM network equalizer as a variable in different transmission optical fibers, test the transmission bit error rate performance BER under different states, and draw BER curves for comparative analysis; among them, at the receiving end, it includes:

[0141] After synchronizing and matching filtering the received adjusted electrical signal using a signal synchronization unit and a matched filter, it is equalized by a feed-forward equalizer and a serial signal is output;

[0142] The serial signal is converted into a multi-dimensional parallel signal and input into a pre-trained LSTM network for equalization to obtain a multi-dimensional PAM4 signal;

[0143] The multi-dimensional PAM4 signal is demapped by a PAM demapping unit to obtain the PRBS signal to be transmitted.

[0144] Transmission bit error rate performance BER curve graphs under different states, specifically including:

[0145] Refer toFigure 6 As shown, it is a comparison chart of the measured BER curves of 30GBd PAM4 signals in back-to-back (BTB), 10-kilometer single-mode fiber (SSMF), 10-kilometer 7-core fiber transmission, and full-core transmission. By comparing the curves PAM4-30G-BTB and PAM4-30G-SSMF, it can be seen that compared with the BTB situation, about 2.5 dB of power loss is introduced in 10-kilometer SSMF transmission. PAM4-30G-7coreFiber-all represents the BER measured at the central core when all 7 cores in the 7-core fiber are used for PAM4 transmission; while PAM4-30G-7coreFiber-core1-only depicts the BER measured at the central core when only this core in the 7-core fiber is used for PAM4 transmission. As can be seen from Figure 6 it that the transmission performance of the 7-core fiber is obviously inferior to that of the SSMF. Due to the crosstalk between the cores in the 7-core fiber, about 1.3 dB of power loss is caused.

[0146] Refer to Figure 7 As shown, it is a comparison chart of the measured BER curves of applying three-dimensional input LSTM equalization in 30G-PAM4 back-to-back, 10-km single-mode fiber, and 10-km seven-core fiber transmission. The solid line and the dashed line in the figure correspond to the situations without using and using the multi-dimensional LSTM equalizer respectively. It can be seen that the situation using the multi-dimensional LSTM equalizer represented by the dashed line in the figure has a significant improvement in the transmission BER. From Figure 7 the comparison of the curves represented by PAM4-30G-SSMF-10KM and PAM4-30G-7coreFiber-10KM in it, it can be seen that the three-dimensional input LSTM equalization not only has a powerful improvement on the ISI problem of the FTN signal, but also can weaken the influence of dispersion on the signal. In this embodiment, a three-dimensional input LSTM network is used. If the input signal dimension is more, the equalization network can make full use of the signal characteristics for training to obtain an excellent correction effect. However, similarly, the complexity of the multi-dimensional input LSTM also increases a lot. Therefore, when using this algorithm in a real-time system, the selection of the input dimension needs to be more cost-effective.

[0147] Refer to Figure 8As shown, it is a comparison chart of measured BER curves under different input dimensions in the LSTM network. Here, the BTB transmission is taken as an example. Obviously, the single-input LSTM equalizer is much worse than the equalizer with multi-dimensional input, as shown by the curve of a = 1 and other curves. As the input dimension of the LSTM network increases, better BER performance can be achieved, as shown by the curves represented by a = 1, a = 3, a = 6, and a = 10. However, the complexity of the multi-dimensional input LSTM also increases significantly. In addition, for more input dimensions, the BER improvement is convergent, as shown by the curves represented by a = 3, a = 6, and a = 10. Therefore, in the work, the LSTM network with three-dimensional input is utilized.

[0148] In summary, in a severely bandwidth-limited PAM4 optical transmission system, as the input dimension of the LSTM network increases, the equalization network can make better use of signal characteristics for training to obtain excellent correction effects, and the system bit error rate performance is better.

[0149] The embodiment of the present invention proposes to jointly equalize the FFE equalizer and the multi-dimensional input LSTM equalizer, and successfully demonstrates it in the 420 Gb / s FTN-PAM4 transmission on a 10-kilometer 7-core optical fiber with a bandwidth limit of 10 GHz. Compared with the single-input LSTM equalizer, the method proposed in this embodiment can achieve an increase in receiver sensitivity of about 6.8 dB under the HD-FEC BER threshold. The present invention aims at the problem of introducing significant system impairments when using the super-Nyquist (FTN) technology or coding technology to improve the spectral efficiency in a low-cost bandwidth-limited system, and proposes a method combining a feed-forward equalizer FFE and a multi-dimensional parallel input LSTM network equalizer to overcome the limitations of traditional equalization methods, enabling the network to better learn the band-limited channel configuration and better compensate for signal impairments, and finally achieving a transmission rate of 420 Gb / s using a 7-core optical fiber on a 10-GHz bandwidth channel.

[0150] The PAM4 equalization transmission system and equalization transmission method under bandwidth limitation provided by the present invention solve the serious signal distortion problem in the bandwidth-limited PAM4 transmission system by adding an equalization unit composed of a feed-forward equalizer FFE and an LSTM network at the receiving end of the system; the feed-forward equalizer is used to adjust the amplitude and phase of the signal, pre-compensate the linear damage suffered by the regulated electrical signal received by the receiver, and generate a one-dimensional serial signal; the one-dimensional serial signal is converted into a multi-dimensional parallel signal and input into the LSTM network for equalization processing; the multi-dimensional parallel signal increases the diversity of data samples, avoids the problem of similar data in adjacent windows, and can make full use of the characteristics of the multi-dimensional parallel signal to predict and correct the current distortion situation through the output of the previous stage, so that the signal equalization is sufficient and the system bit error rate is effectively reduced. And the FFE and the LSTM network are combined and used. The FFE is responsible for initially equalizing and correcting the signal at the physical layer, and the LSTM network is used to deeply process the signal at a higher level to further reduce distortion, so that the signal distortion problem of the PAM4 transmission system is effectively solved under the condition of bandwidth limitation, the signal distortion caused by interference is effectively corrected, the quality of the received signal is significantly improved, and the bit error rate is reduced. And in the severely bandwidth-limited FTN transmission system, as the input dimension of the LSTM network increases, the LSTM network can make full use of signal characteristics for training to obtain an excellent correction effect, and the system bit error rate performance is better.

[0151] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0152] The present application is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, and the combination of flows and / or blocks in the flowchart and / or block diagram can be realized by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for realizing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0153] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to work in a particular manner, such that the instructions stored in the computer-readable memory produce a manufacture including an instruction device that implements the functions specified in one or more of the processes Figure 1 one or more processes and / or blocks Figure 1 specified in the block or blocks.

[0154] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, so that the instructions executed on the computer or other programmable apparatus provide steps for implementing the functions specified in one or more of the processes Figure 1 one or more processes and / or blocks Figure 1 specified in the block or blocks.

[0155] Obviously, the above embodiments are merely examples for clear illustration and are not limitations on the implementation manners. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to exhaustively list all the implementation manners here. And the obvious changes or modifications derived therefrom are still within the protection scope of the present invention.

Claims

1. A PAM4 equalization transmission system under bandwidth limitation, characterized in that Including: A transmitting end that maps the PRBS sequence signal to be transmitted into an initial PAM4 signal, and after digital signal processing, generates and transmits a PAM4 pulse signal; The transmitting end successively includes along the signal transmission direction: A sequence mapping unit that maps the PRBS sequence signal to be transmitted into an initial PAM4 signal for output; An FTN acceleration unit that improves the spectral efficiency of the initial PAM4 signal and outputs an optimized PAM4 signal; An upsampling unit that increases the sampling rate of the optimized PAM4 signal and outputs a smoothed PAM4 signal; A pulse shaping unit that performs pulse shaping on the smoothed PAM4 signal and outputs a PAM4 pulse signal; A transmission link that performs optoelectronic conversion on the PAM4 pulse signal to obtain a transmitted optical signal for transmission, and at the end of transmission, converts it into an adjusted electrical signal and sends it to the receiving end; A receiving end that demodulates the received adjusted electrical signal into the PRBS signal to be transmitted, completing signal transmission, including: A feed-forward equalizer that performs equalization processing on the received adjusted electrical signal after signal synchronization and matching filtering, and generates a serial signal for output; A pre-trained LSTM network that parallel inputs the serial signal converted into a multi-dimensional parallel signal, performs equalization processing, and outputs a multi-dimensional PAM4 signal; A PAM demapping unit that demaps the multi-dimensional PAM4 signal to obtain the PRBS signal to be transmitted.

2. The PAM4 equalization transmission system under bandwidth limitation according to claim 1, wherein The pulse shaping unit is an RRC roll-off filter.

3. The PAM4 equalization transmission system under bandwidth limitation according to claim 1, wherein The transmission link includes, successively connected along the signal propagation direction through an optical fiber: A waveform generator that converts the PAM4 pulse signal into an initial PAM4 electrical signal for output; An MZM modulator that modulates the initial PAM4 electrical signal into a PAM4 optical signal for output; A first variable optical attenuator that adjusts the power of the PAM4 optical signal and outputs an adjusted optical signal; A photodiode that converts the adjusted optical signal into an adjusted electrical signal for output; A real-time oscilloscope that captures and analyzes the output of the adjusted electrical signal.

4. The PAM4 equalized transmission system under bandwidth limitation according to claim 3, wherein, The transmission link further includes a noise control module, including: A second variable optical attenuator whose input end is connected to the output end of the MZM modulator through an optical fiber to adjust the power of the PAM4 optical signal; An erbium-doped fiber amplifier whose input end is connected to the output end of the second variable optical attenuator, amplifies the PAM4 optical signal whose power has been adjusted by the second variable optical attenuator, and outputs it to the first variable optical attenuator.

5. The PAM4 equalized transmission system under bandwidth limitation according to claim 1, wherein The feed-forward equalizer performs equalization processing on the input electrical signal, including: Passing the input electrical signal through multiple delay units to obtain the delayed electrical signals output by each delay unit; Multiplying each delayed electrical signal by its corresponding tap coefficient to obtain multiple multiplication results; Linearly weighting and summing the multiple multiplication results to generate a serial signal for output, expressed as: ； Among them, represents the serial signal output by the feedforward equalizer at time represents the electrical signal input to the feedforward equalizer at time represents the delayed electrical signal output by the th delay unit at time represents the th tap coefficient in the feedforward equalizer, which is iteratively updated using the recursive least squares method. , represents the total number of tap coefficients in the feedforward equalizer.

6. The PAM4 equalization transmission system under bandwidth limitation according to claim 1, wherein Using the LSTM network to perform equalization processing on the multi-dimensional parallel signal and output a multi-dimensional PAM4 signal, including: Inputting the multi-dimensional parallel signal at the current moment into the multi-dimensional input layer of the LSTM network and processing it into a feature vector; Inputting the feature vector into the LSTM layer of the LSTM network, and calculating the input gate, forget gate, output gate, and candidate state at the current moment with the external state and candidate state at the previous moment, respectively expressed as: Input gate: ; Forgotten Gate: ; Output gate: ; Candidate status: , ; External status: ; Updating the candidate state based on the forget gate and the input gate ; updating the external state based on the output gate , in the multi-dimensional output layer of the input LSTM network, converting it into a multi-dimensional PAM4 signal for output; Among them, represents the activation function; , , and respectively represent the input weights of the input gate, forget gate, output gate and candidate state; , , and respectively represent the recurrent weights of the input gate, forget gate, output gate and candidate state; , , and respectively represent the bias parameters of the input gate, forget gate, output gate and candidate state; represents the multi-dimensional parallel signal; represents the external state at the previous moment.

7. The PAM4 equalization transmission system under bandwidth limitation according to claim 1, wherein The training process of the LSTM network includes: Mapping the sequence generated based on the PRBS sequence seed number 1 to obtain multiple initial PAM4 signals; For each initial PAM4 signal, obtaining the output PAM4 signal obtained by demapping through the transmitter, transmission link, and output end; multiple groups of initial PAM4 signals and their corresponding output PAM4 signals form a training set; Presetting the number of layers and the number of neurons in the input layer and output layer of the initial LSTM network, presetting the learning rate, learning rate decay period, learning rate decay factor, and maximum number of learning times, and using the training set to train the initial LSTM network; Based on the PAM4 signal output by the output end after each initial PAM4 signal in the training set passes through the LSTM network and its corresponding initial PAM4 signal, constructing a loss function; using the Adam optimizer to optimize the LSTM network parameters until the loss function converges to obtain a pre-trained LSTM network.

8. The PAM4 equalization transmission system under bandwidth limitation according to claim 1, characterized in that, The receiving end further includes an error rate calculation unit, which compares the initial PAM4 signal and the multi-dimensional PAM4 signal, and obtains the error rate based on the quotient of the number of different bits in the multi-dimensional PAM4 signal and the initial PAM4 signal and the total number of bits of the initial PAM4 signal.

9. An equalization transmission method applied to a PAM4 equalization transmission system under bandwidth limitation as described in any one of claims 1 to 8, characterized in that, Applied to the receiving end, it includes: After synchronizing and matching filtering the received regulated electrical signal using the signal synchronization unit and the matched filter, performing equalization processing through the feed-forward equalizer to output a serial signal; Converting the serial signal into a multi-dimensional parallel signal and inputting it into the pre-trained LSTM network for equalization to obtain a multi-dimensional PAM4 signal; The multi-dimensional PAM4 signal undergoes demapping through the PAM demapping unit to obtain the PRBS signal to be transmitted.

Images