64-QAM WDM-MDM Coherent Transmission System Based on LSTM Nonlinear Equalization
Yao Zhang,Bohan Sang,Chen Wang,Jianyu Long,Kaihui Wang,Wen Zhou,Bo Liu,Jianjun Yu
DOI: https://doi.org/10.3788/cjl231505
2024-01-01
Abstract:Objective Owing to the rapid growth of Internet traffic, the demand for large transmission capacities has increased significantly. To address the issues of limited device bandwidths and high hardware- update costs, researchers have adopted expansion methods, including wavelength division multiplexing (WDM) and mode division multiplexing (MDM). Optical signal impairments caused by devices and fiber-optic- optic channels render advanced digital signal- processing technology crucial for achieving high-speed- speed fiber-optic- optic communications. Parallel with the emergence and development of deep learning, machine learning- based equalization has been widely adopted in optical communications. Deep neural networks (DNNs), convolutional neural networks (CNNs), and recurrent neural networks (RNNs) have been applied for the nonlinear compensation of fiber-optic- optic communication systems. In recent years, nonlinear equalizers (NLE) based on RNNs have been investigated extensively. Long short-term- term memory (LSTM) neural networks, as a type of RNN, can solve the issues of gradient disappearance and explosion during training. In this study, WDM, MDM, polarization multiplexing, and advanced digital- signal processing technology were used to construct a homodyne coherent transmission system based on an LSTM neural network equalizer (LSTM-NNE).- NNE). We successfully equalized 80 channels of 48 GBaud 64-quadrature- quadrature amplitude modulation (QAM) signals after transmitting 650 km of few- mode fibers (FMFs) on four modes: LP11a, LP11b, LP21a, and LP21b. The bit error rate (BER) of the MDM-WDM- WDM system can satisfy the 25% soft- decision forward- error correction (SD-FEC)- FEC) threshold of 4x10(-2). Methods At the transmitter end, 50 GHz spaced 80- channel signals distributed at 1529.9-1561.4 nm are generated. The channels are separated into two groups, where each group used 40 external cavity lasers with 100 GHz spacing (corresponding to C20-C59 and H20-H59 in the ITU standards), and the optical carriers are coupled by two polarization- maintaining arrayed waveguide gratings (PM-AWGs).- AWGs). The coupled carriers are modulated using two independent IQ modulators. The modulators are driven by a four- channel, 64 GSa/s, arbitrary- waveform generator (AWG) programmed to generate 48 GBaud, 64-QAM- QAM signals that are pulse shaped using a 1 % root- raised- cosine (RRC) filter. The transmitted symbols are generated using true random integers obtained via atmospheric noise such that overfitting caused by a pseudo- random binary sequence (PRBS) can be avoided. Two polarization multiplexing emulators (PMEs) are used to generate the PDM signals. The optical signal entering the PME is first split into two beams by a 1x2 coupler, decorrelated through optical fibers of different lengths, and then combined by a polarization beam combiner. The two 40- wavelength PDM signals are coupled and amplified using an erbium- doped fiber amplifier (EDFA), after which the coupled signal is split and delayed to generate four tributaries with a relative delay of 100 ns and then fed into four parallel path- length- aligned recirculating loops. Before adding the 100 ns relative delay, we fine-tuned- tuned the initial delay fiber length such that the mode sync peaks in frame synchronization appeared at the same location, thereby allowing the total delay spread to be reduced. The single- mode outputs of the four loops are amplified to 23 dBm and multiplexed using a mode multiplexer. The obtained four- mode signals are launched into a 50- km- long span of the FMF. The output of the FMF is demultiplexed, and the signals are amplified using four EDFAs. Four wavelength- selective switches operating within the C band with 50 GHz channel spacing are used to add independent attenuation for each channel to compensate for the uneven gain of the EDFA. At the receiver end, the outputs of the four loops passed through four wavelength- division demultiplexers (DWDMs). The output signals are received by four coherent receivers (CRs) for homodyne reception. We used a real-time- time digital storage oscilloscope (DSO) to capture the baseband electrical signal and perform offline DSP, which comprised resampling, CD compensation, frequency- offset compensation, frame synchronization, and the proposed LSTM-NNE,- NNE, which is first initialized in the data- aided mode and then switched to the decision- directed mode for BER calculation. Results and Discussions Figure 4 shows the BER of the conventional MIMO-LMS- LMS (multi- input multi- output least mean square), NNE, and the LSTM-NNE- NNE proposed herein for different transmission distances. The BER yielded by the LSTM-NNE- NNE is approximately 0.01 lower than that yielded by the NNE, whereas it is even lower compared with that yielded by the conventional MIMO-LMS- LMS algorithm. Meanwhile, the LSTM-NNE- NNE enables the BER of the 64-QAM- QAM to be transmitted 650 km lower than the 4.0x10(-2) SD-FEC- FEC threshold. As shown in Figs. 5 and 6, the difference in BER of each mode under different wavelength subchannels is insignificant. Furthermore, the BER of each channel in the 650 km transmission is lower than the 4.0x10(-2) SD-FEC- FEC threshold. Conclusions In this study, we constructed a four- mode 80- wavelength dual- polarization homodyne coherent transmission system. At the receiving end, an LSTM neural network equalizer was used for channel equalization. When transmitting 650 km, the net rate of the transmission system reached 147.4 Tbit/s. By utilizing the LSTM-NNE,- NNE, the BER of the MDM-WDM- WDM system can satisfy the 25 % SD-FEC- FEC threshold of 4x10(-2). The experimental results confirmed the nonlinear equalization potential of the MIMO neural network equalizer for future high- capacity long-distance- distance transmission systems.