Nonlinear equalization method and system for ultra-high order QAM metro optical interconnection

Abstract

The present application relates to the field of optical fiber communication technology, and more particularly to a nonlinear equalization method and system for ultra-high order QAM metropolitan area optical interconnection. The complex signal of two orthogonal polarization channels at each moment in the preprocessed transmission optical signal is extracted; the real part and the imaginary part of the complex signal of two orthogonal polarization channels at each moment are decomposed; the power of two orthogonal polarization channels at each moment is calculated based on the real part and the imaginary part of two orthogonal polarization channels at each moment; the input feature vector at each moment is constructed with the real part, the imaginary part and the power of the orthogonal polarization channel at each moment; the target complex signal of two orthogonal polarization channels at each moment is obtained based on the input feature vector at each moment, and the equalized symbols corresponding to all moments are obtained. The present application reduces the calculation complexity of the neural network equalizer under the premise that the bit error rate of 1024QAM signal is lower than the forward error correction threshold after 80 kilometers of optical fiber transmission.

Description

Technical Field

[0001] This invention relates to the field of optical fiber communication technology, and in particular to a nonlinear equalization method and system for ultra-high-order QAM metropolitan optical interconnection. Background Technology

[0002] With the large-scale deployment of global cloud computing and artificial intelligence computing power clusters, data traffic between distributed data centers is growing exponentially. This trend has created an urgent need for expansion of short- to medium-distance (typically 80 km) optical interconnect links connecting data centers within metropolitan area networks. To meet the evolution of future 800G and 1.6T Ethernet interfaces, the adoption of ultra-high-order orthogonal amplitude modulation (such as 1024QAM) to maximize spectral efficiency has become an industry consensus. However, extending high-sensitivity modulation formats such as 1024QAM from short-distance on-chip / rack applications to 80 km standard single-mode fiber (SSMF) links with significant channel impairments faces severe technical challenges.

[0003] In these high-spectrum-efficiency, medium-distance metropolitan fiber optic transmission systems, signal quality is mainly limited by fiber Kerr nonlinearity, laser phase noise, and imperfections in the transceiver front end. To address the combined impairments caused by the superposition of these fiber nonlinearities and phase noise, various compensation and equalization techniques have been proposed. The overall technological evolution path shows an iterative upgrade trend from traditional linear / nonlinear equalization algorithms to neural network intelligent equalization algorithms.

[0004] In the early stages, channel impairment equalization was mainly achieved using traditional digital signal processing schemes, with core technologies represented by the Digital Back Propagation (DBP) algorithm and linear adaptive equalizers (such as the Constant Modulus Algorithm, CMA). These traditional algorithms rely on classical digital signal processing theory for their framework, have high technical maturity, and simple hardware implementation logic. However, they have inherent defects when dealing with ultra-high-order modulation formats such as 1024QAM: due to the extremely short Euclidean distance of the constellation diagram of ultra-high-order modulation signals, the signal's sensitivity to various noises, especially residual nonlinear noise, is greatly increased. Traditional digital back propagation algorithms can only achieve partial nonlinear impairment compensation, and linear filtering algorithms cannot cope with nonlinear channel distortion. When used alone or in combination, the residual bit error rate of the system can never be reduced below the 20% SD-FEC error correction threshold, which directly leads to strict limitations on the transmission distance and system capacity of 80-kilometer mid-range links, making it difficult to adapt to the actual needs of high-speed transmission of 800G and above.

[0005] To address the shortcomings of traditional equalization techniques, variants of recurrent neural networks, particularly Long Short-Term Memory (LSTM) networks, have been proposed for application in the post-processing of coherent optical receivers. The core working principle of LSTM equalizers lies in their internal memory units and gating mechanisms, effectively capturing the long-range memory effects caused by dispersion and nonlinear coupling in fiber optic channels. This approach trains the LSTM network to learn a nonlinear inverse mapping from the received symbol sequence to the original transmitted symbols, thus significantly outperforming traditional digital backpropagation or linear adaptive equalizers in transmission links dominated by nonlinear noise.

[0006] While recurrent neural network equalization techniques such as LSTM exhibit excellent compensation performance in long-distance transmission scenarios, their inherent structure has unavoidable core shortcomings. The complex recurrent network structure and multi-gating unit design result in extremely high computational complexity and overhead, significantly increasing the difficulty of hardware implementation. Furthermore, 80km metropolitan area DCI transceivers are typical latency-sensitive devices with stringent power consumption budgets. The high complexity of LSTM equalizers not only leads to excessive signal processing latency but also drives up the manufacturing cost and power consumption of hardware chips, completely failing to meet the stringent requirements of low power consumption, low latency, and high engineering feasibility for metropolitan area DCI systems, making large-scale application difficult. Summary of the Invention

[0007] Therefore, the technical problem to be solved by this invention is how to reduce the computational complexity of the neural network equalizer so as to meet the requirements of low cost and low power consumption of metropolitan area DCI, while ensuring that the bit error rate of the 1024QAM signal is lower than the forward error correction threshold after transmission over 80 kilometers of optical fiber.

[0008] To address the aforementioned technical problems, this invention provides a nonlinear equalization method for ultra-high-order QAM metropolitan optical interconnects, comprising: The transmitted optical signal output from the ultra-high-order QAM coherent optical transmission system is preprocessed to obtain the preprocessed transmitted optical signal; wherein, each moment in the transmitted optical signal corresponds to a symbol; Extract the complex signals of the two orthogonal polarization channels at each moment from the preprocessed transmitted optical signal; The real and imaginary parts of the complex signals of the two orthogonal polarization channels at each time moment are decomposed to obtain the real and imaginary components of the two orthogonal polarization channels at each time moment; Based on the real and imaginary components of the two orthogonal polarization channels at each time step, calculate the power of the two orthogonal polarization channels at each time step. The input feature vector for each time step is constructed using the real components, imaginary components, and power in the orthogonal polarization channel at each time step. Based on the input feature vectors at each time step, the target complex signal of the two orthogonal polarization channels at each time step is obtained, and the equalized symbol corresponding to all time steps is obtained.

[0009] Preferably, the method for preprocessing the transmitted optical signal output from an ultra-high-order QAM coherent optical transmission system includes: The transmitted optical signal output from the ultra-high-order QAM coherent optical transmission system is sequentially subjected to dispersion compensation, clock recovery, and resampling to obtain the preprocessed transmitted optical signal.

[0010] Preferably, the formula for calculating the power of the two orthogonal polarization channels at each moment, based on the real and imaginary components of the two orthogonal polarization channels at each moment, is as follows: , , in, The power of the X-polarization channel. The power of the Y-polarized channel. For the real component of the X-polarization channel. This is the imaginary component of the X-polarization channel. For the real component of the Y-polarization channel, This is the imaginary component of the Y-polarization channel.

[0011] Preferably, the method for obtaining the target complex signal of the two orthogonal polarization channels at each time step based on the input feature vector at each time step includes: Based on the input feature vectors at each time step, the target complex signal of two orthogonal polarization channels at each time step is obtained by using a feature extraction backbone network and a multi-symbol output layer.

[0012] Preferably, the feature extraction backbone network includes multiple fully connected hidden layers connected in sequence.

[0013] Preferably, the ReLU activation function of the fully connected hidden layer is replaced with the GELU activation function to obtain the target fully connected hidden layer.

[0014] Preferably, the number of neurons in multiple sequentially connected fully connected hidden layers decreases at each level.

[0015] Preferably, the method for obtaining the target complex signal of two orthogonal polarization channels at each time step based on the input feature vectors at each time step, using a feature extraction backbone network and a multi-symbol output layer, includes: The sliding window length is set to 2L+1, and the sliding step size is fixed at K. There is a gap between two adjacent sliding windows. The overlapping area of ​​the symbols; where L is a set value; Starting from the beginning of the time series, the input feature vectors of all times within the first sliding window are concatenated and integrated in chronological order to obtain the input features of the first sliding window; The input features of the first sliding window are sequentially passed through a feature extraction backbone network and a multi-symbol output layer to obtain the target complex signal of two orthogonal polarization channels at K time points within the first sliding window; where... ; After obtaining the target complex signals of the two orthogonal polarization channels at K times within a single sliding window, the sliding window is shifted backward along the time series according to the set fixed sliding step size K, and the target complex signals of the two orthogonal polarization channels at K times within each sliding window are obtained in turn until all times in the time series are traversed. By integrating the target complex signals of the two orthogonal polarization channels at K time points within all sliding windows, the target complex signals of the two orthogonal polarization channels at each time point are obtained.

[0016] Preferably, the multi-symbol output layer is a linear fully connected layer.

[0017] This invention also provides a nonlinear equalization system for ultra-high-order QAM metropolitan optical interconnects, comprising: The preprocessing module is used to preprocess the transmitted optical signal output by the ultra-high-order QAM coherent optical transmission system to obtain the preprocessed transmitted optical signal; wherein, each moment in the transmitted optical signal corresponds to a symbol; The extraction module is used to extract the complex signals of the two orthogonal polarization channels at each moment in the preprocessed transmitted optical signal; The component acquisition module is used to decompose the complex signals of the two orthogonal polarization channels at each time step into real and imaginary components, and obtain the real and imaginary components of the two orthogonal polarization channels at each time step. The power acquisition module is used to calculate the power of the two orthogonal polarization channels at each time step based on the real and imaginary components of the two orthogonal polarization channels at each time step. The input feature vector construction module is used to construct the input feature vector at each time step using the real components, imaginary components, and power in the orthogonal polarization channel at each time step. The recovery module is used to obtain the target complex signal of the two orthogonal polarization channels at each time step based on the input feature vector at each time step, and obtain the equalized symbol corresponding to all time steps.

[0018] Compared with the prior art, the above-described technical solution of the present invention has the following advantages: This invention discloses a nonlinear equalization method and system for ultra-high-order QAM metropolitan optical interconnection. It extracts independent complex signals from two orthogonal polarization channels at each moment, fully preserving the complete signal information of the dual polarization channels in coherent optical transmission and avoiding feature loss caused by crosstalk between polarization channels. Based on this, the complex signals of each channel are decomposed into real and imaginary parts, obtaining independent real and imaginary components of the dual polarization channels at each moment. This fully preserves the basic phase and amplitude characteristics of the signal, achieving precise decomposition of the core modulation information of the signal. Furthermore, based on the real and imaginary components, the signal power of the dual polarization channels at the corresponding moment is calculated synchronously, introducing an additional amplitude-correlated power feature dimension. Finally, the real and imaginary components of the orthogonal polarization channels at each moment are fused with the power features to construct an enhanced input feature vector for a dedicated optical fiber channel. Through this systematic feature construction method, explicit and precise capture of physical damage in optical fiber transmission is achieved. Furthermore, by using a feature extraction backbone network and a multi-symbol output layer, it is possible to simultaneously output equalized symbols corresponding to multiple time points. This effectively reduces the computational complexity of the neural network equalizer while ensuring that the bit error rate of the 1024QAM signal is lower than the forward error correction threshold after transmission over 80 kilometers of optical fiber. Attached Figure Description

[0019] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein:

[0020] Figure 1 This is a flowchart illustrating a nonlinear equalization method for ultra-high-order QAM metropolitan optical interconnection according to the present invention.

[0021] Figure 2 This is a schematic diagram of the internal structure of the FA-MSO-NN equalizer.

[0022] Figure 3 This is the complexity-performance trade-off curve of the FA-MSO-NN of this invention.

[0023] Figure 4 The curves show the bit error rate of FA-MSO-NN as a function of input optical power under different parallelization factors.

[0024] Figure 5 This is a flowchart of digital signal processing (DSP) for a coherent optical receiver that includes FA-MSO-NN. Detailed Implementation

[0025] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0026] To reduce the computational overhead of highly complex deep learning models, a multi-symbol output (MSO) strategy has been disclosed in the prior art. In terahertz wireless communication systems, this scheme modifies the output layer structure of a neural network, enabling a single forward propagation process to simultaneously output multiple consecutive equalized symbols. This "one-time computation, multiple-symbol output" architecture effectively amortizes the computational complexity required per symbol.

[0027] However, existing MSO technology is mainly used in wireless channels, and its direct migration to fiber optic channels has shortcomings. First, the traditional MSO architecture does not feature-optimize for the amplitude-dependent nonlinearity impairments of fiber optic links. During 80-kilometer transmission, the accumulated nonlinearity and phase rotation are severe, and existing MSO networks struggle to explicitly capture this physical mechanism. Second, existing complex-valued neural network schemes bundle the I / Q signals for processing, assuming the modulator operates in the ideal linear region. However, when driving ultra-high-order 1024QAM signals, the electrical amplifier and lithium niobate modulator operate near high bias points, resulting in severe I / Q independent nonlinear distortion. The complex-valued network lacks the freedom to independently adjust the I and Q branches, leading to insufficient modeling accuracy and increased residual bit error rate.

[0028] Therefore, this application aims to overcome the shortcomings of the prior art, and significantly reduce the computational complexity of the neural network equalizer while ensuring that the bit error rate of the 1024QAM signal is lower than the forward error correction threshold after transmission over 80 kilometers of optical fiber, so as to meet the requirements of low cost and low power consumption for metro DCI. It captures and compensates for amplitude-dependent nonlinear impairments in the metro transmission link and I / Q asymmetric nonlinear distortion introduced by the transceiver, thereby improving equalization accuracy.

[0029] Reference Figure 1 As shown, this embodiment provides a nonlinear equalization method for ultra-high-order QAM metropolitan optical interconnects, including: Step S1: Preprocess the transmitted optical signal output by the ultra-high-order QAM coherent optical transmission system to obtain the preprocessed transmitted optical signal; wherein, each moment in the transmitted optical signal corresponds to a symbol; In this embodiment, specifically, the method for preprocessing the transmitted optical signal output by the ultra-high-order QAM coherent optical transmission system includes: The transmitted optical signal output from the ultra-high-order QAM coherent optical transmission system is sequentially subjected to dispersion compensation, clock recovery, and resampling to obtain the preprocessed transmitted optical signal.

[0030] To explicitly capture amplitude-dependent nonlinear phase noise caused by the optical Kerr effect in optical fiber transmission, this invention adds a deterministic feature enhancement unit before the input layer, namely steps S2-S5; the deterministic feature enhancement unit has no training parameters and only performs fixed quadratic operations.

[0031] Step S2: Extract the complex signals of the two orthogonal polarization channels at each moment from the preprocessed transmitted optical signal; Step S3: Solve the complex signals of the two orthogonal polarization channels at each time step by solving for the real part (I) and the imaginary part (Q) to obtain the real and imaginary components of the two orthogonal polarization channels at each time step. , , , ,in, for The real component of the X-polarization channel at time t. for The imaginary component of the X-polarization channel at time t. for The real component of the Y-polarization channel at time t. for The imaginary component of the Y-polarization channel at time; Step S4: Calculate the power of the two orthogonal polarization channels at each time step based on the real and imaginary components of the two orthogonal polarization channels at each time step; The formula for calculating the power of the two orthogonal polarization channels at each moment, based on the real and imaginary components of the two orthogonal polarization channels at each moment, is as follows: , , in, The power of the X-polarization channel. The power of the Y-polarized channel. For the real component of the X-polarization channel. This is the imaginary component of the X-polarization channel. For the real component of the Y-polarization channel, This is the imaginary component of the Y-polarization channel.

[0032] Step S5: Construct the input feature vector for each time step using the real components, imaginary components, and power in the orthogonal polarization channel. ; Input feature vector at each time step It not only includes the raw I / Q sample values, but also explicitly stitches together the instantaneous power characteristics. , , The specific definitions are as follows: .

[0033] in, for The input feature vector at time step 1. The design of the input feature vector allows the neural network to acquire physical quantities directly related to nonlinear phase rotation, avoiding the need for hidden layers to consume additional neuron resources for fitting square operations. Explicitly encoding the amplitude-phase nonlinear coupling mechanism dominated by the optical Kerr effect in fiber optic transmission into the neural network input space eliminates the need for the network to expend neuron resources on "implicitly fitting" square operations through hidden layers, thereby significantly accelerating convergence speed (reducing training epochs by approximately 40%), improving data efficiency, and enhancing model generalization ability. This feature enhancement layer can be used as a standalone, plug-and-play preprocessing module, adaptable to various real-valued neural network equilibrium architectures.

[0034] Step S6: Based on the input feature vectors at each time step, obtain the target complex signal of the two orthogonal polarization channels at each time step, and obtain the equalized symbol corresponding to all time steps.

[0035] In this embodiment, specifically, the method for obtaining the target complex signal of the two orthogonal polarization channels at each time step based on the input feature vector at each time step includes: Based on the input feature vectors at each time step, the target complex signal of two orthogonal polarization channels at each time step is obtained by using a feature extraction backbone network and a multi-symbol output (MSO) layer.

[0036] In this embodiment, specifically, the feature extraction backbone network includes multiple sequentially connected fully connected hidden layers to perform nonlinear feature compression and reconstruction.

[0037] In this embodiment, preferably, the ReLU activation function of the fully connected hidden layer is replaced with the GELU activation function to obtain the target fully connected hidden layer. Each fully connected layer in the target fully connected hidden layer is followed by a batch normalization (BN) layer and a Gaussian Error Linear Unit (GELU) activation function. The BN layer normalizes the input of each layer, eliminating internal covariate bias, accelerating network convergence, and allowing for a higher learning rate. The GELU activation function retains non-zero gradients in the negative interval, avoiding the neuron death phenomenon of ReLU, and its smooth characteristics are more suitable for fitting continuous nonlinear functions of fiber optic channels.

[0038] In this embodiment, preferably, the number of neurons in the multiple sequentially connected fully connected hidden layers decreases step by step.

[0039] To balance nonlinear fitting capability and computational cost, this embodiment preferably employs a dual-hidden-layer structure, that is, using two sequentially connected fully connected hidden layers as the feature extraction backbone network. The number of neurons in the first fully connected hidden layer is... The number of neurons in the second fully connected hidden layer is The hidden layer output is progressively reduced in dimensionality, ultimately compressing the high-dimensional features into a compact nonlinear compensated signal representation. In this embodiment, preferably, the method for obtaining the target complex signal of two orthogonal polarization channels at each time step based on the input feature vector at each time step, using a feature extraction backbone network and a multi-symbol output layer, includes: The sliding window length is set to 2L+1, and the sliding step size is fixed at K. There is a gap between two adjacent sliding windows. The overlapping area of ​​the symbols; where L is a set value; Starting from the beginning of the time series, the input feature vectors of all times within the first sliding window are concatenated and integrated in chronological order to obtain the input features of the first sliding window; the dimension of the input features of each sliding window is 6×(2L+1) (i.e., three channels I, Q, and P × 2 polarizations × window length), where P is the power channel; The input features of the first sliding window are sequentially passed through a feature extraction backbone network and a multi-symbol output layer to obtain the target complex signal of two orthogonal polarization channels at K time points within the first sliding window; where... ; After obtaining the target complex signals of the two orthogonal polarization channels at K times within a single sliding window, the sliding window is shifted backward along the time series according to the set fixed sliding step size K, and the target complex signals of the two orthogonal polarization channels at K times within each sliding window are obtained in turn until all times in the time series are traversed. By integrating the target complex signals of the two orthogonal polarization channels at K time points within all sliding windows, the target complex signals of the two orthogonal polarization channels at each time point are obtained.

[0040] This invention proposes a sliding window batch inference deployment method co-designed with a multi-symbol output architecture. Its key feature is that, during the real-time inference phase, the receiver DSP slides an input window across the symbol sequence with a fixed step size K, and adjacent windows are spaced apart. The network performs a forward computation on each window input, outputting K balanced symbols in parallel. The windows are then slid together to form a complete output sequence. The technical advantage of this deployment method is that it avoids the frequent context loading / storage overhead of symbol-by-symbol inference, fully releasing the hardware's parallel computing capabilities. The measured throughput can reach K times that of the symbol-by-symbol inference mode. This method is a key guarantee for ensuring that the theoretical complexity advantage of MSO is truly realized on a practical hardware platform.

[0041] In this embodiment, specifically, the multi-symbol output layer is a linear fully connected layer with 4K neurons (where K≥2), corresponding to two polarizations and four paths of real and imaginary parts, outputting K consecutive equalized symbols at once. During the training phase, the loss function is jointly optimized based on the mean square error (MSE) of these K symbols and the corresponding transmitted symbol labels. During the inference phase, the network slides the received symbol sequence with a step size of K to achieve low-overhead real-time equalization.

[0042] This invention uses Real Multiplications Per Symbol (RMPS) as a complexity metric. For the aforementioned double hidden layer structure, the RMPS calculation formula is: , in, For the input feature dimension, This represents the number of neurons in the first hidden layer. This represents the number of neurons in the second hidden layer. For the output layer dimension, the first term represents the computational cost of the backbone network (from the input layer to the second hidden layer), amortized by K symbols; the second term represents the computational cost of the output layer, which is independent of K. As K increases, the average cost per symbol of the backbone network decreases significantly, thus achieving the best trade-off between complexity and performance.

[0043] like Figure 2 As shown, Figure 2 This is a schematic diagram of the internal structure of the Feature Enhancement Multi-Symbol Output Neural Network Equalizer (FA-MSO-NN).

[0044] The FA-MSO-NN equalizer consists of the aforementioned deterministic feature enhancement unit, feature extraction backbone network, and multi-symbol output layer. The connection relationship and data flow of each layer are clearly shown in the figure.

[0045] To verify the practical effectiveness of the feature-enhanced multi-symbol output neural network equalizer proposed in this invention, an experimental platform for transmission over 80 kilometers of standard single-mode fiber (SSMF) using 20-Gbaud dual-polarization 1024QAM (DP-1024QAM) was constructed. The experiment evaluated the invention from two dimensions: computational complexity reduction capability and nonlinear impairment compensation performance. The experimental results fully demonstrate the significant advancements of the technical solution presented in this invention.

[0046] like Figure 3 As shown, Figure 3 The figure shows the complexity-performance trade-off curve of the FA-MSO-NN of this invention. The horizontal axis represents the number of parallel output symbols K (i.e., the number of parallel output symbols in the multi-symbol output layer), the left vertical axis represents the number of actual multiplications per symbol (RMPS), and the right vertical axis represents the bit error rate (BER). The input optical power was fixed at the optimal value of -2dBm in the experiment.

[0047] The computational complexity is inversely proportional to K. This is because the multi-symbol output strategy adopted in this invention successfully amortizes the computational overhead of the backbone network onto the precise characterization of K output symbols.

[0048] The BER degrades slightly with increasing K. For all configurations where K ≤ 6, the BER remains consistently below the 20% soft-decision forward error correction (SD-FEC) threshold (2.4 × 10⁻²). This demonstrates that the present invention, through its ingenious architectural design, achieves an order-of-magnitude reduction in complexity with minimal performance cost, perfectly meeting the core requirements of metro DCI for a balanced solution of low power consumption, low cost, and high performance.

[0049] like Figure 4 As shown, Figure 4 The figure shows the bit error rate (BER) of the FA-MSO-NN as a function of input optical power under different parallelization factors. Experimental results demonstrate that the FA-MSO-NN of this invention achieves the optimal trade-off between complexity and performance. Within a wide dynamic range of -6dBm to +3dBm, the BER at all operating points is below the FEC threshold.

[0050] like Figure 5 As shown, Figure 5 This is a flowchart of digital signal processing (DSP) for a coherent optical receiver that includes FA-MSO-NN.

[0051] This embodiment verifies the effectiveness of the invention by constructing a 20-GBaud probabilistic shaping (PS) 1024QAM seven-core fiber optic transmission system. The experimental setup is as follows: Figure 2As shown. At the transmitting end, an external cavity laser (ECL) with a linewidth of 1 kHz and a center wavelength of 1550.112 nm is used as the light source. In the digital signal processing (DSP) section, a binary pseudo-random bit sequence (PRBS) is generated and mapped to PS-1024QAM symbols using a probabilistic shaping technique with an entropy of 8.5 bits / symbol. Subsequently, pulse shaping is performed using a root raised cosine (RRC) filter with a roll-off factor of 0.05 to limit the signal bandwidth. The generated digital baseband signal is loaded into an arbitrary waveform generator (AWG) with a sampling rate of 64 GSa / s to generate four analog electrical signals to drive the I / Q modulator. The modulated optical signal is pre-amplified by an erbium-doped fiber amplifier (EDFA) and its power is adjusted by a variable optical attenuator (VOA). It is then split into seven decorrelation signals by a 1×7 optical splitter and coupled into a 10 km long weakly coupled seven-core fiber (MCF) for transmission via a fan-in device. The inter-core crosstalk of this fiber is less than -45 dB, ensuring that the inter-core interference mainly exhibits linear characteristics. At the receiving end, the signal is demultiplexed by a fan-out device, then the optical signal-to-noise ratio (OSNR) is adjusted by the receiving end EDFA and VOA. It then enters a 90-degree optical mixer for mixing with a local oscillator, and is photodetected by four balanced photodetectors (BPDs). The local oscillator is derived from the same source as the transmitting laser. The output analog electrical signal is acquired and digitized by a 50 GSa / s real-time oscilloscope. In the offline DSP processing flow, the signal undergoes 2x resampling, RRC matched filtering, and clock recovery based on the Gardner algorithm. After preprocessing, the signal enters a 4×4 real-valued MIMO-LMS equalizer for linear polarization demultiplexing and residual dispersion compensation. The filter tap length is set to 41, and the convergence step size is... The equalized output is then processed sequentially based on frequency offset estimation and compensation, followed by carrier phase retrieval (CPE). At this point, the signal is restored to a dual-polarization complex baseband form. The processed signals are then combined into the aforementioned vector, which is sequentially fed into two fully connected hidden layers: the first hidden layer contains 1024 neurons, and the second contains 512 neurons. Each layer is followed by a cascaded batch normalization (BN) layer and a Gaussian error linear unit (GELU) activation function to accelerate convergence and improve nonlinear fitting capability. The output layer is a linear fully connected layer employing a multi-symbol output (MSO) strategy, outputting equalized symbols for K consecutive time steps in parallel. This amortizes the computational overhead of the backbone network across K symbols, reducing the number of real multiplications per symbol (RMPS) by approximately 80% compared to the single-symbol output scheme.

[0052] This invention proposes a low-complexity neural network equalizer architecture specifically designed for 80km-level metropolitan data center interconnection scenarios and adapted to ultra-high-order QAM (typically 1024QAM) coherent optical transmission systems. It employs a pure real-valued computation base, explicitly decomposing the received complex signal into independent in-phase (I) and quadrature (Q) components. A deterministic feature enhancement unit is added between the input and hidden layers to explicitly inject instantaneous power features strongly correlated with the physical mechanism of fiber Kerr nonlinearity. A multi-symbol parallel output (MSO) topology is used at the output layer, estimating K consecutive equalization symbols in parallel during a single forward propagation. Through a three-pronged collaborative design of "real-valued computation + physical feature injection + computational amortization," this architecture achieves a breakthrough for the first time in 1024QAM 80km transmission experiments, achieving a bit error rate below the 20% SD-FEC threshold and a computational complexity reduction of 80% compared to the single-symbol benchmark. This solves the core bottleneck of the inability to engineer deep learning models in power-sensitive metropolitan optical modules.

[0053] This second embodiment provides a nonlinear equalization system for ultra-high-order QAM metropolitan optical interconnects, including: The preprocessing module is used to preprocess the transmitted optical signal output by the ultra-high-order QAM coherent optical transmission system to obtain the preprocessed transmitted optical signal; wherein, each moment in the transmitted optical signal corresponds to a symbol; The extraction module is used to extract the complex signals of the two orthogonal polarization channels at each moment in the preprocessed transmitted optical signal; The component acquisition module is used to decompose the complex signals of the two orthogonal polarization channels at each time step into real and imaginary components, and obtain the real and imaginary components of the two orthogonal polarization channels at each time step. The power acquisition module is used to calculate the power of the two orthogonal polarization channels at each time step based on the real and imaginary components of the two orthogonal polarization channels at each time step. The input feature vector construction module is used to construct the input feature vector at each time step using the real components, imaginary components, and power in the orthogonal polarization channel at each time step. The recovery module is used to obtain the target complex signal of the two orthogonal polarization channels at each time step based on the input feature vector at each time step, and obtain the equalized symbol corresponding to all time steps.

[0054] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied 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.

[0055] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0056] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0057] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0058] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A nonlinear equalization method for ultra-high-order QAM metropolitan optical interconnects, characterized in that, include: The transmitted optical signal output from the ultra-high-order QAM coherent optical transmission system is preprocessed to obtain the preprocessed transmitted optical signal; wherein, each moment in the transmitted optical signal corresponds to a symbol; Extract the complex signals of the two orthogonal polarization channels at each moment from the preprocessed transmitted optical signal; The real and imaginary parts of the complex signals of the two orthogonal polarization channels at each time moment are decomposed to obtain the real and imaginary components of the two orthogonal polarization channels at each time moment; Based on the real and imaginary components of the two orthogonal polarization channels at each time step, calculate the power of the two orthogonal polarization channels at each time step. The input feature vector for each time step is constructed using the real components, imaginary components, and power in the orthogonal polarization channel at each time step. Based on the input feature vectors at each time step, the target complex signal of the two orthogonal polarization channels at each time step is obtained, and the equalized symbol corresponding to all time steps is obtained.

2. The nonlinear equalization method for ultra-high-order QAM metropolitan optical interconnects according to claim 1, characterized in that, Methods for preprocessing the transmitted optical signal output from an ultra-high-order QAM coherent optical transmission system include: The transmitted optical signal output from the ultra-high-order QAM coherent optical transmission system is sequentially subjected to dispersion compensation, clock recovery, and resampling to obtain the preprocessed transmitted optical signal.

3. The nonlinear equalization method for ultra-high-order QAM metropolitan optical interconnects according to claim 1, characterized in that, The formula for calculating the power of the two orthogonal polarization channels at each moment, based on the real and imaginary components of the two orthogonal polarization channels at each moment, is as follows: ， ， in, The power of the X-polarization channel. The power of the Y-polarized channel. For the real component of the X-polarization channel. This is the imaginary component of the X-polarization channel. For the real component of the Y-polarization channel, This is the imaginary component of the Y-polarization channel.

4. The nonlinear equalization method for ultra-high-order QAM metropolitan optical interconnects according to claim 1, characterized in that, The methods for obtaining the target complex signal of two orthogonal polarization channels at each time step based on the input feature vector at each time step include: Based on the input feature vectors at each time step, the target complex signal of two orthogonal polarization channels at each time step is obtained by using a feature extraction backbone network and a multi-symbol output layer.

5. A nonlinear equalization method for ultra-high-order QAM metropolitan optical interconnects according to claim 1, characterized in that, The feature extraction backbone network consists of multiple fully connected hidden layers connected in sequence.

6. The nonlinear equalization method for ultra-high-order QAM metropolitan optical interconnection according to claim 5, wherein the ReLU activation function of the fully connected hidden layer is replaced with the GELU activation function to obtain the target fully connected hidden layer.

7. A nonlinear equalization method for ultra-high-order QAM metropolitan optical interconnects according to claim 5, characterized in that, The number of neurons in multiple sequentially connected fully connected hidden layers decreases at each level.

8. A nonlinear equalization method for ultra-high-order QAM metropolitan optical interconnects according to claim 5, characterized in that, The method for obtaining the target complex signal of two orthogonal polarization channels at each time step based on the input feature vectors at each time step, using a feature extraction backbone network and a multi-symbol output layer, includes: The sliding window length is set to 2L+1, and the sliding step size is fixed at K. There is a gap between two adjacent sliding windows. The overlapping area of ​​the symbols; where L is a set value; Starting from the beginning of the time series, the input feature vectors of all times within the first sliding window are concatenated and integrated in chronological order to obtain the input features of the first sliding window; The input features of the first sliding window are sequentially passed through a feature extraction backbone network and a multi-symbol output layer to obtain the target complex signal of two orthogonal polarization channels at K time points within the first sliding window; where... ; After obtaining the target complex signals of the two orthogonal polarization channels at K times within a single sliding window, the sliding window is shifted backward along the time series according to the set fixed sliding step size K, and the target complex signals of the two orthogonal polarization channels at K times within each sliding window are obtained in turn until all times in the time series are traversed. By integrating the target complex signals of the two orthogonal polarization channels at K time points within all sliding windows, the target complex signals of the two orthogonal polarization channels at each time point are obtained.

9. A nonlinear equalization method for ultra-high-order QAM metropolitan optical interconnects according to claim 5, characterized in that, The multi-symbol output layer is a linear fully connected layer.

10. A nonlinear equalization system for ultra-high-order QAM metropolitan optical interconnects, characterized in that, include: The preprocessing module is used to preprocess the transmitted optical signal output by the ultra-high-order QAM coherent optical transmission system to obtain the preprocessed transmitted optical signal; wherein, each moment in the transmitted optical signal corresponds to a symbol; The extraction module is used to extract the complex signals of the two orthogonal polarization channels at each moment in the preprocessed transmitted optical signal; The component acquisition module is used to decompose the complex signals of the two orthogonal polarization channels at each time step into real and imaginary components, and obtain the real and imaginary components of the two orthogonal polarization channels at each time step. The power acquisition module is used to calculate the power of the two orthogonal polarization channels at each time step based on the real and imaginary components of the two orthogonal polarization channels at each time step. The input feature vector construction module is used to construct the input feature vector at each time step using the real components, imaginary components, and power in the orthogonal polarization channel at each time step. The recovery module is used to obtain the target complex signal of the two orthogonal polarization channels at each time step based on the input feature vector at each time step, and obtain the equalized symbol corresponding to all time steps.

Images