Signal compensation method and device in optical communication, equipment and medium

Abstract

The invention discloses a signal compensation method and device in optical communication, equipment and a medium, relates to the technical field of network communication, is applied to a receiving end, and constructs a training set and a verification set comprising a plurality of sample pairs for a current optical communication scene; the sample pair is constructed based on a first historical digital signal acquired by the receiving end and a second historical digital signal needing to be sent by the sending end; respectively utilizing the training set and the verification set to alternately update the calculation weight of a candidate compensation operator and the selection weight of the candidate compensation operator of each architecture search unit in an initial equalizer so as to obtain a target equalizer; and obtaining a to-be-compensated digital signal of an optical signal currently received by the receiving end, and performing signal compensation on the to-be-compensated digital signal by using the target equalizer to obtain a target compensated signal. The generalization capability and the precision of a nonlinear equalization scheme in signal compensation in ultra-high-speed optical communication are improved, and the cost is reduced.

Description

Technical Field

[0001] This invention relates to the field of network communication technology, and in particular to signal compensation methods, devices, equipment and media in optical communication. Background Technology

[0002] In ultra-high-speed optical communication systems, as transmission rates increase to 800Gb / s and 1.6Tb / s, nonlinear impairments such as Kerr nonlinearity, phase noise, and dispersion coupling generated during fiber transmission become increasingly prominent. These impairments severely degrade optical signal quality, leading to signal distortion, increased bit error rate, and limiting system transmission performance and reliability. Traditional linear compensation methods in digital signal processing (DSP) can only compensate for simple linear impairments and cannot effectively suppress complex nonlinear impairments, making it difficult to meet the transmission requirements of ultra-high-speed optical communication.

[0003] While current neural network-based nonlinear equalization schemes can compensate for nonlinear damage to some extent, they have significant drawbacks. Their network structures largely rely on manual design, and the selection of candidate operators and parameter tuning require substantial manual labor. Furthermore, they cannot adaptively match optical communication scenarios with different transmission distances and modulation formats, resulting in weak generalization capabilities. At the same time, their network structures are redundant and computationally complex, making it difficult to adapt to the resource limitations of high-speed DSP and other real-time processing platforms. They cannot achieve a balance between compensation accuracy and engineering practicality, thus restricting their deployment and application in actual ultra-high-speed optical communication receivers.

[0004] In summary, improving the generalization ability and accuracy of nonlinear equalization schemes while reducing costs in signal compensation in ultra-high-speed optical communication is a problem that needs to be solved in this field. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a signal compensation method, apparatus, device, and medium in optical communication, which improves the generalization ability and accuracy of nonlinear equalization schemes and reduces costs in signal compensation in ultra-high-speed optical communication. The specific solution is as follows: In a first aspect, this application discloses a signal compensation method in optical communication, applied at the receiving end, comprising: A training set and a validation set, comprising multiple sample pairs, are constructed for the current optical communication scenario; the sample pairs are constructed based on the first historical digital signal acquired by the receiver and the second historical digital signal to be transmitted by the transmitter. The computational weights and selection weights of the candidate compensation operators of each architecture search unit in the initial equalizer are alternately updated using the training set and the validation set, respectively, to obtain the target equalizer; The digital signal to be compensated of the optical signal currently received by the receiving end is obtained, and the target equalizer is used to perform signal compensation on the digital signal to be compensated in order to obtain the target compensated signal.

[0006] Optionally, the construction of a training set and a validation set comprising multiple sample pairs for the current optical communication scenario includes: The original historical optical signal received by the receiving end in the current optical communication scenario is collected, and the original historical optical signal is preprocessed to obtain the preprocessed historical optical signal. The first historical digital signal of the preprocessed historical optical signal and the second historical digital signal to be transmitted by the transmitting end are obtained, and multiple sample pairs are constructed based on the first historical digital signal and the second historical digital signal. The sample pairs are divided to obtain training and validation sets.

[0007] Optionally, acquiring the first historical digital signal of the preprocessed historical optical signal and the second historical digital signal to be transmitted by the transmitting end includes: I / Q sequences are extracted from the preprocessed historical optical signals, and time windows of a preset window size are slid within the I / Q sequences to extract the first sub-sequence groups under different time windows. The first four-way windowed I / Q subsequences in the first subsequence group are concatenated to obtain the first historical digital signal in real vector form; Based on a preset window size, the original historical digital signal to be transmitted by the transmitting end is synchronously framed with the first sub-sequence group to obtain each second sub-sequence group, and the second sub-sequence group is determined as the second historical digital signal; wherein, the second sub-sequence group includes the second four-channel windowed I / Q sub-sequence.

[0008] Optionally, before alternately updating the computational weights and selection weights of the candidate compensation operators for each architecture search unit in the initial equalizer using the training set and the validation set respectively, the method further includes: An initial equalizer is constructed, comprising multiple reusable architecture search units; wherein the architecture search units include multiple candidate compensation operators, which are used to perform compensation operations on digital signals.

[0009] Optionally, the step of alternately updating the computational weights and selection weights of the candidate compensation operators of each architecture search unit in the initial equalizer using the training set and the validation set, respectively, to obtain the target equalizer, includes: The initial equalizer is determined as the current equalizer. The selection weights of the candidate compensation operators of each architecture search unit in the current equalizer are fixed, and the calculation weights of the candidate compensation operators of each architecture search unit in the current equalizer are updated using the training set to obtain the first updated equalizer. The computational weights of the candidate compensation operators of each architecture search unit in the first updated equalizer are fixed, and the selection weights of the candidate compensation operators of each architecture search unit in the first updated equalizer are updated using the validation set to obtain the second updated equalizer. The second updated equalizer is determined as the new current equalizer, and the process jumps back to the step of fixing the selection weights of the candidate compensation operators of each architecture search unit in the current equalizer until the preset stopping condition is met, and the output second updated equalizer is determined as the target equalizer.

[0010] Optionally, determining the second updated equalizer as the target equalizer includes: Each architecture search unit in the second updated equalizer output is sequentially determined as the current architecture search unit; The selection weights of each candidate compensation operator in the current architecture search unit are sorted, and the candidate compensation operators of the current architecture search unit are lightweighted according to the sorted selection weights to obtain the current target architecture search unit. The target equalizer is obtained based on the search units of each target architecture.

[0011] Optionally, updating the computational weights of the candidate compensation operators for each architecture search unit in the current equalizer using the training set includes: The first historical digital signal in the training set is input into each architecture search unit of the current equalizer to obtain the output results of each candidate compensation operator in each architecture search unit. A normalized exponential function is applied to the computational weights of each candidate compensation operator to obtain the target weights of each candidate compensation operator. The target weights are then used to perform a weighted summation of the output results in each architecture search unit to obtain the weighted result of each architecture search unit. The computational weights of the candidate compensation operators of each architecture search unit in the current equalizer are updated based on the loss function value between the second historical digital signal in the training set and the weighted result.

[0012] Secondly, this application discloses a signal compensation device for optical communication, applied at the receiving end, comprising: The dataset construction module is used to construct a training set and a validation set including multiple sample pairs for the current optical communication scenario; the sample pairs are constructed based on the first historical digital signal obtained by the receiver and the second historical digital signal to be transmitted by the transmitter. The equalizer update module is used to alternately update the computation weights and selection weights of the candidate compensation operators of each architecture search unit in the initial equalizer using the training set and the validation set, respectively, in order to obtain the target equalizer. The signal compensation module is used to acquire the digital signal to be compensated of the optical signal currently received by the receiver, and to use the target equalizer to perform signal compensation on the digital signal to be compensated in order to obtain the target compensated signal.

[0013] Thirdly, this application discloses an electronic device, comprising: Memory, used to store computer programs; A processor is configured to execute the computer program to implement the steps of the aforementioned disclosed signal compensation method in optical communication.

[0014] Fourthly, this application discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the steps of the aforementioned disclosed signal compensation method in optical communication.

[0015] The beneficial effects of this application are as follows: This application is applied to the receiving end, and constructs a training set and a validation set including multiple sample pairs for the current optical communication scenario; the sample pairs are constructed based on the first historical digital signal obtained by the receiving end and the second historical digital signal to be transmitted by the transmitting end; the calculation weights and selection weights of the candidate compensation operators of each architecture search unit in the initial equalizer are alternately updated using the training set and the validation set to obtain the target equalizer; the digital signal to be compensated of the optical signal currently received by the receiving end is obtained, and the target equalizer is used to perform signal compensation on the digital signal to be compensated to obtain the target compensated signal. Therefore, this application constructs sample pairs based on the first historical digital signal from the receiver and the second historical digital signal from the transmitter, and uses the training set and validation set to alternately update the computational weights and selection weights of the candidate compensation operators in each architecture search unit of the initial equalizer to obtain the target equalizer. This can improve the signal compensation accuracy by optimizing the computational weights within the candidate compensation operators, while adaptively learning and determining the selection weights of each candidate compensation operator, achieving joint optimization of the equalizer structure and parameters. This avoids the problems of poor adaptability and limited compensation effect caused by relying on manual experience to design the network structure. The training set ensures the model's fitting and compensation ability for the damaged signal, while the validation set ensures the model's generalization performance, preventing overfitting and improving adaptability to different optical communication scenarios. Then, the target equalizer is used to compensate the digital signal to be compensated. This enables adaptive determination of the optimal equalizer structure and parameters for different optical communication scenarios, improving the accuracy and effect of optical signal nonlinear damage compensation, simplifying the equalizer structure design and manual debugging process, reducing model complexity and computational overhead, making the equalizer more adaptable to the real-time operation requirements of high-speed DSP and other hardware platforms, improving the reliability and stability of signal recovery at the receiver of the optical communication system, and improving signal transmission quality. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0017] Figure 1 This is a flowchart of a signal compensation method in optical communication disclosed in this application; Figure 2 This is a schematic diagram of a specific optical communication system disclosed in this application; Figure 3 This is a schematic diagram of a specific equalizer architecture search principle disclosed in this application; Figure 4This is a schematic diagram illustrating the test results of a specific optical transmission system disclosed in this application; Figure 5 This is a schematic diagram of the structure of a signal compensation device in optical communication disclosed in this application; Figure 6 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] In ultra-high-speed optical communication systems, as transmission rates increase to 800Gb / s and 1.6Tb / s, nonlinear impairments such as Kerr nonlinearity, phase noise, and dispersion coupling generated during fiber transmission become increasingly prominent. These impairments severely degrade optical signal quality, leading to signal distortion, increased bit error rate, and limiting system transmission performance and reliability. Traditional linear compensation methods in digital signal processing (DSP) can only compensate for simple linear impairments and cannot effectively suppress complex nonlinear impairments, making it difficult to meet the transmission requirements of ultra-high-speed optical communication.

[0020] While current neural network-based nonlinear equalization schemes can compensate for nonlinear damage to some extent, they have significant drawbacks. Their network structures largely rely on manual design, and the selection of candidate operators and parameter tuning require substantial manual labor. Furthermore, they cannot adaptively match optical communication scenarios with different transmission distances and modulation formats, resulting in weak generalization capabilities. At the same time, their network structures are redundant and computationally complex, making it difficult to adapt to the resource limitations of high-speed DSP and other real-time processing platforms. They cannot achieve a balance between compensation accuracy and engineering practicality, thus restricting their deployment and application in actual ultra-high-speed optical communication receivers.

[0021] Therefore, this application provides a signal compensation scheme in optical communication, which improves the generalization ability and accuracy of nonlinear equalization schemes and reduces costs in signal compensation in ultra-high-speed optical communication.

[0022] See Figure 1 As shown in the figure, this application discloses a signal compensation method in optical communication, applied at the receiving end, including: Step S11: Construct a training set and a validation set including multiple sample pairs for the current optical communication scenario; the sample pairs are constructed based on the first historical digital signal obtained by the receiver and the second historical digital signal to be transmitted by the transmitter.

[0023] In this embodiment, the construction of a training set and a validation set comprising multiple sample pairs for the current optical communication scenario includes: collecting the original historical optical signal received by the receiving end in the current optical communication scenario; preprocessing the original historical optical signal to obtain a preprocessed historical optical signal; acquiring a first historical digital signal of the preprocessed historical optical signal and a second historical digital signal to be transmitted by the transmitting end; and constructing multiple sample pairs based on the first historical digital signal and the second historical digital signal; and dividing each sample pair to obtain a training set and a validation set.

[0024] When constructing training and validation sets including multiple sample pairs for current optical communication scenarios, the first step is to collect the original historical optical signals received by the receiver in the current optical communication scenario. These original historical optical signals, after processing by traditional DSP stages such as photoelectric detection, analog-to-digital conversion, dispersion compensation, frequency offset correction, multi-path multi-output adaptive equalization, and carrier phase recovery, still contain nonlinear impairments such as Kerr nonlinearity, phase noise, and dispersion coupling. The original historical optical signals undergo preprocessing operations consistent with those at the actual optical communication receiver, including signal conditioning, noise filtering, and format normalization, to obtain preprocessed historical optical signals that eliminate invalid interference and meet the input format requirements of the neural network. Then, through signal demodulation and digitization, the corresponding signals of the preprocessed historical optical signals are obtained. The first historical digital signal reflects the actual reception status of the receiver. At the same time, the original, lossless second historical digital signal that the transmitter actually needs to send is extracted from the signal transmission record of the transmitter and corresponds one-to-one with the first historical digital signal. Each pair of corresponding first and second historical digital signals is paired. Multiple sample pairs are constructed based on multiple pairs of such corresponding signals. In each sample pair, the first historical digital signal is used as the input sample for the neural network architecture search, and the second historical digital signal is used as the corresponding label sample. Then, all the constructed sample pairs are divided. Most of the sample pairs can be divided into training sets, and the remaining sample pairs can be divided into validation sets. During the division process, it is ensured that both the training set and the validation set can cover the typical link nonlinear characteristics and noise statistics in the current optical communication scenario.

[0025] In the first specific sample pair partitioning embodiment, different sample pairs can be divided into training set and validation set based on a preset ratio. That is, there are no identical sample pairs in the training set and validation set. The training set is used to train the super network functional parameters θ, and the validation set is used to optimize the super network architecture parameters α. In this way, it is ensured that the network can fully fit the signal characteristics of the current optical communication scenario during the two-layer optimization process, and also has good generalization ability on unseen data, adapting to the actual link conditions of the current optical communication scenario.

[0026] In the second specific sample pair partitioning embodiment, each sample pair is adaptively partitioned based on the real-time channel status of the current optical communication link, the characteristic distribution of sample data, and the fitting effect and generalization performance during model training to obtain a training set and a validation set. That is, an adaptive data selection mechanism is introduced for dynamic partitioning. This mechanism can adaptively adjust the sample composition, quantity ratio, and sample selection criteria of the training set and validation set according to the real-time channel status of the current optical communication link, the characteristic distribution of sample data, and the fitting effect and generalization performance during model training. Sample pairs that can characterize the typical nonlinear characteristics, noise statistics, and channel mutation characteristics of the current link are preferentially included in the training set. At the same time, sample pairs that cover different working states of the link, different channel quality levels, and have not participated in model parameter optimization are selected to form the validation set.

[0027] In the third specific sample pair partitioning embodiment, each sample pair is partitioned based on the collected physical layer performance indicators of the optical communication link to obtain a training set and a validation set. That is, performance feedback based on channel quality is added during the partitioning process. The physical layer performance indicators of the optical communication link are collected in real time and used as an important basis for sample partitioning. This ensures that the partitioned training set and validation set can match the physical layer performance requirements. This allows the subsequent architecture search process to not only optimize the signal symbol-level loss, but also to carry out multi-objective optimization by combining physical layer performance indicators unique to optical communication, such as bit error rate and Q factor. This ensures that the partitioned training set and validation set can support the two-layer optimization mechanism in optimizing network functional parameters and architecture parameters. This ensures that the model's ability to fit the symbol-level signal is guaranteed, and that the searched network architecture adapts to the actual physical layer performance requirements of the optical communication link, thereby improving the robustness and adaptability of the architecture in actual deployment.

[0028] In this embodiment, acquiring the first historical digital signal of the preprocessed historical optical signal and the second historical digital signal to be transmitted by the transmitter includes: extracting I / Q sequences from the preprocessed historical optical signal; controlling a time window of a preset window size to slide within the I / Q sequences to extract first sub-sequence groups under different time windows; splicing the first four-way windowed I / Q sub-sequences in the first sub-sequence group to obtain the first historical digital signal in real vector form; synchronizing and framing the original historical digital signal to be transmitted by the transmitter with the first sub-sequence group based on the preset window size to obtain each second sub-sequence group, and determining the second sub-sequence group as the second historical digital signal; wherein, the second sub-sequence group includes the second four-way windowed I / Q sub-sequences.

[0029] I / Q sequences representing the phase and amplitude information of optical signals are extracted from historical optical signals that have been preprocessed to eliminate invalid interference and conform to the input format requirements of neural networks. A pre-set window size is set according to the Kerr nonlinearity effect of the optical communication link and the temporal neighborhood correlation characteristics caused by phase noise. This fixed-size time window is controlled to slide symbol-by-symbol across the continuous I / Q sequences to extract first sub-sequence groups containing adjacent symbol features under different time windows. These first sub-sequence groups contain windowed I / Q sub-sequences of the real and imaginary parts of four signals corresponding to the two polarization directions of the optical signal, i.e., the first four-way windowed I / Q sub-sequences. The first four-way windowed I / Q sub-sequences in the first sub-sequence group are then dimensionally concatenated to transform them into high-dimensional feature vectors in real vector form. These real vectors can fully represent the nonlinear coupling relationship between symbols in the optical fiber link and are used as the first historical digital signal for neural network architecture search and training. Simultaneously, according to the same preset window size as the first sub-sequence group, the original, lossless historical digital signal transmitted by the transmitter is subjected to frame processing that is completely synchronized with the time axis of the first sub-sequence group. This ensures that the framed sub-sequences correspond one-to-one with the first sub-sequence group at the receiver in the time dimension, resulting in each second sub-sequence group. This second sub-sequence group also contains the second four-way windowed I / Q sub-sequences corresponding to the real and imaginary parts of the two polarization directions. The second sub-sequence group is directly identified as the tag-class second historical digital signal that matches the first historical digital signal, achieving precise alignment between the input features at the receiver and the original tags at the transmitter in the windowing dimension. This lays the data foundation for subsequent sample pair construction and two-layer optimization training.

[0030] In other words, this embodiment uses a sliding window to extract features from the symbol sequence to construct a multidimensional input vector suitable for nonlinear modeling. Specifically, a time-domain window of length Nin is constructed based on the received I / Q sequence, and the real and imaginary parts of the two polarizations are concatenated to form a 4Nin-dimensional real vector, which is used as the input to the neural network architecture search equalizer. This window can cover a certain range of adjacent symbols to utilize the temporal neighborhood correlation generated by Kerr nonlinearity and phase noise in the optical fiber link to achieve cross-symbol nonlinear coupling modeling. It does not rely on fixed, manually designed pre-stage feature engineering, has high data-driven characteristics, and is strictly matched with the coherent signal structure of the optical communication link.

[0031] Step S12: Alternately update the computation weights and selection weights of the candidate compensation operators of each architecture search unit in the initial equalizer using the training set and the validation set respectively, in order to obtain the target equalizer.

[0032] In this embodiment, before alternately updating the computation weights and selection weights of the candidate compensation operators of each architecture search unit in the initial equalizer using the training set and the validation set respectively, the method further includes: constructing an initial equalizer comprising multiple reusable architecture search units; wherein, the architecture search unit comprises multiple candidate compensation operators, which are used to perform compensation operations on digital signals.

[0033] Construct an initial equalizer and embed it directly into the nonlinear compensation module before decision-making at the optical communication receiver, such as... Figure 2 As shown, compensation processing is performed on digital signals that still have residual nonlinear impairments after traditional DSP processing, thereby achieving high-performance symbol recovery. The mapping relationship between the input X and output Y of the overall network structure can be expressed as: ; Where θ represents all trainable weights in the supernetwork, i.e., the computational weights of the candidate compensation operators, and α represents the architecture parameters, i.e., the selection weights of the candidate compensation operators. The input X is constructed using a data-driven window-style feature model, enabling the equalizer to learn the nonlinear coupling relationships between symbols caused by the Kerr effect, residual phase noise, etc.

[0034] like Figure 3 As shown, an initial equalizer is constructed, comprising multiple reusable architecture search units. Each architecture search unit includes multiple candidate compensation operators, which are used to perform compensation operations on digital signals. Specifically, based on the actual needs of nonlinear compensation in optical communication systems, a super network formed by several reusable architecture search units (Cells) is built as the initial equalizer. This super network is constructed to adapt to the strong physical constraints and high temporal correlation characteristics of optical communication links, and has the ability to cover multiple nonlinear modeling paths. Each architecture search unit serves as the core basic module constituting the initial equalizer, internally integrating multiple preset candidate compensation operators, defined as op( ).

[0035] Furthermore, the candidate compensation operators configured in the super network are designed with the nonlinear impairment compensation requirements of optical communication links as the core. The basic operators include feedforward neural network fully connected layer operators of different sizes (FNN-128, FNN-256, FNN-512), one-dimensional convolutional layer operators with different numbers of channels (CNN-8, CNN-16, CNN-64), and identity mapping operators used to realize cross-layer information transmission and improve feature reuse efficiency. At the same time, the candidate compensation operators support flexible expansion and replacement. For optical communication systems with more complex channel effects, more rich convolutional operators with different kernel sizes, depthwise separable convolutional operators, residual block operators, and attention mechanism unit operators can be dynamically introduced. Lightweight operators such as low-rank fully connected layers and channel-wise convolution can also be introduced according to hardware resource constraints. All kinds of candidate compensation operators can complete the corresponding digital signal compensation operation for the nonlinear characteristics of different optical communication links, adapting to different compensation requirements and hardware deployment conditions.

[0036] As a core component of the super network, the Cell is a reusable architecture search unit, composed of various candidate compensation operators. The super network as a whole is formed by combining multiple such Cells in series. At the same time, the configuration parameters of the Cell can be flexibly adjusted according to the actual scenario of the optical communication system. The number of Cells and the connection method between Cells can be adaptively changed according to the system's transmission rate, modulation format, or spectrum utilization method. The sliding window length of the optical signal input can also be adjusted. By adjusting the Cell-related configurations, the overall scale and model capacity of the super network can be changed, thereby forming an architecture search space that adapts to different optical communication scenarios and covers different nonlinear modeling paths, meeting the balanced architecture search requirements under different link conditions.

[0037] In this embodiment, the step of alternately updating the computational weights and selection weights of the candidate compensation operators of each architecture search unit in the initial equalizer using the training set and the validation set to obtain the target equalizer includes: determining the initial equalizer as the current equalizer; fixing the selection weights of the candidate compensation operators of each architecture search unit in the current equalizer; updating the computational weights of the candidate compensation operators of each architecture search unit in the current equalizer using the training set to obtain a first updated equalizer; fixing the computational weights of the candidate compensation operators of each architecture search unit in the first updated equalizer; updating the selection weights of the candidate compensation operators of each architecture search unit in the first updated equalizer using the validation set to obtain a second updated equalizer; determining the second updated equalizer as the new current equalizer; and jumping back to the step of fixing the selection weights of the candidate compensation operators of each architecture search unit in the current equalizer until a preset stopping condition is met; and determining the output second updated equalizer as the target equalizer.

[0038] The super network formed by multiple reusable architecture search units is determined as the initial equalizer and set as the current equalizer. First, the selection weights of the candidate compensation operators in each architecture search unit in the current equalizer, i.e., the architecture parameters α, are fixed. Based on the real data training set of the optical communication link, the mean square error between the received symbol and the transmitted symbol is minimized as the optimization objective. The trainable computation weights of all candidate compensation operators in each architecture search unit in the current equalizer, i.e., the network function parameters θ, are updated through gradient backpropagation. The internal optimization of the network parameters is completed and the first updated equalizer is obtained.

[0039] Then, fix the updated computational weights θ of the candidate compensation operators in each architecture search unit of the first updated equalizer. Based on the validation set that did not participate in parameter training, with the optimization goal of improving the network's generalization ability on unseen data, update the selection weights α of the candidate compensation operators in each architecture search unit of the first updated equalizer to achieve external optimization of architecture parameters and obtain the second updated equalizer.

[0040] Subsequently, the second updated equalizer is determined as the new current equalizer, and the process jumps back to the step of fixing the selection weights of candidate compensation operators in each architecture search unit of the current equalizer. The alternating update process of calculating and selecting weights of candidate compensation operators is repeated. This process is an iterative execution of the Bilevel Optimization (BO) mechanism. During the iteration, the performance indicators of the optical communication link are always the core optimization target, allowing the equalizer to continuously adapt to match the nonlinear characteristics and noise statistics of the optical fiber link until a preset stopping condition is met. The preset stopping condition includes any one or more combinations of the following: the number of iterations reaches a set threshold, the change in architecture parameters and network parameters is less than a preset threshold, and the optical communication physical layer performance indicators, namely bit error rate and Q factor, on the validation set reach a preset optimal standard. When the preset stopping condition is met, the iteration stops, and the final output second updated equalizer is determined as the target equalizer adapted to the characteristics of the current optical communication link. The selection weights of candidate compensation operators in each architecture search unit of the target equalizer have shown a significant bias, which can accurately match the nonlinear compensation requirements of the current optical communication scenario, and has good fitting ability and generalization performance. It can be directly used for subsequent structural discretization processing and high-speed DSP platform deployment.

[0041] In other words, this embodiment proposes a two-layer optimization mechanism to simultaneously optimize the network's functional parameter θ and structural parameter α. First, α is fixed on the training set, and θ is updated to minimize the mean square error between received and transmitted symbols. Then, θ is fixed on the validation set, and α is updated to give the corresponding network structure stronger generalization ability on unseen data. By alternately optimizing the training and validation datasets, it is possible to ensure that the searched network architecture has both high fitting ability and avoids overfitting to specific link conditions, maintaining robustness under different fiber parameters and different operating points. Compared with the traditional manual parameter tuning process, this two-layer optimization method fully automates the architecture selection process, not only reducing the subjectivity and inefficient iteration of manual design and improving model development efficiency, but also systematically exploring different architecture combinations in the search space. Finally, after the search process converges, the Softmax weights of each Cell will show a significant bias. By selecting the candidate operation with the largest corresponding weight to discretize the Cell, the super network is transformed into a clear, compact final network architecture suitable for practical DSP deployment.

[0042] In this embodiment, determining the output second updated equalizer as the target equalizer includes: sequentially determining each architecture search unit in the output second updated equalizer as the current architecture search unit; sorting the selection weights of each candidate compensation operator of the current architecture search unit, and performing lightweight processing on each candidate compensation operator of the current architecture search unit according to the sorted selection weights to obtain the current target architecture search unit; and obtaining the target equalizer based on each target architecture search unit.

[0043] Each architecture search unit in the output of the second updated equalizer is sequentially determined as the current architecture search unit. This second updated equalizer is a super network adapted to the nonlinear characteristics of the current optical communication link after iterative optimization through a two-layer optimization mechanism. The candidate compensation operators in each architecture search unit have formed selection weights with obvious biases. The selection weights of each candidate compensation operator in the current architecture search unit are sorted in descending order. Taking into account the actual deployment requirements of the optical communication system's high-speed DSP, real-time platform hardware computing power, latency, etc., the candidate compensation operators in the current architecture search unit are subjected to targeted lightweight processing based on the sorted selection weights.

[0044] In a specific embodiment of targeted lightweighting of candidate compensation operators of the current architecture search unit based on the sorted selection weights, a preset number of candidate compensation operators corresponding to the sorted selection weights are selected as target compensation operators. Operators other than the target compensation operators are removed from the candidate compensation operators, retaining only the target compensation operators. It can be understood that the number of target compensation operators can be multiple or a single one. That is, for candidate compensation operators with high selection weights, a single optimal operator or multiple high-weight candidate compensation operators can be retained to form a hybrid structure, depending on hardware resources. For candidate compensation operators with low selection weights, automatic pruning and sparsification operations are directly performed to remove them from the current architecture search unit. Simultaneously, quantization and distillation operations are performed on the retained target compensation operators and their corresponding network parameters to further compress the network size and reduce computational complexity. This ensures that the processed architecture search unit fully matches the operating requirements of the hardware platform while maintaining nonlinear compensation performance. The current target architecture search unit is obtained after the above lightweighting process.

[0045] After completing the lightweighting of all architecture search units in the second updated equalizer in the above manner, each target architecture search unit is obtained. All target architecture search units are then combined in the same way as the original super network. Based on the combined target architecture search units, a target equalizer that can be directly deployed on an actual optical communication hardware platform is obtained. This target equalizer is a compact and computationally efficient discrete network architecture that retains the optimal nonlinear compensation capability to adapt to link characteristics and meets the engineering operation requirements of high-speed DSP and real-time platforms.

[0046] In this embodiment, updating the computational weights of candidate compensation operators for each architecture search unit in the current equalizer using the training set includes: inputting the first historical digital signal from the training set into each architecture search unit of the current equalizer to obtain the output results of each candidate compensation operator in each architecture search unit; applying a normalized exponential function to the computational weights of each candidate compensation operator to obtain the target weights of each candidate compensation operator, and using the target weights to perform a weighted summation of the output results in each architecture search unit to obtain the weighted result of each architecture search unit; and updating the computational weights of candidate compensation operators for each architecture search unit in the current equalizer based on the loss function value between the second historical digital signal from the training set and the weighted result.

[0047] The first historical digital signal, in real vector form extracted and spliced ​​from the preprocessed historical optical signals in the training set, is input batch by batch into each architecture search unit of the current equalizer. Each candidate compensation operator within each architecture search unit performs a corresponding nonlinear compensation calculation operation on the input first historical digital signal, thereby obtaining the independent signal compensation output result of each candidate compensation operator in each architecture search unit. It is important to note that during the forward process, the Cell does not directly select a single operator, but rather performs a weighted sum of the outputs of all candidate operations, with the weights... The algorithm is obtained by applying the Softmax function to α. In other words, the initial computational weights of each candidate compensation operator are processed by applying the Softmax normalized exponential function, mapping the computational weights to target weights of each candidate compensation operator in the form of a probability distribution. These target weights reflect the suitability of each candidate compensation operator under the current link conditions. Then, these target weights are used to perform a weighted summation operation on the signal compensation results output by each candidate compensation operator in each architecture search unit, resulting in a weighted result of the multi-operator compensation effect for each architecture search unit. This result is the overall output of the architecture search unit. Specifically, the output h of the i-th cell can be represented as a weighted combination of a set of candidate operation results to ensure that the architecture can be optimized during gradient backpropagation. By stacking several cells, the super network forms a huge search space that can cover multiple possible structures. The task of the structure search process is to find the optimal-performing and practically deployable equilibrium network within this space. The computation method within a single cell is as follows: ; In the formula, i represents the i-th architecture search unit, and j represents the j-th candidate compensation operator. denoted by , h represents the weighted result, and N represents the total number of candidate compensation operators.

[0048] Using the second historical digital signal in the training set that precisely matches the first historical digital signal as the true label, the loss function value between the true label and the weighted result of each architecture search unit is calculated. This loss function value characterizes the degree of deviation between the current equalizer output and the actual transmitted signal. Then, based on the gradient backpropagation algorithm, with the goal of minimizing this loss function value, the trainable computational weights of the candidate compensation operators of each architecture search unit in the current equalizer are iteratively updated and optimized. This allows the computational weights of the candidate compensation operators to adapt to the nonlinear characteristics of the current optical communication link, thereby improving the nonlinear compensation fitting capability of the received digital signal.

[0049] like Figure 4As shown, the candidate compensation operators included in each architecture search unit are Skip (identity mapping operator), CNN_8 / CNN_16 / CNN_64 (8 / 16 / 64-channel one-dimensional convolutional layer operator), and FNN_128 / FNN_256 / FNN_512 (128 / 256 / 512-scale feedforward neural network fully connected layer operator). According to experimental results, in a polarization-multiplexed coherent optical transmission system experiment with a transmission distance of 960km and a rate of 800Gb / s, the NASEE method proposed in this embodiment autonomously selects a three-layer network structure formed by combining CNN-8, CNN-16, and FNN-256 (e.g., ...). Figure 4 As shown in (a), the architecture selection weight value corresponding to this combined structure is the largest. The convolutional layer operator is mainly responsible for extracting the local temporal correlation features of the optical signal, while the fully connected layer operator can achieve stronger global nonlinear modeling. Together, they complete the accurate compensation for nonlinear damage to the fiber optic link. This combined structure cannot be directly obtained through manual experience design, but the NASEE method can autonomously mine and determine the optimal structure scheme from a massive network structure combination space. The equalization schemes include W / o NLC (no nonlinear compensation), W / FNN (using a feedforward neural network equalization scheme), w / CNN (using a convolutional neural network equalization scheme), and W / NAS (using a neural network architecture search scheme). After structural discretization, this network architecture not only has lower computational complexity but also achieves a Q-factor improvement of approximately 0.3dB in this communication link (e.g., ...). Figure 4 As shown in (b), w / NAS is the system performance curve obtained by adopting the NASEE scheme of this embodiment, and its performance is significantly better than that of manually designed equalizer architecture. The above experimental results fully verify that the method of this embodiment can not only automatically search for the optimal network structure adapted to the characteristics of optical fiber links, but also achieve better nonlinear compensation performance under the premise of controlling computational complexity to be basically the same or even lower. In summary, the NASEE method proposed in this embodiment, by constructing a unified cell-level super network, adopting a differentiable structure search optimization mechanism, and combining the input feature modeling method of sliding window and architecture discretization strategy, creates an automated, low-complexity, and high-performance equalizer design process. It effectively solves the technical problems of low efficiency, poor cross-system mobility, and difficulty in balancing complexity and performance in the existing technology of manually designed network architecture, and provides a new architectural implementation path for nonlinear compensation of future ultra-high-speed, ultra-long-distance, and multi-channel optical fiber communication systems.

[0050] Step S13: Obtain the digital signal to be compensated of the optical signal currently received by the receiver, and use the target equalizer to perform signal compensation on the digital signal to be compensated to obtain the target compensated signal.

[0051] The optical signal currently received in real time by the receiver of the optical communication system is acquired. First, traditional DSP processing operations such as photoelectric detection, analog-to-digital conversion, dispersion compensation, frequency offset correction, multi-channel multi-output adaptive equalization, and carrier phase recovery are sequentially performed on this optical signal. This results in a digital signal that still contains residual nonlinear damage such as Kerr nonlinearity, phase noise, and dispersion coupling. This digital signal is used as the digital signal to be compensated. Then, following the same processing method as when constructing the training set, I / Q sequences are extracted from this digital signal to be compensated. A time window of a preset size slides symbol-by-symbol within the I / Q sequence, extracting windowed subsequences containing features of adjacent symbols. The four windowed I / Q subsequences corresponding to the real and imaginary parts of the two polarization directions are concatenated to transform the signal into a signal compatible with the target signal. The equalizer input dimension is a real vector form that matches the digital signal to be compensated. This real vector form of the digital signal to be compensated is input into the target equalizer, which is obtained after two-layer optimization iteration and structural discretization processing. The target equalizer works collaboratively with the candidate compensation operators selected in each target architecture search unit. The convolutional layer operator extracts the local time correlation features of the digital signal to be compensated, and the fully connected layer operator realizes global nonlinear modeling. Each operator completes the accurate compensation calculation for residual nonlinear damage according to the predetermined architecture of the target equalizer. Finally, the compensated digital signal is output and identified as the target compensated signal. This signal effectively eliminates the nonlinear distortion caused by the fiber optic link, realizes high-performance symbol recovery, and can be directly used in the subsequent decision-making process to ensure the transmission performance of the optical communication system.

[0052] This embodiment can be extended to more optical communication application scenarios, covering various types such as direct-modulation and direct-detection short-range interconnect systems, multi-core / multi-mode space-division multiplexing systems, ultra-wideband wavelength-division multiplexing systems, and free-space optical communication systems. For different optical communication systems, the signal structure characteristics of each scenario can be adapted by adjusting the input data dimension, changing the complex number representation, optimizing the feature splicing method, or adjusting the multi-channel sampling strategy. This allows the equalizer in this embodiment to autonomously complete the search and construction of the optimal network architecture under various complex signal structure conditions, ensuring the nonlinear compensation effect in different scenarios. Furthermore, the technical solution of this embodiment can also be cascaded with physical model-driven compensation methods such as digital backpropagation (DBP). By searching the network structure, the optimal combination of the target equalizer and the physical model-driven compensation method in this embodiment is automatically determined, constructing a hybrid compensation scheme that coordinates data-driven and physical-driven optimization. Based on this optimal combination, the target equalizer is used to compensate the digital signal to be compensated to obtain the pre-compensated signal. Then, physical model-driven compensation is performed on the pre-compensated signal to obtain the target compensated signal, further enriching the engineering application forms of the technology and greatly expanding the engineering application scope of this embodiment in the field of optical communication.

[0053] The beneficial effects of this application are as follows: This application is applied to the receiving end, and constructs a training set and a validation set including multiple sample pairs for the current optical communication scenario; the sample pairs are constructed based on the first historical digital signal obtained by the receiving end and the second historical digital signal to be transmitted by the transmitting end; the calculation weights and selection weights of the candidate compensation operators of each architecture search unit in the initial equalizer are alternately updated using the training set and the validation set to obtain the target equalizer; the digital signal to be compensated of the optical signal currently received by the receiving end is obtained, and the target equalizer is used to perform signal compensation on the digital signal to be compensated to obtain the target compensated signal. Therefore, this application constructs sample pairs based on the first historical digital signal from the receiver and the second historical digital signal from the transmitter, and uses the training set and validation set to alternately update the computational weights and selection weights of the candidate compensation operators in each architecture search unit of the initial equalizer to obtain the target equalizer. This can improve the signal compensation accuracy by optimizing the computational weights within the candidate compensation operators, while adaptively learning and determining the selection weights of each candidate compensation operator, achieving joint optimization of the equalizer structure and parameters. This avoids the problems of poor adaptability and limited compensation effect caused by relying on manual experience to design the network structure. The training set ensures the model's fitting and compensation ability for the damaged signal, while the validation set ensures the model's generalization performance, preventing overfitting and improving adaptability to different optical communication scenarios. Then, the target equalizer is used to compensate the digital signal to be compensated. This enables adaptive determination of the optimal equalizer structure and parameters for different optical communication scenarios, improving the accuracy and effect of optical signal nonlinear damage compensation, simplifying the equalizer structure design and manual debugging process, reducing model complexity and computational overhead, making the equalizer more adaptable to the real-time operation requirements of high-speed DSP and other hardware platforms, improving the reliability and stability of signal recovery at the receiver of the optical communication system, and improving signal transmission quality.

[0054] See Figure 5 As shown in the figure, this application discloses a signal compensation device for optical communication, applied at the receiving end, comprising: The dataset construction module 11 is used to construct a training set and a validation set including multiple sample pairs for the current optical communication scenario; the sample pairs are constructed based on the first historical digital signal obtained by the receiver and the second historical digital signal to be transmitted by the transmitter. The equalizer update module 12 is used to alternately update the computation weights and selection weights of the candidate compensation operators of each architecture search unit in the initial equalizer using the training set and the validation set, respectively, in order to obtain the target equalizer. The signal compensation module 13 is used to acquire the digital signal to be compensated of the optical signal currently received by the receiver, and to use the target equalizer to perform signal compensation on the digital signal to be compensated in order to obtain the target compensated signal.

[0055] Furthermore, embodiments of this application also provide an electronic device. Figure 6 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application.

[0056] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Specifically, it may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the signal compensation method for optical communication performed by the electronic device disclosed in any of the foregoing embodiments.

[0057] In this embodiment, the power supply 23 is used to provide operating voltage for various hardware devices on the electronic device; the communication interface 24 can create a data transmission channel between the electronic device and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0058] The processor 21 may include one or more processing cores, such as a quad-core processor or an octa-core processor. The processor 21 may be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). The processor 21 may also include a main processor and a coprocessor. The main processor, also known as a CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, the processor 21 may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, the processor 21 may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning.

[0059] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored on it include operating system 221, computer program 222 and data 223, etc., and the storage method can be temporary storage or permanent storage.

[0060] The operating system 221 manages and controls the various hardware devices and computer programs 222 on the electronic device to enable the processor 21 to perform calculations and processing on the massive amounts of data 223 in the memory 22. The operating system can be Windows, Unix, Linux, etc. The computer program 222, in addition to including a computer program capable of performing the signal compensation method in optical communication disclosed in any of the foregoing embodiments, may further include computer programs capable of performing other specific tasks. The data 223 may include data received by the electronic device from external devices, as well as data collected by its own input / output interface 25.

[0061] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned signal compensation method in optical communication. The specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0062] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0063] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application. The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly in hardware, software modules executed by a processor, or a combination of both. The software module may be located in random access memory (RAM), memory, read-only memory (ROM), electrically programmable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), register, hard disk, removable disk, CD-ROM (Compact Disc Read-Only Memory), or any other form of storage medium known in the art.

[0064] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0065] The above provides a detailed description of a signal compensation method, apparatus, device, and medium in optical communication provided by the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only intended to help understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A signal compensation method in optical communication, characterized in that, Applied to the receiving end, including: A training set and a validation set, comprising multiple sample pairs, are constructed for the current optical communication scenario; the sample pairs are constructed based on the first historical digital signal acquired by the receiver and the second historical digital signal to be transmitted by the transmitter. The computational weights and selection weights of the candidate compensation operators of each architecture search unit in the initial equalizer are alternately updated using the training set and the validation set, respectively, to obtain the target equalizer; The digital signal to be compensated of the optical signal currently received by the receiving end is obtained, and the target equalizer is used to perform signal compensation on the digital signal to be compensated in order to obtain the target compensated signal.

2. The signal compensation method in optical communication according to claim 1, characterized in that, The construction of a training set and a validation set, comprising multiple sample pairs, for the current optical communication scenario includes: The original historical optical signal received by the receiving end in the current optical communication scenario is collected, and the original historical optical signal is preprocessed to obtain the preprocessed historical optical signal. The first historical digital signal of the preprocessed historical optical signal and the second historical digital signal to be transmitted by the transmitting end are obtained, and multiple sample pairs are constructed based on the first historical digital signal and the second historical digital signal. The sample pairs are divided to obtain training and validation sets.

3. The signal compensation method in optical communication according to claim 2, characterized in that, The acquisition of the first historical digital signal of the preprocessed historical optical signal and the second historical digital signal to be transmitted by the transmitting end includes: I / Q sequences are extracted from the preprocessed historical optical signals, and time windows of a preset window size are slid within the I / Q sequences to extract the first sub-sequence groups under different time windows. The first four-way windowed I / Q subsequences in the first subsequence group are concatenated to obtain the first historical digital signal in real vector form; Based on a preset window size, the original historical digital signal to be transmitted by the transmitting end is synchronously framed with the first sub-sequence group to obtain each second sub-sequence group, and the second sub-sequence group is determined as the second historical digital signal; wherein, the second sub-sequence group includes the second four-channel windowed I / Q sub-sequence.

4. The signal compensation method in optical communication according to claim 1, characterized in that, Before alternately updating the computational weights and selection weights of the candidate compensation operators for each architecture search unit in the initial equalizer using the training set and the validation set respectively, the method further includes: An initial equalizer is constructed, comprising multiple reusable architecture search units; wherein the architecture search units include multiple candidate compensation operators, which are used to perform compensation operations on digital signals.

5. The signal compensation method in optical communication according to any one of claims 1 to 4, characterized in that, The step of alternately updating the computational weights and selection weights of the candidate compensation operators of each architecture search unit in the initial equalizer using the training set and the validation set, respectively, to obtain the target equalizer, includes: The initial equalizer is determined as the current equalizer. The selection weights of the candidate compensation operators of each architecture search unit in the current equalizer are fixed, and the calculation weights of the candidate compensation operators of each architecture search unit in the current equalizer are updated using the training set to obtain the first updated equalizer. The computation weights of the candidate compensation operators of each architecture search unit in the first updated equalizer are fixed, and the selection weights of the candidate compensation operators of each architecture search unit in the first updated equalizer are updated using the verification set to obtain the second updated equalizer. The second updated equalizer is determined as the new current equalizer, and the process jumps back to the step of fixing the selection weights of the candidate compensation operators of each architecture search unit in the current equalizer until the preset stopping condition is met, and the output second updated equalizer is determined as the target equalizer.

6. The signal compensation method in optical communication according to claim 5, characterized in that, The step of determining the second updated equalizer as the target equalizer includes: Each architecture search unit in the second updated equalizer output is sequentially determined as the current architecture search unit; The selection weights of each candidate compensation operator in the current architecture search unit are sorted, and the candidate compensation operators of the current architecture search unit are lightweighted according to the sorted selection weights to obtain the current target architecture search unit. The target equalizer is obtained based on the search units of each target architecture.

7. The signal compensation method in optical communication according to claim 5, characterized in that, The step of updating the computational weights of the candidate compensation operators for each architecture search unit in the current equalizer using the training set includes: The first historical digital signal in the training set is input into each architecture search unit of the current equalizer to obtain the output results of each candidate compensation operator in each architecture search unit. A normalized exponential function is applied to the computational weights of each candidate compensation operator to obtain the target weights of each candidate compensation operator. The target weights are then used to perform a weighted summation of the output results in each architecture search unit to obtain the weighted result of each architecture search unit. The computational weights of the candidate compensation operators of each architecture search unit in the current equalizer are updated based on the loss function value between the second historical digital signal in the training set and the weighted result.

8. A signal compensation device for optical communication, characterized in that, Applied to the receiving end, including: The dataset construction module is used to construct a training set and a validation set including multiple sample pairs for the current optical communication scenario; the sample pairs are constructed based on the first historical digital signal obtained by the receiver and the second historical digital signal to be transmitted by the transmitter. The equalizer update module is used to alternately update the computation weights and selection weights of the candidate compensation operators of each architecture search unit in the initial equalizer using the training set and the validation set, respectively, in order to obtain the target equalizer. The signal compensation module is used to acquire the digital signal to be compensated of the optical signal currently received by the receiver, and to use the target equalizer to perform signal compensation on the digital signal to be compensated in order to obtain the target compensated signal.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the steps of the signal compensation method in optical communication as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store a computer program; wherein, when the computer program is executed by a processor, it implements the steps of the signal compensation method in optical communication as described in any one of claims 1 to 7.

Images