Digital signal compensation method and system for optical module

Abstract

The invention discloses a digital signal compensation method and system for an optical module, and relates to the field of network transmission, and the method comprises the steps: dividing an input ideal driving signal into a static basis and a dynamic basis in an initialization stage, respectively constructing a static basis matrix and a dynamic basis matrix, and obtaining a static basis coefficient and a dynamic basis coefficient; performing QR decomposition orthogonalization on the dynamic basis, and removing the part overlapped with the static basis to obtain a sparse set; an output compensation signal is obtained; and updating the static basis coefficient. According to the invention, an input signal is divided into a static base and a dynamic base, so that accurate compensation of instantaneous nonlinear and dynamic nonlinear distortion is realized. The number of static bases and the number of dynamic bases are adjusted in a self-adaptive mode by combining hardware resources and channel performance parameters, performance targets and resource constraints are considered, and good universality is achieved. In the operation stage, static base self-adaptive updating is triggered through periodic error monitoring, the dynamic base learning step length is reduced and returned, and it is ensured that dynamic response is stable and signal distortion is controllable.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of network transmission, in particular to a digital signal compensation method and system for optical modules. BACKGROUND

[0002] An optical transceiver module is a core device for converting electrical signals and optical signals in an optical communication system. It is generally inserted into the interface of a switch, router, server or optical transmission equipment for data transmission through an optical fiber. The optical module includes an optical transmitting assembly, an optical receiving assembly, a DSP chip and a control circuit and interface.

[0003] In the process of converting electrical signals and optical signals back and forth by the optical module, various distortions and noises are introduced, resulting in a decrease in signal quality. In order to ensure stable transmission at high speed, digital signal compensation is required, that is, real-time correction of distortions and noises is performed through a DSP (digital signal processing) algorithm.

[0004] In a high-speed optical module (such as a 400G / 800G PAM4), the nonlinear distortion of the electrical driver and the modulator can severely compress the eye diagram opening, causing inter-symbol interference (ISI) and an increase in bit error rate.

[0005] The existing Volterra filter or LUT pre-distortion has high computational complexity and is difficult to realize in real time in a low-power DSP.

[0006] Traditional LUT pre-distortion can only compensate for static nonlinearity and is usually based on the mapping of input amplitude and output amplitude. It cannot effectively handle dynamic distortion (memory effect), such as bandwidth limitations of modulators / drivers and hysteresis effects. Moreover, the number of LUT points is limited, and when the resolution is not enough, step effects or error amplification can occur at the symbol edges. The LUT curve changes with temperature, current, and device aging, and frequent recalibration is required, resulting in high maintenance costs. High-precision LUT requires a large amount of storage space (such as several hundred to several thousand points), which is not friendly to the limited DSP / storage resources inside the optical module.

[0007] Traditional Volterra filter expansion contains a large number of high-order and cross terms. For example, the calculation amount and the number of taps increase exponentially with the third-order kernel, making it difficult to run in real time on a low-power chip. The large number of coefficients slows down the training process and makes it easy to diverge, requiring a long time of data and complex algorithms. Each kernel coefficient requires multiplication and addition operations, which requires a large number of DSP slices and significantly increases power consumption. It is difficult to meet the convergence speed and delay requirements in high-speed (50Gbaud~100Gbaud) optical module scenarios.

[0008] In the prior technical solution, the LUT and the Volterra are directly cascaded, without distinguishing static and dynamic, which may cause resource waste, and the coefficients of the two are updated independently, which may cause parameter adjustment conflict and convergence shock.

[0009] The prior art such as the invention patent with the announcement number CN108718217B is a compensation method based on a coherent optical communication system, which comprises: an electrical domain OFDM transmitting module for transmitting an electrical domain OFDM signal; an optical domain modulation module for modulating the electrical domain OFDM signal; a digital conversion module of an electrical domain OFDM receiving module for converting into a digital signal; performing IQ compensation on the digital signal through a specific matrix; and outputting the IQ compensated signal according to signal frequency domain equalization and signal demodulation.

[0010] The prior art such as the invention patent with the announcement number CN112615679B is a spatial coherent optical communication frequency tracking system and a frequency shift tracking compensation method, relating to the field of spatial optical communication, which comprises: a local laser for outputting local light for mixing; a digital signal processing module for calculating the frequency difference between the carrier frequency of a received optical signal and the optical frequency of the local light; a control module for receiving the frequency difference calculated by the digital signal processing module; and the digital signal processing module is configured to: when the frequency difference does not exceed a demodulation threshold, directly demodulate the received optical signal; when the frequency difference exceeds the demodulation threshold but does not exceed a compensation range, demodulate the received optical signal after compensation; and when the frequency difference exceeds the demodulation threshold and the compensation range, feed back the frequency difference to the control module and drive the local laser to adjust the optical frequency of the local light through the control module.

[0011] Based on the above content, it can be seen that the prior art relies on a fixed matrix for IQ compensation, which has poor adaptability when the channel characteristics change dynamically, especially in the case of strong fiber nonlinearity, intersymbol interference or frequency selective fading, and the compensation effect may be insufficient. At the same time, the signal is mainly processed by frequency domain equalization, which may not be able to track the rapidly changing channel characteristics, resulting in an increase in the bit error rate. The prior art has limitations in adapting to high-speed dynamic channel changes and processing nonlinear distortion, which limits its application in more complex optical communication systems. SUMMARY

[0012] In view of the deficiencies of the prior art, the present application provides a digital signal compensation method and system for optical modules, in order to achieve the above purpose, the present application is implemented by the following technical solutions: a digital signal compensation method for optical modules, comprising: In the initialization stage, the input ideal drive signal is divided into static bases and dynamic bases, and static base matrix and dynamic base matrix are constructed respectively, the number of static bases of the static base matrix is determined based on input amplitude and hardware resource parameters, and static base coefficients are obtained, the number of dynamic bases is dynamically adjusted based on channel performance parameters, and dynamic base coefficients are obtained.

[0013] In the initialization stage, the input signal is decomposed into static base subspace and dynamic base subspace, the dynamic bases are orthogonalized by QR decomposition, the overlapping part with the static bases is removed, and sparse sets are obtained by sparsifying the dynamic base coefficients.

[0014] In the running stage, the digital signal sequence from the optical module ADC is received, the static compensation component is calculated based on the static base coefficients obtained in the initialization stage, the dynamic compensation component is calculated based on the dynamic base coefficients and the sparse sets, and the static compensation component and the dynamic compensation component are superimposed to obtain the output compensation signal.

[0015] In the periodic updating stage, when the static base coefficients are updated, the static nonlinear change rate is adaptively adjusted, the dynamic base coefficient learning step is temporarily saved, the dynamic base coefficient learning step is reduced in proportion, and the dynamic base coefficient learning step is reduced according to the performance index.

[0016] In addition, a system for a digital signal compensation method for an optical module is also provided, which specifically comprises: The initialization module is configured to divide the input ideal drive signal into static bases and dynamic bases in the initialization stage, construct static base matrix and dynamic base matrix respectively, determine the number of static bases of the static base matrix based on input amplitude and hardware resource parameters, and obtain static base coefficients, dynamically adjust the number of dynamic bases based on channel performance parameters, and obtain dynamic base coefficients. The input signal is decomposed into static base subspace and dynamic base subspace, the dynamic bases are orthogonalized by QR decomposition, the overlapping part with the static bases is removed, and sparse sets are obtained by sparsifying the dynamic base coefficients.

[0017] The running module is configured to receive the digital signal sequence from the optical module ADC, calculate the static compensation component based on the static base coefficients obtained in the initialization stage, calculate the dynamic compensation component based on the dynamic base coefficients and the sparse sets, and superimpose the static compensation component and the dynamic compensation component to obtain the output compensation signal.

[0018] The updating module is configured to adaptively adjust the static nonlinear change rate when the static base coefficients are updated, temporarily save the dynamic base coefficient learning step, reduce the dynamic base coefficient learning step in proportion, and reduce the dynamic base coefficient learning step according to the performance index.

[0019] Compared with the prior art, the embodiments of the present application have at least the following beneficial effects: (1) The application provides a digital signal compensation method for an optical module, which realizes fine modeling of the nonlinear characteristics of the optical module by dividing the input ideal drive signal into a static base and a dynamic base. The static base focuses on describing the nonlinear influence of instantaneous amplitude on the output, while the dynamic base considers the memory effect and intersymbol interference of the signal, capturing dynamic nonlinearities and frequency-selective distortions. This separation strategy can ensure targeted compensation for static nonlinearities and dynamic nonlinearities, avoiding the fitting problem of a single model when processing complex signals, thereby significantly improving the compensation accuracy and signal recovery quality.

[0020] (2) In the dynamic base construction process, the application introduces QR decomposition to orthogonalize the dynamic base vectors and eliminate the overlapping part with the static base, ensuring that the static base and the dynamic base do not interfere with each other. At the same time, by sparsifying and screening the dynamic base coefficients, only the dynamic base functions with obvious compensation contribution are retained, reducing redundant features. This strategy not only improves the stability and generalization ability of the model, but also effectively reduces the computational complexity and hardware resource consumption, making the compensation method practical for high data rate optical modules.

[0021] (3) The application combines the hardware resource parameters and channel performance parameters of the optical module to adaptively adjust the number of static bases and dynamic bases. Through database mapping and weight coupling mechanism, it can reasonably allocate hardware resources while ensuring performance targets, achieving a balance between compensation accuracy and resource consumption. This design enables the system to flexibly adapt to optical module environments with different transmission rates and signal bandwidths, improving the universality and practicality of the scheme.

[0022] (4) In the running phase, the application periodically monitors the signal error and static nonlinear drift, triggers adaptive update of the static base coefficients, and temporarily reduces the learning step of the dynamic base coefficients and performs performance rollback processing. This mechanism ensures coordination between static and dynamic compensation, avoiding signal distortion or oscillation caused by excessive response of the dynamic base, while restoring the learning ability of the dynamic base after the static base update is completed, achieving long-term stable signal compensation. This adaptive update mechanism significantly improves the reliability and stability of the system in actual running environment, especially suitable for high-speed, long-distance optical communication scenarios.

[0023] Of course, any product implementing the present application does not necessarily need to achieve all the above advantages at the same time. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 The method flowchart of the application.

[0025] Figure 2 The system module schematic diagram of the application.

[0026] Figure 3The logic flowchart of the present application.

[0027] Figure 4 The initialization flowchart of the present application.

[0028] Figure 5 The signal compensation flowchart of the running stage of the present application.

[0029] Figure 6 The coefficient adaptive updating flowchart of the present application. DETAILED DESCRIPTION

[0030] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0031] In the description of the present application, it should be understood that the terms "opening", "upper", "lower", "thickness", "top", "middle", "length", "inner", "periphery" and the like indicate the orientation or positional relationship, which are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the components or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation of the present application.

[0032] Please refer to Figure 1 The embodiment of the present application provides a digital signal compensation method for an optical module, which specifically comprises: As Figure 3 The logic flowchart involved in the embodiment of the present application is shown, which shows the complete processing flow of the digital signal compensation of the optical module, including three stages of initialization, running and periodic updating. In the initialization stage, the system inputs the ideal driving signal and decomposes it into static base and dynamic base, constructs the static base matrix to obtain the static base coefficient, constructs the dynamic base matrix at the same time and dynamically adjusts the number and coefficient, and then projects the signal to the static and dynamic base subspace, orthogonalizes and sparsifies the dynamic base, and provides parameters for subsequent compensation. In the running stage, the system receives the digital signal collected by the ADC, calculates the static compensation component and the dynamic compensation component respectively, superimposes the two to generate the complete output compensation signal, and realizes the real-time compensation for the static and dynamic nonlinear distortion. In the periodic updating stage, the system monitors the static nonlinear drift, and if the drift occurs, the static base coefficient is updated adaptively, the dynamic base coefficient is fine-tuned and the learning step is rolled back according to the performance index, and after the static updating is completed, the dynamic base learning step is restored, so as to ensure the long-term stable signal compensation effect.

[0033] As Figure 4The initialization flowchart in the embodiment of the application is shown. An input signal generates a digital signal sequence through sampling, and each sampling point corresponds to an amplitude. Then, the sampling point sequence is subjected to feature extraction to form a candidate feature sequence, and a self-correlation coefficient or a delay dependency index is calculated. According to a dependency threshold, the feature sequence is divided into a static base and a dynamic base to form a static base matrix and a dynamic base matrix, respectively. The core purpose of the diagram is to explain the whole logic from sampling to feature division, and then to base matrix construction, which provides a basis for subsequent compensation calculation. In the static base coefficient solving process, the system first determines the number of static bases according to input amplitude statistics, and solves the static base coefficients by using the least square method or the regularization method. At the same time, the performance index is calculated in combination with the hardware resource information and the performance target, and the hardware constraint is mapped to ensure that the number of static bases and the coefficients are within the bearable range. In the dynamic base number adjustment process, the system calculates the number of dynamic bases according to the performance index or the channel state, and solves the dynamic base coefficients by fitting the static error signal through the least square method. The core purpose of the diagram is to explain how to dynamically adjust the number of bases and solve the corresponding coefficients under the hardware constraint and the performance target to ensure the signal compensation performance.

[0034] In the initialization stage, the input ideal drive signal is divided into a static base and a dynamic base, and a static base matrix and a dynamic base matrix are constructed, respectively. The number of static bases of the static base matrix is determined based on the input amplitude and the hardware resource parameters, and the static base coefficients are obtained. The number of dynamic bases is dynamically adjusted based on the channel performance parameters, and the dynamic base coefficients are obtained.

[0035] The input ideal drive signal is divided into a static base and a dynamic base, and the specific process is as follows: The static base refers to a nonlinear feature that depends only on the current input signal, without considering the influence of the previous and subsequent symbols, without a delay term, and only describes the nonlinear influence of the input instantaneous amplitude on the output, with a small amount of calculation and easy real-time processing.

[0036] The dynamic base refers to a combination of the current and past input signals, which is used to describe the memory effect and the symbol-to-symbol crosstalk, and includes a delay term, which can capture dynamic nonlinearities, symbol-to-symbol interference and frequency-selective distortions.

[0037] The sampling frequency is determined based on the Nyquist sampling theorem according to the transmission rate and the signal bandwidth of the optical module. The time interval of the sampling points is fixed by the sampling frequency. A clock signal at the sampling frequency is generated using a crystal oscillator. The ADC performs acquisition based on the clock signal. The time points of the clock signal are the sampling points, and each sampling point corresponds to a signal amplitude at a time position. The sampling point is the basic unit for subsequent feature extraction, static base and dynamic base construction. It should be noted that the higher the sampling frequency, the denser the sampling points, and the more details can be captured, but the amount of calculation also increases.

[0038] The feature extraction is performed on the digital signal of the sample point sequence of the input signal to obtain each candidate feature sequence, and the autocorrelation coefficients of each candidate feature sequence are calculated through the autocorrelation function, the autocorrelation coefficient is used to quantify the relationship between each candidate feature sequence and the delayed version of itself, the larger the autocorrelation coefficient is, the more obvious the dependence of the feature on the past symbol is, and it should be noted that the autocorrelation coefficient ranges from -1 to 1.

[0039] Based on the preset delay time length, the autocorrelation coefficients of all non-zero delays are accumulated (that is, the autocorrelation coefficients of the historical sampling points are added), which represents the overall dependence of the candidate sequence feature on the past signal, and the ratio calculation is performed with the zero-delay autocorrelation coefficient to obtain the delay dependence index of each candidate feature sequence, it should be noted that the zero-delay autocorrelation coefficient is the correlation degree of the candidate feature sequence itself at the current time with itself, and the non-zero-delay autocorrelation coefficient is the correlation degree between the candidate feature sequence and the corresponding values at past time, the delay dependence index reflects the influence degree of the past symbol on each candidate feature sequence, if the delay dependence of a certain candidate feature sequence is less than the preset delay dependence threshold, it is determined that the candidate feature sequence hardly depends on the past symbol, and is classified as a static base, if the delay dependence of a certain candidate feature sequence is greater than or equal to the preset delay dependence threshold, the candidate feature sequence is obviously affected by the past symbol, and is classified as a dynamic base.

[0040] The static base matrix and the dynamic base matrix are constructed respectively, and the specific process is as follows: The candidate feature sequence classified as a static base is extracted, denoted as a static feature sequence, the values of the static feature sequence at each sampling point are sequentially arranged based on the sampling point sequence as a column of the static base matrix, and the sampling point sequence is used as the row of the static base matrix, the static base matrix is used to describe the instantaneous nonlinear interference, denoted as , wherein is the static base matrix, is the sampling point, S is the number of static bases, and the elements in the static base are denoted as , wherein denotes the element in the nth row and the ith column of the static base matrix, n corresponds to the sampling point index, n=1, 2, 3,..., N, i corresponds to the static base function index, i=1, 2, 3,..., S, denotes the mapping of the ith static base function from the input x[n] to the feature space.

[0041] The candidate feature sequence classified as a dynamic base is extracted, denoted as a dynamic feature sequence, the values of the dynamic feature sequence at each sampling point are sequentially arranged based on the sampling point sequence as a column of the dynamic base matrix, and the sampling point sequence is used as the row of the dynamic base matrix, the dynamic base matrix is used to represent the dynamic nonlinear interference, denoted as , wherein is a dynamic basis matrix, N is a sampling point, D is a dynamic basis number, and an element in the dynamic basis is denoted as wherein, denotes an element in the dynamic basis matrix, n corresponds to a sampling point index, n = 1, 2, 3, …, N, j corresponds to a dynamic basis function index, j = 1, 2, 3, …, D, denotes that the nth dynamic basis function maps the input x[n] to the feature space.

[0042] The static basis number of the static basis matrix is determined based on the input amplitude and the hardware resource parameters, and the static basis coefficients are obtained, specifically including: The hardware resource parameters are obtained, including available storage capacity, available multiply-add unit number, available throughput, and delay budget. It should be noted that the available storage capacity refers to the storage resource size available for compensation algorithm calling in the optical module digital signal processing unit, including the available space of on-chip cache, external memory, and intermediate register. Its acquisition method is usually determined through a hardware resource management interface or a design specification table, for example, the available storage block number and its capacity are given by the synthesis report on the FPGA or ASIC platform. The available multiply-add unit number refers to the number of multipliers and adders that can be used in parallel in the processor, FPGA or ASIC. These units directly determine the parallel computing capability of the algorithm on the hardware. Its acquisition method depends on the structure description and resource allocation of the hardware implementation platform, for example, the FPGA design tool reports the remaining DSP unit number; in the DSP chip, it can be monitored by a hardware counter; in the special chip, it is explicitly given by the design specification. The available throughput refers to the amount of data that the system can process per unit time, which is one of the core indicators of optical communication link performance optimization. Its acquisition method is generally determined by system design goals and business requirements, through link budget and rate planning setting, for example, in the design of 400G or 800G optical modules, the throughput target is directly given by the standard protocol. The delay budget refers to the upper limit of the total processing delay from the time when the signal enters the digital compensation module to the time when it is output, including storage access delay, operation delay, and queue waiting time, etc., which is an important constraint for measuring the real-time performance of the system. Its acquisition method is usually determined by the real-time requirement in the system design stage, for example, the requirement for delay in high-speed optical communication systems is in the order of microseconds.

[0043] The performance target parameter includes an error upper limit, a bit error rate upper limit, and an error vector amplitude. The error upper limit refers to a maximum deviation range allowed between a system output signal after compensation and an ideal target response, which is used to ensure overall compensation accuracy. The error upper limit is obtained from a system design specification, for example, a maximum mean square error threshold set by link tolerance analysis. The bit error rate upper limit refers to a maximum bit error rate (BER) acceptable in an optical communication link. The bit error rate upper limit is usually determined by a communication standard (such as an OTN or Ethernet standard) or service requirements, and can also be obtained by performing bit error detection on an actual transmission signal by using a link test instrument. The error vector amplitude (EVM) is used to measure the degree of shift of a modulation signal constellation point, and reflects modulation quality and system linearity. The EVM is defined as the ratio of the root mean square value of an error vector between a received signal and an ideal signal to an ideal reference amplitude, and is usually expressed as a percentage. The EVM can be obtained by sampling and calculating a received signal by using an oscilloscope or a vector signal analyzer, or by comparing and analyzing a received signal with a reference signal in a digital processing module.

[0044] The performance target parameter unit weighting factors corresponding to the performance target parameters are extracted from the database, including an error upper limit unit weighting factor, a bit error rate upper limit unit weighting factor, and an error vector amplitude unit weighting factor. The performance target parameters are multiplied by the performance target parameter unit weighting factors one by one to obtain the performance target index. The performance target parameter unit weighting factors are used to eliminate the dimensions and normalize the orders of magnitude of different performance target parameters, so that the performance target parameters are in a comparable scale, and the weights of the performance target parameters in the performance target index are adjusted. The specific calculation process includes: ; wherein, is the performance target index, is the bth performance target parameter. In the embodiment of the present application, the performance target parameters include an error upper limit, a bit error rate upper limit, and an error vector amplitude, that is, is the error upper limit, is the bit error rate upper limit, is the error vector amplitude. In other embodiments of the present application, other related parameters can be selected for calculation, is the performance target parameter unit weighting factor corresponding to the performance target parameter. In the embodiment of the present application, the performance target parameter unit weighting factors include an error upper limit unit weighting factor, a bit error rate upper limit unit weighting factor, and an error vector amplitude unit weighting factor, that is, is the error upper limit unit weighting factor, is the bit error rate upper limit unit weighting factor, is the error vector amplitude unit weighting factor. b is the number of the performance target parameter, b = 1, 2, 3,..., B, and B is the total number of the performance target parameters. In the embodiment of the present application, B is 3.

[0045] It should be noted that there is a certain correlation between the error upper limit, the error rate upper limit and the error vector magnitude value, the error upper limit is the most basic constraint, which directly limits the deviation between the output signal after compensation and the ideal signal. If the error is too large, it will lead to serious deviation, thereby directly affecting the error vector magnitude (EVM). Therefore, EVM can be regarded as a specific quantitative representation of the error upper limit in the modulation domain. There is a strong correlation between EVM and bit error rate (BER). Smaller EVM means lower probability of error in symbol decision, thereby reducing BER; on the contrary, if EVM is too large, noise and nonlinear interference are more likely to cause symbol decision error, which increases BER.

[0046] The performance target index is input into a preset mapping set of performance target index-hardware target resource parameter in a database to obtain hardware target resource parameters, including target available storage capacity, target available multiply-add unit quantity, target available throughput and target latency budget, and the hardware resource parameters and the hardware target resource parameters are compared and coupled to obtain a hardware resource representation value, and the specific calculation process includes: ; Among them, the hardware resource representation value is C, the u-th hardware resource parameter, in the embodiment of the application, including available storage capacity, available multiply-add unit quantity, available throughput and latency budget, that is, C1 is available storage capacity, C2 is available multiply-add unit quantity, C3 is available throughput, and C4 is latency budget, the hardware target resource parameter corresponding to the u-th hardware resource parameter, including target available storage capacity, target available multiply-add unit quantity, target available throughput and target latency budget, wherein, the target available storage capacity is C, the target available multiply-add unit quantity is C, the target available throughput is C, the target latency budget is C, the weighting factor corresponding to the u-th hardware resource parameter, including available storage capacity weighting factor, available multiply-add unit quantity weighting factor, available throughput weighting factor and latency budget weighting factor, wherein, the available storage capacity weighting factor is C, the available multiply-add unit quantity weighting factor is C, the available throughput weighting factor is C, the latency budget weighting factor is C, and u is the hardware resource parameter number, u=1, 2, 3,..., U, U is the total number of hardware resource parameters, in the embodiment of the application, U is 4.

[0047] It should be noted that the available storage capacity weighting factor, the available multiply-add unit quantity weighting factor, the available throughput weighting factor and the delay budget weighting factor act to adjust the weight of each parameter in the calculation process of the hardware resource representation value. The available storage capacity weighting factor is mainly obtained by modeling the storage utilization of the hardware platform and the storage complexity of the algorithm. The specific method is to establish the relationship curve between the capacity and the performance by statistically analyzing the change of the system performance indicators such as the error convergence speed or the compensation accuracy under different storage capacities, and then to obtain the weighting factor through normalization processing. The available multiply-add unit quantity weighting factor is obtained by combining the hardware architecture characteristics and the parallel computing capability evaluation. On the FPGA, ASIC or DSP platform, the processing delay and throughput improvement under different multiply-add unit quantities are measured through simulation or experiment, so as to extract the marginal effect, and then the factor value is formed through normalization. The available throughput weighting factor is obtained by comparing the error rate or error vector amplitude improvement effect that the system can achieve under different throughput targets. The specific method is to establish the correspondence between the throughput and the performance improvement by using link budget and system simulation, and then to convert the weighting factor through numerical normalization. The delay budget weighting factor is obtained by measuring the influence of delay on the system error rate and link stability. In the experiment or simulation, the mapping curve of delay and performance loss is established, and the degree of performance degradation when the compensation delay increases is analyzed. Then the factor value is obtained through normalization.

[0048] It should be noted that there is a certain correlation between the available storage capacity, the available multiply-add unit quantity, the available throughput and the delay budget. The available storage capacity and the available multiply-add unit quantity directly determine the algorithm complexity that the system can achieve: larger storage capacity allows more historical symbols and intermediate results to be cached, thereby supporting higher-order dynamic basis compensation, and more multiply-add units improve the parallel computing capability, so that complex compensation calculation can be completed within a limited delay. The available throughput imposes a joint constraint on the storage capacity and the multiply-add unit. In order to ensure that a predetermined number of symbols or bits can be processed per second, the system needs to meet the demand on the calculation unit and the storage bandwidth at the same time, otherwise it will cause data blocking or processing speed reduction. The delay budget plays a role in the final convergence constraint on the above three. Even if the storage and calculation resources are sufficient, if the calculation path of the compensation algorithm is too long or the memory access frequency is too high, the delay budget may be exceeded, resulting in that the compensation result cannot be output in real time. Therefore, the storage capacity, the number of calculation units, the throughput target and the delay budget must be overall balanced. Increasing the number of basis functions can improve the compensation accuracy, but it will occupy more storage and operation resources and may increase the processing delay; on the contrary, excessively compressing the resource overhead can reduce the delay, but it may sacrifice the throughput and the compensation accuracy.

[0049] The input amplitude refers to the amplitude range of the sampling signal before digital compensation of the optical module, that is, the size distribution of the digital signal collected from the ADC in numerical value.

[0050] The input amplitude is counted, and the distribution parameters of the input amplitude are obtained through MATLAB for the digital signal obtained by sampling, including the maximum amplitude, the minimum amplitude, and the average amplitude. Based on the distribution parameters of the input amplitude, the initial number of static bases required for each amplitude interval is determined, specifically including: Based on the distribution parameters of the input amplitude, the discrete sampling signal amplitude is mapped into a continuous probability distribution. The probability density function is constructed based on the histogram method. The probability density function is used to reflect the probability distribution of different amplitude intervals. If the signal amplitude distribution is concentrated in the low-amplitude interval, it means that most of the symbol energy is small, otherwise, it means that the strong nonlinear component is more obvious. The higher the amplitude, the easier it is to trigger the nonlinear distortion of the optical module, and more base functions are needed to describe. The probability distribution of each amplitude interval is input into the mapping set of amplitude interval-required static base number pre-stored in the database, and the initial number of static bases required for each amplitude interval is obtained through mapping matching. The mapping set of amplitude interval-required static base number is a mapping relationship set obtained through offline simulation.

[0051] The hardware resource representation value is input into the mapping set of hardware resource representation value-static base number upper limit pre-stored in the database for mapping matching to obtain the static base number upper limit.

[0052] If the initial static base number is less than the static base number upper limit, the initial static base number is taken as the final static base number and applied.

[0053] If the initial static base number is greater than or equal to the static base number upper limit, the static base number upper limit is taken as the final static base number and applied.

[0054] The ideal target response of the optical module is taken as the ideal output, the static base matrix is input into the optical module, and the actual output is obtained. By minimizing the error between the actual output and the ideal output, the coefficients of each static base function are solved, which are denoted as static base coefficients. It should be noted that the minimization process is used to find a set of coefficients of static base functions, so that the difference between the output signal generated by them and the ideal target response of the optical module is minimized. The static base coefficient is the optimal weight of the static base function when fitting the ideal output signal of the optical module. By minimizing the sum of squares of errors, the coefficient solving process is converted into a standard least squares fitting problem.

[0055] In the embodiment of the present application, the optimal solution is obtained by matrix operation in a closed form solution, which is suitable for the case where the number of static basis functions is small and the matrix condition is good. In the specific solution of other embodiments of the present application, when the number of basis functions is large, the matrix size is large, or there is a pathological condition, iterative optimization methods such as gradient descent and quasi-Newton method can be used to gradually approach the optimal coefficients. In some scenarios where overfitting needs to be avoided or a sparse solution is desired, a regularization term is added to the error function, such as a ridge regression with L2 norm constraint or a Lasso with L1 norm constraint, thereby controlling the coefficient size or sparsity.

[0056] The above process can simultaneously consider performance targets and hardware resource constraints in the determination of the number of static basis matrices and the coefficient solving process, which has a significant advantage, avoiding the performance deficiency or resource waste problem that easily occurs in the traditional fixed basis function number scheme, and achieving a dynamic balance between performance and resources. Based on the constraint mechanism of the hardware resource representation value and the upper limit of the number of static bases, the number of basis functions will not exceed the actual range that the hardware can bear, thereby ensuring the feasibility and real-time performance of the compensation calculation, and avoiding the time delay problem caused by excessive calculation. This process enables the present application to adapt to different application scenarios and hardware conditions, ensuring efficiency in low-resource scenarios and improving accuracy in high-performance requirements, so that the output compensation result is closer to the real characteristics of the optical module, significantly improving the link performance.

[0057] The number of dynamic bases is dynamically adjusted based on the channel performance parameters to obtain dynamic base coefficients, specifically including: The hardware resource representation value is extracted and mapped and matched in the pre-stored mapping set of hardware resource representation value-upper limit of the number of dynamic bases in the database to obtain the upper limit of the number of dynamic bases.

[0058] The channel performance parameters include the order of channel nonlinearity, bit error rate and signal-to-noise ratio. The channel performance parameter unit weighting factors are extracted from the database, including the order of channel nonlinearity unit weighting factor, bit error rate unit weighting factor and signal-to-noise ratio unit weighting factor. The channel performance parameters are multiplied by the channel performance parameter unit weighting factors and then coupled to obtain the channel performance limitation index. It should be noted that the algorithm model of the channel performance limitation index is the same as that of the performance target index, which will not be described here.

[0059] The number of dynamic bases is adjusted by multiplying the initial number of dynamic bases based on the channel performance limitation index. It should be noted that the initial number of dynamic bases is the number of sampling points. When the adjusted number of dynamic bases is greater than or equal to the upper limit of the number of dynamic bases, the upper limit of the number of dynamic bases is taken as the final number of dynamic bases. When the adjusted number of dynamic bases is less than the upper limit of the number of dynamic bases, the adjusted number of dynamic bases is taken as the final number of dynamic bases.

[0060] The dynamic base matrix is fitted with a static error residual signal by using a least square method, the static error residual signal is a signal removing a static base error part, and coefficients of each dynamic base function are solved and recorded as dynamic base coefficients, a specific process is that a set of coefficients are solved to make a linear combination of the dynamic base matrix after weighting closest to the residual signal, a specific solving method is that a transpose of the dynamic base matrix is multiplied by itself and inverse, and then the transpose of the dynamic base matrix is multiplied by the residual signal to obtain the dynamic base coefficients, the dynamic base coefficients are used to reflect contributions of corresponding dynamic base functions to dynamic nonlinear compensation.

[0061] The above content realizes adaptive optimization of the nonlinear compensation capability of the optical module by closely associating the number of dynamic bases with the channel performance parameters. By extracting the hardware resource characterization value and matching the pre-stored hardware resource-dynamic base number upper limit, it can be ensured that the number of dynamic bases is reasonably allocated within the range of available hardware resources, avoiding exceeding the system calculation and storage capacity, while fully utilizing the hardware potential. The channel nonlinear order, bit error rate and signal-to-noise ratio and other performance parameters are coupled after being multiplied by the corresponding unit weighting factor to obtain the channel performance limiting index, so that the number of dynamic bases can be dynamically adjusted according to the actual channel conditions. In the case of high nonlinearity or low signal-to-noise ratio, the number of dynamic bases is automatically increased to improve the compensation accuracy, while in the case of good channel conditions, the number of dynamic bases is moderately reduced to save computing resources, thereby achieving a balance between performance and resources. The dynamic base matrix is fitted with a static error residual signal by using a least square method, and the dynamic base coefficients are solved, which not only accurately reflects the contribution of each dynamic base function in compensation, but also ensures that the dynamic compensation is optimized for the residual part after the static error is eliminated, improves the overall compensation effect, enhances the adaptive response capability of the system to nonlinear and inter-symbol interference, and takes into account the calculation efficiency and hardware constraints.

[0062] In the initialization phase, the input signal is decomposed into a static base subspace and a dynamic base subspace, the dynamic base is orthogonalized by QR decomposition, the overlapping part with the static base is removed, and the dynamic base coefficients are sparsified to obtain a sparse set.

[0063] The input signal is decomposed into a static base subspace and a dynamic base subspace, and the dynamic base is orthogonalized by QR decomposition, which specifically includes: The static base subspace and the dynamic base subspace are mathematical space concepts used to describe and decompose different characteristics of the input signal in the signal compensation method.

[0064] The static basis subspace is a vector space spanned by all static basis functions. It is mainly used to represent the nonlinear characteristics in the input signal that are related to the current time instant amplitude. Such characteristics do not depend on the past or future sign information, nor do they contain time-delayed terms. The vector combination in the static basis subspace can fit the nonlinear response caused by the current amplitude in the input signal. By projecting the signal into the static basis subspace, the part of the signal that can be compensated by the static basis can be extracted.

[0065] The dynamic basis subspace is a vector space spanned by all dynamic basis functions. The dynamic basis subspace is used to represent the nonlinear characteristics in the input signal that depend on the current and past input at several time instants, including memory effects, intersymbol interference, and frequency-selective distortion. The dynamic basis subspace is usually constructed after the static basis subspace is orthogonalized to ensure that the dynamic basis only captures the nonlinear part that the static basis fails to compensate.

[0066] The input signal is projected into the static basis subspace, the input signal is the actual collected signal containing nonlinear interference for initialization, the input signal is fitted based on the static basis matrix and the static basis coefficient to obtain the static basis output signal, and the residual signal is obtained after the static basis output signal is subtracted from the input signal, the residual signal is the fitting object of the dynamic basis.

[0067] It should be noted that the fitting input signal is realized by linear combination of the static basis matrix and the static basis coefficient. Specifically, each column of the static basis matrix corresponds to the value of a static basis function at each sampling point, and the static basis coefficient represents the weight of each static basis function in the fitting input signal. The columns of the static basis matrix are summed by weighting the corresponding coefficients, and the static basis output signal is obtained, which reproduces the components of the static nonlinear part in the input signal as much as possible. From the perspective of linear algebra, this process is equivalent to projecting the input signal into the static basis subspace, and the static basis output signal is the vector after projection, which only contains the part of the input signal that can be represented by the static basis function. After obtaining the static basis output signal, it is subtracted from the original input signal to obtain the dynamic error signal, which provides the target signal for subsequent fitting of the dynamic basis.

[0068] The dynamic basis matrix is orthogonalized with the static basis subspace, and the column vectors of the dynamic basis are converted into mutually orthogonal vector groups through QR decomposition to ensure that the dynamic basis does not overlap with the static basis.

[0069] It should be noted that the dynamic basis matrix is orthogonalized with the static basis subspace, and the specific process includes: Each column vector of the dynamic basis matrix is projected into the static basis subspace to obtain the component on the static subspace, and the component is subtracted from the dynamic basis vector to obtain a vector orthogonal to the static basis.

[0070] The QR decomposition is performed on the dynamic basis matrix after eliminating the static component, and the matrix is decomposed into an orthogonal matrix Q and an upper triangular matrix R. The column vectors of the Q matrix obtained by QR decomposition are orthogonal to each other, and the overlapping part with the static basis has been removed. Through the above operation, the dynamic basis matrix and the static basis matrix are orthogonal to each other in the vector space, which ensures that the dynamic compensation is only for the nonlinear part that the static basis fails to cover. On the mathematical level, it is equivalent to orthogonally projecting the dynamic basis vector on the static basis subspace, and the remaining part constructs a new dynamic basis vector, so that = 0, where is the new dynamic basis matrix. In wireless communication and optical communication, the orthogonal basis decomposition or Gram-Schmidt orthogonalization is often used to separate static nonlinearity and dynamic nonlinearity, and reference can be made to the existing technologies related to digital feedforward compensation (DFE) and Volterra filter. Through the above-mentioned manner, the compensation accuracy can be improved, and the dynamic basis focuses on compensating the residual nonlinearity, avoiding repeated adjustment of the static part.

[0071] The dynamic basis coefficients are sparsified and collected, specifically including: The performance target index is mapped and matched in the mapping set of the performance target index-coefficient threshold adjustment value pre-stored in the database to obtain the coefficient threshold adjustment value in the sparsification processing. The coefficient threshold is used for screening the retained dynamic basis function. The coefficient threshold adjustment value is corrected with the initial coefficient threshold pre-stored in the database to obtain the coefficient threshold.

[0072] The dynamic basis coefficients are compared with the coefficient threshold. If the dynamic basis coefficient of a certain dynamic basis function is less than the coefficient threshold, the dynamic basis function is excluded. If the dynamic basis coefficient of a certain dynamic basis function is greater than or equal to the coefficient threshold, the dynamic basis function is retained.

[0073] The retained dynamic basis function and the corresponding coefficient form a sparse set.

[0074] As shown in Figure 5 the running stage signal compensation flowchart of the present application. It includes static compensation, dynamic compensation and dynamic basis sparsification processing. First, the real-time signal sequence is obtained from the ADC, and the static compensation component is calculated by using the static basis matrix and the static basis coefficient. Then, the static residual signal is fitted by using the dynamic basis matrix to obtain the dynamic basis coefficient, and the dynamic basis function with small contribution is excluded by the coefficient threshold to form a sparse set. Finally, the static compensation and dynamic compensation components are superimposed at each sampling point to generate the final output compensation signal. The periodic update mechanism can adjust the static basis coefficient and the number of dynamic bases according to the channel state or the change rate of the static nonlinearity.

[0075] In the running phase, a sequence of digital signals from the optical module ADC is received, a static compensation component is calculated based on the static basis coefficients obtained in the initialization phase, a dynamic compensation component is calculated based on the dynamic basis coefficients and the sparse set, and the static compensation component and the dynamic compensation component are superimposed to obtain an output compensation signal.

[0076] The static compensation component and the dynamic compensation component are added to obtain an output compensation signal, and the specific processing condition is: A sequence of real-time sampled digital signals is obtained from the ADC of the optical module, each sampling point corresponds to a signal amplitude at a specific time position, and the sequence of digital signals is the original signal that needs to be compensated and is affected by the static and dynamic nonlinear distortion of the optical module.

[0077] The sequence of digital signals is substituted into the static basis matrix obtained in the initialization phase, and a static compensation component is obtained by weighted summation of the static basis coefficients, the static compensation component is used to compensate the static nonlinear distortion in the signal, and the specific expression is , wherein is the static compensation component, n corresponds to the sampling point index, i corresponds to the static basis function index, represents that the ith static basis function maps the input x[n] to the feature space, and S is the number of static bases, is the static basis coefficient of the ith static basis function.

[0078] The sequence of digital signals is substituted into the dynamic basis matrix, and a dynamic compensation component is obtained by weighted summation of the dynamic basis coefficients in the sparse set, the dynamic compensation component is used to compensate the dynamic nonlinear distortion in the signal, and the specific expression is , wherein is the dynamic compensation component, n corresponds to the sampling point index, and j is the index of the dynamic basis function, represents the index of the dynamic basis function remaining after sparsification, represents that the nth dynamic basis function maps the input x[n] to the feature space, is the dynamic basis coefficient of the jth dynamic basis function.

[0079] The static compensation component and the dynamic compensation component are superimposed at the corresponding sampling points to generate a complete output compensation signal, and the output compensation signal is the digital signal after static nonlinear compensation and dynamic nonlinear compensation, and the specific expression is .

[0080] As Figure 6The coefficient adaptive updating flowchart in the embodiment of the application is shown, and the complete flow of adaptive updating of the static base and the dynamic base in the signal compensation system is described. First, a new input signal or link data is received, which is the starting point of the entire updating process. Then, the static nonlinear variation rate is calculated according to the input signal, and the static base coefficient is adaptively updated using the rate, so that the static base can accurately reflect the long-term characteristics and nonlinear variation of the input signal. At the same time, the system extracts the latest channel performance parameters from the input signal, and updates the number and coefficients of the dynamic base based on these parameters to capture the short-term fluctuations and sudden changes of the input signal. The updated results of the static base and the dynamic base are combined to generate the updated static base coefficient and the dynamic base sparse set, and to prepare for compensation calculation. After the updating is completed, the system enters the running stage, and uses the latest static base and dynamic base coefficients to compensate the signal, thereby optimizing the output performance.

[0081] In the periodic updating stage, when the static base coefficient is updated, the adaptive adjustment is performed based on the static nonlinear variation rate, and the dynamic base coefficient learning step is temporarily saved, the dynamic base coefficient learning step is reduced in proportion and reduced according to the performance index.

[0082] When the output compensation signal is output, the normalized error calculation is performed on the output compensation signal and the comparison signal without compensation, the compensation difference is obtained, and based on the variation rate of the time sequence of the compensation difference in the running period, it is judged whether the static nonlinear characteristic drifts. In the embodiment of the application, a variation rate threshold is set, and when the variation rate of the time sequence of the compensation difference in the running period exceeds the variation rate threshold, it is determined that the static nonlinear characteristic drifts.

[0083] If the static nonlinear characteristic drifts, the adaptive updating of the static base coefficient is triggered immediately, and if the static nonlinear characteristic does not drift, the periodic updating is waited.

[0084] After the updating of the static base coefficient is completed, the dynamic base coefficient is fine-tuned using the compensation difference. Specifically, the compensation difference is decomposed into the dynamic base subspace by fitting the compensation difference through the least square method, and the dynamic base coefficient is updated.

[0085] When the static base coefficient is updated, the dynamic base coefficient learning step is temporarily saved. The dynamic base coefficient learning step refers to the size of the coefficient adjustment amplitude allowed when the dynamic base coefficient is updated.

[0086] Based on the preset safety ratio in the database, the current dynamic base coefficient learning step is reduced through the Gain module, the response speed of the dynamic base to the input signal is reduced, after reducing the dynamic base coefficient learning step, the signal distortion index is monitored by using the BER tester, if the signal distortion index exceeds the signal distortion index threshold, the signal distortion index and the signal distortion index threshold are subtracted to obtain the signal distortion deviation, based on the signal distortion deviation, the signal distortion deviation-retreat amplitude mapping set pre-stored in the database is mapped to obtain the retreat amplitude and the Gain module is retreated, the dynamic base coefficient learning step is further reduced, if the signal distortion index does not exceed the signal distortion index threshold, no retreat operation is performed, and the static base coefficient update is directly waited to be completed.

[0087] After the static base coefficient update is completed, the learning step of the dynamic base coefficient is restored through the Gain module based on the temporarily saved dynamic base coefficient learning step.

[0088] In the embodiment, as Figure 2 The application provides a digital signal compensation system for an optical module, which comprises: An initialization module is configured to divide an input ideal driving signal into static bases and dynamic bases in an initialization stage, construct static base matrices and dynamic base matrices, determine the number of static bases of the static base matrices based on input amplitudes and hardware resource parameters, obtain static base coefficients, dynamically adjust the number of dynamic bases based on channel performance parameters, obtain dynamic base coefficients, decompose the input signal into static base subspaces and dynamic base subspaces, orthogonalize the dynamic bases by QR decomposition, remove the overlapping parts with the static bases, and perform sparse processing on the dynamic base coefficients to obtain a sparse set.

[0089] An operation module is configured to receive a digital signal sequence from an optical module ADC, calculate a static compensation component based on the static base coefficients obtained in the initialization stage, calculate a dynamic compensation component based on the dynamic base coefficients and the sparse set, and superimpose the static compensation component and the dynamic compensation component to obtain an output compensation signal.

[0090] An update module is configured to adaptively adjust based on a static nonlinear change rate when the static base coefficient is updated, temporarily save a dynamic base coefficient learning step, reduce the dynamic base coefficient learning step in proportion, and reduce the dynamic base coefficient learning step according to a performance index.

[0091] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting; it is not intended to exclude myriad other embodiments of the present application that other present or future devices, platforms, modules, components and systems can utilize, nor is it intended to exclude prior art contrary to this application. It must be noted that as used herein, the singular form "a", "an" and "the" include plural references unless the context clearly dictates otherwise. As such, the terms "comprise" (and grammatical variations thereof, such as "comprising" and "comprises"), "have" (and grammatical variations thereof, such as "having" and "has"), "include" (and grammatical variations thereof, such as "including" and "includes") or the like, are used herein not to limit the component, element, or method or process or steps as defined or implicit by such terms, but rather only to discern that such components or elements or methods or processes or steps are either included in, or encompassed by, the described embodiments. As used herein, "and / or" means and. Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. It will be further understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art and the present disclosure, and will not be interpreted in an overly literal or overly formal sense unless expressly so defined herein.

[0092] The preferred embodiments of the application disclosed above are only to help explain the principles of the present application. The preferred embodiments do not describe all the details of the present application, nor limit the present application to the specific embodiments. It is apparent that many modifications and variations can be made to the present application according to the contents of the present specification. The present specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present application, so that those skilled in the art can well understand and utilize the present application. Any modifications and variations that do not deviate from the structure of the present application or exceed the scope defined by the present application shall be within the scope of protection of the present application.

Claims

1. A digital signal compensation method for optical modules, characterized in that, include: During the initialization phase, the input ideal driving signal is divided into static basis and dynamic basis, and static basis matrix and dynamic basis matrix are constructed respectively. The number of static basis in the static basis matrix is ​​determined based on the input amplitude and hardware resource parameters, and the static basis coefficients are obtained. The number of dynamic basis is dynamically adjusted based on the channel performance parameters to obtain the dynamic basis coefficients. During the initialization phase, the input signal is decomposed into a static basis subspace and a dynamic basis subspace. The dynamic basis is orthogonalized by QR decomposition, and the overlapping part with the static basis is removed. The coefficients of the dynamic basis are then sparsified to obtain a sparse set. During the operation phase, the digital signal sequence from the optical module ADC is received. Based on the static basis coefficients obtained in the initialization phase, the static compensation component is calculated. Based on the dynamic basis coefficients and sparse set, the dynamic compensation component is calculated. The static compensation component and the dynamic compensation component are superimposed to obtain the output compensation signal. During the periodic update phase, when the static base coefficients are updated, adaptive adjustment is performed based on the static nonlinear change rate. At the same time, the dynamic base coefficient learning step size is temporarily saved, scaled down proportionally, and then reduced back according to the performance index.

2. The digital signal compensation method for optical modules according to claim 1, characterized in that: The specific process of dividing the input ideal driving signal into static basis and dynamic basis is as follows: The static basis refers to a basis that relies only on the nonlinear characteristics of the input signal at the current moment, without considering the influence of previous and subsequent signs, has no delay term, and only describes the nonlinear influence of the instantaneous amplitude of the input on the output. The dynamic basis refers to a combination of the current and past moments of the input signal, used to describe memory effects and inter-symbol crosstalk, including a delay term, and capable of capturing dynamic nonlinearity, inter-symbol interference and frequency-selective distortion. The sampling frequency is determined based on the transmission rate and signal bandwidth of the optical module. The time interval between sampling points is fixed by the sampling frequency. Each sampling point corresponds to the signal amplitude at a time position. The sampling point is the basic unit for subsequent feature extraction, static basis and dynamic basis construction. Feature extraction is performed on the sampling point sequence of the input signal to obtain each candidate feature sequence. The autocorrelation coefficient of each candidate feature sequence is calculated, and the relationship between each candidate feature sequence and its own delayed version is quantified by the autocorrelation function. The autocorrelation coefficients of all non-zero delays are summed and compared with the autocorrelation coefficients of zero delays to obtain the delay dependency index of each candidate feature sequence. The delay dependency index reflects the degree of influence of past symbols on each candidate feature sequence. If the delay dependency of a candidate feature sequence is less than the preset delay dependency threshold, the candidate feature sequence is determined to be a static basis. If the delay dependency of a candidate feature sequence is greater than or equal to the preset delay dependency threshold, the candidate feature sequence is determined to be a dynamic basis.

3. The digital signal compensation method for optical modules according to claim 1, characterized in that: The specific process for constructing the static basis matrix and the dynamic basis matrix is ​​as follows: Candidate feature sequences that are classified as static bases are extracted and denoted as static feature sequences. The values ​​of the static feature sequences at each sampling point are arranged sequentially as a column of the static basis matrix, and the sampling point sequence is used as a row of the static basis matrix. The static basis matrix is ​​used to describe instantaneous nonlinear disturbances. Candidate feature sequences that are classified as dynamic basis are extracted and denoted as dynamic feature sequences. The maximum delay length and nonlinear order are limited. The values ​​of the dynamic feature sequences at each sampling point are taken as a column of the dynamic basis matrix, and the sampling point sequence is taken as a row of the dynamic basis matrix. The dynamic basis matrix is ​​used to characterize dynamic nonlinear interference.

4. The digital signal compensation method for optical modules according to claim 1, characterized in that: The process of determining the number of static bases of the static basis matrix and obtaining the static base coefficients based on the input amplitude and hardware resource parameters specifically includes: Obtain hardware resource parameters, including available storage capacity, number of available multiply-accumulate units, available throughput, and latency budget; The performance target parameters are obtained, including the upper limit of error, the upper limit of bit error rate, and the magnitude of the error vector. The unit weighting factors of the performance target parameters corresponding to the performance target parameters are extracted from the database, including the unit weighting factors of the upper limit of error, the upper limit of bit error rate, and the magnitude of the error vector. The performance target parameters are multiplied one by one by the unit weighting factors of the performance target parameters and then coupled to obtain the performance target index. The unit weighting factors of the performance target parameters are used to eliminate the dimensions and normalize the magnitude of different performance target parameters, so that the performance target parameters are on a comparable scale, and at the same time, the weight of each performance target parameter in the performance target index is adjusted. The performance target indicators are input into the preset performance target indicator-hardware target resource parameter mapping set in the database and mapped and matched to obtain the hardware target resource parameters, including the target available storage capacity, the target available number of multiply-accumulate units, the target available throughput, and the target latency budget. The hardware resource parameters are compared with the hardware target resource parameters and then weighted and coupled to obtain the hardware resource characterization value. The input amplitude refers to the range of the sampled signal amplitude before entering the digital compensation of the optical module, that is, the magnitude distribution of the digital signal acquired from the ADC. The input amplitude is statistically analyzed. For the sampled digital signal, the distribution parameters of the input amplitude are obtained, including the maximum amplitude, minimum amplitude, and average amplitude. Based on the distribution parameters of the input amplitude, the number of initial static bases required for each amplitude interval is determined. The upper limit of the number of static bases is obtained by inputting the hardware resource representation value into the pre-stored mapping set of hardware resource representation value - upper limit of the number of static bases in the database and performing mapping matching. If the initial number of static bases is less than the upper limit of the number of static bases, then the initial number of static bases is used as the final number of static bases and applied accordingly. If the initial number of static bases is greater than or equal to the upper limit of the number of static bases, then the upper limit of the number of static bases is used as the final number of static bases and applied accordingly. The ideal target response of the optical module is taken as the ideal output. The static basis matrix is ​​input into the optical module to obtain the actual output. By minimizing the error between the actual output and the ideal output, the coefficients of each static basis function are solved and denoted as static basis coefficients. The static basis coefficients are the optimal weights of the static basis functions when fitting the ideal output signal of the optical module.

5. The digital signal compensation method for optical modules according to claim 1, characterized in that: The dynamic adjustment of the number of dynamic bases based on channel performance parameters to obtain dynamic base coefficients specifically includes: Extract the hardware resource representation value, input it into the pre-stored mapping set of hardware resource representation value - dynamic base number upper limit in the database, and perform mapping matching to obtain the dynamic base number upper limit; The channel performance parameters include the channel nonlinearity order, bit error rate, and signal-to-noise ratio. The unit weighting factor of the channel performance parameters is extracted from the database. The channel performance parameters are multiplied by the unit weighting factor and then coupled to obtain the channel performance constraint index. Based on the channel performance constraint index, the dynamic basis number is adjusted by multiplying by the initial dynamic basis number. When the adjusted dynamic basis number is greater than or equal to the upper limit of the dynamic basis number, the upper limit of the dynamic basis number is used as the final dynamic basis number. When the adjusted dynamic basis number is less than the upper limit of the dynamic basis number, the adjusted dynamic basis number is used as the final dynamic basis number. The dynamic basis matrix is ​​fitted to the static error residual signal using the least squares method. The static error residual signal is the signal after removing the static basis error part. The coefficients of each dynamic basis function are obtained by solving the problem and are denoted as dynamic basis coefficients. The dynamic basis coefficients are used to reflect the contribution of the corresponding dynamic basis function to the dynamic nonlinear compensation.

6. The digital signal compensation method for optical modules according to claim 1, characterized in that: The step of decomposing the input signal into a static basis subspace and a dynamic basis subspace, and performing QR decomposition orthogonalization on the dynamic basis, specifically includes: The input signal is projected into the static basis subspace. The input signal is the actual acquired signal containing nonlinear interference used for initialization. The input signal is fitted based on the static basis matrix and static basis coefficients to obtain the static basis output signal. The static basis output signal is subtracted from the input signal to obtain the residual signal. The residual signal is the dynamic basis fitting object. The dynamic basis matrix is ​​orthogonalized to the static basis subspace. The column vectors of the dynamic basis are transformed into mutually orthogonal vector groups through QR decomposition, ensuring that the dynamic basis and the static basis do not overlap.

7. The digital signal compensation method for optical modules according to claim 1, characterized in that: The process of sparsifying the dynamic basis coefficients to obtain a sparse set specifically includes: The performance target index is input into the pre-stored performance target index-coefficient threshold adjustment value mapping set in the database and the mapping matching is performed to obtain the coefficient threshold adjustment value in the sparsity processing. The coefficient threshold is used to filter the retained dynamic basis functions. The coefficient threshold adjustment value is corrected with the preset initial coefficient threshold in the database to obtain the coefficient threshold. Each dynamic basis coefficient is compared with a coefficient threshold. If the dynamic basis coefficient of a dynamic basis function is less than the coefficient threshold, the dynamic basis function is filtered out. If the dynamic basis coefficient of a dynamic basis function is greater than or equal to the coefficient threshold, the dynamic basis function is retained. The retained dynamic basis functions and their corresponding coefficients are combined into a sparse set.

8. The digital signal compensation method for optical modules according to claim 1, characterized in that: The specific processing conditions for superimposing the static compensation component and the dynamic compensation component to obtain the output compensation signal are as follows: The digital signal sequence sampled in real time is obtained from the ADC of the optical module. Each sampling point corresponds to the signal amplitude at a specific time position. The digital signal sequence refers to the original signal that needs to be compensated for due to the static and dynamic nonlinear distortion of the optical module. Substitute the digital signal sequence into the static basis matrix obtained in the initialization stage, and obtain the static compensation component by weighted summation of the static basis coefficients. The static compensation component is used to compensate for static nonlinear distortion in the signal. The digital signal sequence is mapped to the dynamic basis matrix, and the dynamic compensation component is obtained by weighted summation of the dynamic basis coefficients in the sparse set. The dynamic compensation component is used to compensate for the dynamic nonlinear distortion in the signal. The static compensation component and the dynamic compensation component are superimposed at the corresponding sampling points to generate a complete output compensation signal, which is a digital signal after static nonlinear compensation and dynamic nonlinear compensation.

9. The digital signal compensation method for optical modules according to claim 1, characterized in that: The periodic updates specifically include: When outputting the compensation signal, the error is calculated by comparing it with the uncompensated output signal to obtain the compensation difference. Based on the rate of change of the compensation difference in the time series within the operating cycle, it is determined whether the static nonlinear characteristics have drifted. If the static nonlinear characteristic drifts, the adaptive update of the static basis coefficients is immediately triggered; if the static nonlinear characteristic does not drift, it waits for periodic updates. After the static basis coefficients are updated, the dynamic basis coefficients are fine-tuned using the compensation difference. Specifically, the compensation difference is fitted by the least squares method, the compensation difference is decomposed into the dynamic basis subspace, and the dynamic basis coefficients are updated. When updating the static base coefficients, the learning step size of the dynamic base coefficients is temporarily saved. The learning step size of the dynamic base coefficients refers to the allowable adjustment range of the coefficients when updating the dynamic base coefficients. Based on the preset safety ratio in the database, the current dynamic base coefficient learning step size is reduced to decrease the response speed of the dynamic base to the input signal. After reducing the dynamic base coefficient learning step size, the signal distortion index is monitored. If the signal distortion index exceeds the signal distortion index threshold, the difference between the signal distortion index and the signal distortion index threshold is used to obtain the signal distortion deviation. Based on the signal distortion deviation, the mapping set of signal distortion deviation and backoff amplitude pre-stored in the database is used for mapping and matching to obtain the backoff amplitude and perform backoff to further reduce the dynamic base coefficient learning step size. After the static base coefficient is updated, the dynamic base coefficient is restored based on the temporarily saved dynamic base coefficient learning step size.

10. A system applying the digital signal compensation method for optical modules as described in any one of claims 1-9, characterized in that: The initialization module is used to divide the input ideal driving signal into static basis and dynamic basis during the initialization phase, construct static basis matrix and dynamic basis matrix respectively, determine the number of static basis in the static basis matrix and obtain static basis coefficients based on input amplitude and hardware resource parameters, dynamically adjust the number of dynamic basis based on channel performance parameters to obtain dynamic basis coefficients, decompose the input signal into static basis subspace and dynamic basis subspace, perform QR decomposition orthogonalization on the dynamic basis, remove the overlapping part with the static basis, and perform sparsification processing on the dynamic basis coefficients to obtain a sparse set; The operation module is used to receive the digital signal sequence from the optical module ADC, calculate the static compensation component based on the static basis coefficients obtained in the initialization stage, calculate the dynamic compensation component based on the dynamic basis coefficients and sparse set, and superimpose the static compensation component and the dynamic compensation component to obtain the output compensation signal. The update module is used to adaptively adjust the static nonlinear change rate when updating the static base coefficients, while temporarily saving the dynamic base coefficient learning step size, scaling it down proportionally and then backing it down according to the performance index.

Images