A testing system and method for optical interconnect links for computing power clusters

Abstract

This invention relates to the field of optical communication testing technology and discloses an optical interconnect link testing system and method for computing power clusters. The optical interconnect link testing method for computing power clusters acquires wavelength configuration data, broadband spectral data, and bidirectional polarization state evolution trajectory data of a DWDM system. It then uses a narrowband filter window sequence to separate the independent spectral curves of each wavelength channel, extracts single-wavelength spectral fingerprint vectors and calculates the coupling relationship matrix between channels, calculates bidirectional polarization dynamic characteristics and generates polarization symmetry difference vectors, calculates the cross-correlation coefficients of spectral and polarization characteristics to generate a channel-level coupling degree matrix, analyzes the temporal causal relationship of coupling faults, matches a composite fault mode library to generate fault classification results, and finally outputs a spectral-polarization joint diagnostic report. This achieves accurate identification and root cause diagnosis of spectral degradation and polarization degradation coupling faults in DWDM coherent optical systems.

Description

Technical Field

[0001] This invention relates to the field of optical communication testing technology, and more specifically, to a testing system and method for optical interconnect links for computing power clusters. Background Technology

[0002] Dense wavelength division multiplexing (DWDM) coherent optical communication systems are widely used for data transmission in high-speed computing clusters, achieving 400G / 800G high-speed links by transmitting dozens of wavelength channels in a single optical fiber. The spacing between each wavelength channel is typically 0.8 nm or less, and the signal simultaneously utilizes the amplitude, phase, and polarization state of light to carry information. Physical effects such as polarization mode dispersion (PMD) and polarization dependent loss (PDL) exist in the optical fiber link, leading to differences in the polarization state evolution path of bidirectional transmission. Each wavelength channel may experience two independent or coupled failure modes: spectral degradation and polarization state degradation.

[0003] Currently, fault diagnosis of DWDM systems mainly adopts broadband spectral analysis and power monitoring methods. Broadband spectral data covering all wavelength channels is obtained through a spectral analyzer, and the power value of each channel is measured through an optical power meter. Fault determination is then performed by combining the signal-to-noise ratio differential method.

[0004] The existing technology has the following drawbacks: First, the spectral characteristics of each wavelength channel are highly similar and overlap, making it difficult for broadband spectral fingerprints to distinguish the independent degradation of a single wavelength channel; Second, nonlinear crosstalk between adjacent channels produces spectral characteristic changes similar to device degradation, making it difficult to determine the fault source; Third, traditional power or signal-to-noise ratio differential methods cannot detect polarization-state related bidirectional asymmetry; Fourth, degradation of a single wavelength channel may involve both spectral feature shift and polarization-state asymmetric drift. Rapid polarization-state drift can cause instantaneous fluctuations in spectral measurements, and spectral distortion can also affect the accuracy of polarization-state estimation. When these two types of degradation are coupled, single-dimensional analysis is prone to misjudgment and cannot accurately locate the root cause of the fault. Summary of the Invention

[0005] This invention provides a testing system and method for optical interconnect links for computing power clusters, which solves the technical problems of insufficient accuracy in fault diagnosis and inability to effectively identify spectral and polarization coupling faults in related technologies.

[0006] This invention discloses a testing system and method for optical interconnect links in computing power clusters, comprising the following steps: acquiring wavelength configuration data and broadband spectral data of a DWDM system; simultaneously acquiring Stokes parameter time-series data of bidirectional transmission signals from coherent transceiver modules at both ends of the optical link to generate a bidirectional polarization state evolution trajectory dataset; based on the center wavelength and spacing parameters of each wavelength channel, applying a narrowband filtering window sequence to separate the broadband spectral data to generate independent spectral curve sets for each wavelength channel; extracting the center wavelength offset, power value, and sideband steepness from the independent spectral curves of each wavelength channel to generate single-wavelength spectral fingerprint vector sets; calculating the correlation coefficient of spectral fingerprint vectors of adjacent wavelength channels to generate a coupling relationship matrix between channels. The process involves calculating the drift rate, drift range, and drift periodicity of the polarization state on the Poincaré sphere for each bidirectional polarization state evolution trajectory, generating a bidirectional polarization dynamic feature vector; calculating the difference between the bidirectional polarization dynamic feature vectors to generate a polarization symmetry difference vector; calculating the cross-correlation coefficient between the temporal variation of the single-wavelength spectral fingerprint of each wavelength channel and the temporal variation of the polarization dynamic feature, identifying strongly coupled and weakly coupled spectral-polarization channels based on the cross-correlation coefficient threshold, and generating a channel-level spectral-polarization coupling degree matrix; connecting the inter-channel coupling relationship matrix with the polarization symmetry difference vector to construct a spectral-polarization multidimensional fault feature vector, matching it with a preset composite fault mode library, generating fault classification results, and outputting a spectral-polarization joint diagnostic report.

[0007] Furthermore, the narrowband filtering window sequence refers to a set of bandpass filtering functions centered on the center wavelength of each wavelength channel and with the channel interval as the window width. The bandpass filtering function adopts a super-Gaussian window, and the window width parameter of the super-Gaussian window ranges from 0.3 to 0.5 times the channel interval, while the super-Gaussian order ranges from 2 to 6. Before applying the narrowband filtering window, the broadband spectral data is preprocessed by deconvolution, and the broadband spectral data is deconvolved using Wiener filtering or the Richardson-Lucy iterative algorithm.

[0008] Furthermore, the sideband steepness is calculated as follows: determine the wavelength positions where the power drops to 50% of the peak power on both sides of the spectral peak position, calculate the slope on the left and the slope on the right respectively, and take the average of the absolute values ​​of the slopes on both sides as the sideband steepness; the elements of the inter-channel coupling relationship matrix are the Pearson correlation coefficients of the spectral fingerprint vectors of the two corresponding wavelength channels.

[0009] Furthermore, the drift rate is calculated as follows: the arc length displacement of the polarization state point on the Poincaré sphere at adjacent time points is calculated, the arc length displacement at each time point is removed to obtain the instantaneous drift rate by time interval, and the average value of all instantaneous drift rates is calculated; the drift range is calculated as follows: the centroid coordinates of the polarization state trajectory point sequence are calculated, the spherical angular distance between each trajectory point and the centroid is calculated, and the maximum value of all spherical angular distances is taken; the drift periodicity characteristic is calculated as follows: autocorrelation analysis is performed on the polarization state time series data, and the time delay corresponding to the first local maximum value in the autocorrelation function is searched as the principal periodic component.

[0010] Furthermore, when calculating the polarization dynamic characteristics, the distribution ellipticity and principal axis orientation angle of the polarization trajectory on the Poincaré sphere are also extracted. The distribution ellipticity is characterized by the ratio of the maximum eigenvalue to the minimum eigenvalue obtained by principal component analysis of the coordinates of the polarization trajectory points. The principal axis orientation angle is characterized by the azimuth angle of the first principal component vector obtained by principal component analysis in the Poincaré spherical coordinate system. The distribution ellipticity is related to PDL, and the principal axis orientation angle is related to the principal polarization state of PMD.

[0011] Furthermore, the cross-correlation coefficient is calculated as follows: Pearson correlation coefficients are calculated between each component of the spectral fingerprint vector time series and each component of the polarization dynamic feature vector time series to obtain a sub-correlation coefficient matrix. The maximum absolute value of all elements in the sub-correlation coefficient matrix is ​​taken as the cross-correlation coefficient. When the cross-correlation coefficient is greater than a preset threshold, it is determined to be a strongly coupled spectral-polarization channel. When the cross-correlation coefficient is less than or equal to the preset threshold, it is determined to be a weakly coupled channel. The preset threshold ranges from 0.6 to 0.8.

[0012] Furthermore, the method also includes a step to analyze the temporal causal relationship of coupling faults: for the spectral-polarization strongly coupled channel, principal component analysis is used to extract the spectral comprehensive characteristic time series and the polarization comprehensive characteristic time series respectively, calculate the cross-correlation function of the two comprehensive characteristic time series, and determine the time delay corresponding to the peak value of the cross-correlation function; if the time delay is greater than zero, it is determined that the device degradation is dominant; if the time delay is less than zero, it is determined that the link polarization effect is dominant; if the absolute value of the time delay is less than the synchronization determination threshold, it is determined that the environmental factors are dominant, and the value range of the synchronization determination threshold is 2 to 5 sampling points.

[0013] Furthermore, the Granger causality test method is used to analyze the time-series lead-lag relationship. The Granger causality test method constructs a vector autoregressive model of spectral features and polarization features, respectively constructing an autoregressive model that only contains historical information of target features and a vector autoregressive model that contains historical information of both types of features. The prediction residuals of the two models are compared, and the F test is used to determine whether the improvement in prediction accuracy is statistically significant.

[0014] Furthermore, the composite fault mode library includes feature templates for the following fault modes: single-channel device faults are characterized by abnormal spectral fingerprints of a single wavelength channel and coupling coefficients with adjacent channels below 0.3; crosstalk faults are characterized by synchronous changes in the spectral fingerprints of at least two adjacent wavelength channels and coupling coefficients above 0.7; PMD-dominated faults are characterized by drift rate differences and periodicity differences in the polarization symmetry difference vector exceeding a preset multiple standard deviation of the normal fluctuation range; PDL-dominated faults are characterized by drift range differences in the polarization symmetry difference vector exceeding a preset multiple standard deviation of the normal fluctuation range; when matching the composite fault mode library, a multi-label classification neural network is used, and the output layer of the multi-label classification neural network uses a Sigmoid activation function to output the matching confidence of each fault mode.

[0015] This invention discloses an optical interconnect link testing system for computing power clusters, comprising: a multi-source data acquisition module for acquiring wavelength configuration data and broadband spectral data of a DWDM system, simultaneously acquiring Stokes parameter time-series data of bidirectional transmission signals from coherent transceiver modules at both ends of the optical link, and generating a bidirectional polarization state evolution trajectory dataset; a spectral separation module for separating the broadband spectral data using a narrowband filtering window sequence based on the center wavelength and spacing parameters of each wavelength channel, generating an independent spectral curve set for each wavelength channel; and a spectral feature extraction module for extracting the center wavelength offset, power value, and sideband steepness from the independent spectral curves of each wavelength channel, generating a single-wavelength spectral fingerprint vector set, and calculating the correlation coefficient of spectral fingerprint vectors of adjacent wavelength channels to generate a channel coupling relationship matrix; polarization The feature extraction module calculates the drift rate, drift range, and drift periodicity of the polarization state on the Poincaré sphere for the bidirectional polarization evolution trajectory, generates a bidirectional polarization dynamic feature vector, and calculates the difference between the bidirectional polarization dynamic feature vectors to generate a polarization symmetry difference vector. The coupling degree calculation module calculates the cross-correlation coefficient between the temporal changes of the single-wavelength spectral fingerprint of each wavelength channel and the temporal changes of the polarization dynamic feature, identifies strongly coupled and weakly coupled spectral-polarization channels based on the cross-correlation coefficient threshold, and generates a channel-level spectral-polarization coupling degree matrix. The fault diagnosis module connects the inter-channel coupling relationship matrix with the polarization symmetry difference vector to construct a spectral-polarization multi-dimensional fault feature vector, matches it with a preset composite fault mode library, generates fault classification results, and outputs a spectral-polarization joint diagnostic report.

[0016] This invention separates broadband spectral data using a narrowband filtering window sequence, independently extracting the spectral features of each wavelength channel, thus solving the technical problem of indistinguishable single-channel degradation due to spectral overlap in dense wavelength channels. It distinguishes between single-channel independent faults and multi-channel crosstalk faults by calculating the inter-channel coupling relationship matrix, solving the technical problem of misjudgment of fault sources due to the similarity between nonlinear crosstalk and device degradation characteristics. It characterizes the asymmetry of polarization correlation effects by extracting a bidirectional polarization symmetry difference vector, solving the technical problem that traditional methods cannot detect bidirectional asymmetry of polarization state correlation. Finally, it solves the technical problem of misjudgment caused by single-dimensional analysis when spectral degradation and polarization degradation are coupled by calculating the cross-correlation coefficient between spectral and polarization features and analyzing the temporal causal relationship, achieving accurate classification and root cause localization of composite faults in DWDM coherent optical systems. Attached Figure Description

[0017] Figure 1 This is a flowchart of a testing method for optical interconnect links for computing power clusters provided in an embodiment of the present invention. Detailed Implementation

[0018] Dense wavelength division multiplexing (DWDM) coherent optical communication systems are widely used for data transmission in high-speed computing clusters, achieving 400G / 800G high-speed links by transmitting dozens of wavelength channels in a single optical fiber. The spacing between each wavelength channel is typically 0.8 nm or less, and the signal simultaneously utilizes the amplitude, phase, and polarization state of light to carry information. Physical effects such as polarization mode dispersion (PMD) and polarization dependent loss (PDL) exist in the optical fiber link, leading to differences in the polarization state evolution path of bidirectional transmission. Each wavelength channel may experience two independent or coupled failure modes: spectral degradation and polarization state degradation.

[0019] Current DWDM system fault diagnosis faces the following technical challenges: First, the spectral characteristics of each wavelength channel are highly similar and overlap, making it difficult to distinguish the independent degradation of a single wavelength channel using broadband spectral fingerprinting. Second, nonlinear crosstalk between adjacent channels produces spectral changes similar to device degradation, making fault source identification difficult. Third, traditional power or signal-to-noise ratio differential methods cannot detect polarization-related bidirectional asymmetry. Fourth, degradation of a single wavelength channel may involve both spectral feature shift and polarization-state asymmetric drift. Rapid polarization-state drift can cause instantaneous fluctuations in spectral measurements, and spectral distortion can also affect the accuracy of polarization-state estimation. When these two types of degradation are coupled, single-dimensional analysis is prone to misjudgment.

[0020] like Figure 1 As shown, the optical interconnect link testing method for computing power clusters provided in this embodiment of the invention includes the following steps:

[0021] Step 1: Acquire multi-source monitoring data from the DWDM system;

[0022] Acquire wavelength configuration data and broadband spectral data of the DWDM system, and simultaneously acquire Stokes parameter timing data of bidirectional transmission signals from coherent transceiver modules at both ends of the optical link to generate a bidirectional polarization state evolution trajectory dataset.

[0023] It should be noted that the wavelength configuration data includes the center wavelength of each wavelength channel. and channel interval ,in For wavelength channel indexing; broadband spectral data is a continuous spectral power distribution covering all wavelength channels. Stokes parameter timing data includes Stokes parameters for both forward and reverse transmission. , , , Time series within the sampling period.

[0024] Furthermore, Stokes parameters This represents the total light intensity of the optical signal. This represents the intensity difference between horizontally polarized and vertically polarized light. This represents the intensity difference between light polarized at 45 degrees and light polarized at 135 degrees. These four Stokes parameters, representing the intensity difference between left-handed and right-handed circularly polarized light, fully describe the polarization state of the optical signal.

[0025] It should be noted that the bidirectional polarization state evolution trajectory refers to the sequence of trajectory points formed by mapping the normalized forward propagation Stokes parameters onto the Poincaré sphere. and the trajectory point sequence formed by reverse transmission ,in For time indexing, , , .

[0026] Furthermore, the Poincaré sphere is a three-dimensional unit sphere, and any point on the Poincaré sphere corresponds to a polarization state. By using the three components of the Stokes parameter after normalization as the three coordinate axis values ​​of the spherical coordinate system, the polarization state can be mapped to a point on the Poincaré sphere, and the change of the polarization state over time forms a continuous trajectory on the Poincaré sphere.

[0027] In this embodiment of the application, in order to ensure the time alignment accuracy of spectral data and polarization state data, broadband spectral data and Stokes parameter time series data are marked with a unified timestamp, and a linear interpolation algorithm is used to eliminate the time offset caused by the sampling frequency difference, thereby generating a time-synchronized multi-source monitoring dataset.

[0028] Step 2: Separate the independent spectral curves of each wavelength channel;

[0029] Based on the center wavelength and spacing parameters of each wavelength channel, the broadband spectral data is separated by a narrowband filter window sequence to generate independent spectral curve sets for each wavelength channel.

[0030] It should be noted that the narrowband filter window sequence refers to the sequence based on the center wavelength of each wavelength channel. Centered on, with channels as intervals A set of bandpass filtering functions for the window width For broadband spectral data Perform convolution operations with each filter window to obtain the first... Independent spectral curves of each wavelength channel .

[0031] It should be noted that the filtering window function Rectangular windows, Gaussian windows, or super-Gaussian windows can be used, with the expression for a super-Gaussian window being:

[0032]

[0033] in Represents an exponential function. For window width parameter, It is a super-Gaussian order.

[0034] Furthermore, the window width parameter The value range is the channel interval. 0.3 to 0.5 times that of the super-Gaussian order The value ranges from 2 to 6; a larger super-Gaussian order is chosen when the channel spacing is small. The value is used to improve the selectivity of the filter window.

[0035] Furthermore, a rectangular window refers to a window function that takes a value of 1 within a fixed width range near the center wavelength and a value of 0 outside the range. A Gaussian window refers to a filtering window whose window function values ​​follow a Gaussian distribution. A super-Gaussian window narrows the transition region at the edge of the window by introducing a higher-order exponential term, thereby providing better channel isolation when the interval between adjacent channels is small.

[0036] In this embodiment of the application, in order to reduce the crosstalk effect in the overlapping region of adjacent channel spectra, the broadband spectral data is deconvolved before applying the narrowband filter window, and the original spectrum is deconvolved using the superposition response function of the filter window of each channel, thereby improving the purity of the independent spectral curves of each channel after separation.

[0037] Furthermore, deconvolution preprocessing refers to treating broadband spectral data as the result of convolving the true spectrum of each channel with the system response function, and then performing deconvolution operations on the observed broadband spectral data using Wiener filtering or the Richardson-Lucy iterative algorithm to restore the true spectral distribution of each channel, thereby eliminating spectral broadening and distortion caused by the overlap of the system response function and adjacent channels.

[0038] Step 3: Extract single-wavelength spectral fingerprint vectors and calculate inter-channel coupling relationships;

[0039] The center wavelength shift, power value, and sideband steepness are extracted from the independent spectral curves of each wavelength channel to generate a single-wavelength spectral fingerprint vector set; the correlation coefficient of the spectral fingerprint vectors of adjacent wavelength channels is calculated to generate the coupling relationship matrix between channels.

[0040] It should be noted that the center wavelength offset It refers to the first Measured spectral peak positions and nominal center wavelengths of each wavelength channel The difference; power value It refers to the first Peak power of the spectral curves for each wavelength channel; sideband steepness The sideband kurtosis refers to the average slope of the falling edges on both sides of the peak of the spectral curve. The specific calculation method is as follows: First, determine the wavelength positions where the power drops to 50% of the peak power on both sides of the spectral peak position, and record them as the left position. and the right side position ;

[0041] Then calculate the slope on the left side. and the slope on the right ,in The peak wavelength is used as the reference value, and the average of the absolute values ​​of the slopes on both sides is taken as the sideband steepness. The single-wavelength spectral fingerprint vector is represented as... ,in This indicates transpose.

[0042] Furthermore, the center wavelength offset reflects the degree of frequency drift of the wavelength channel, the power value reflects the signal strength of the wavelength channel, and the sideband steepness reflects the degree of distortion of the spectral shape. The smaller the sideband steepness value, the smoother the spectral edge and the more degraded the filter characteristics.

[0043] It should be noted that the inter-channel coupling matrix elements For the first The and the first Pearson correlation coefficients of spectral fingerprint vectors for each wavelength channel:

[0044]

[0045] in For covariance calculation, The standard deviation is denoted as .

[0046] Furthermore, the range of values ​​for the Pearson correlation coefficient is as follows: The closer the Pearson correlation coefficient is to 1, the more consistent the spectral fingerprint trends of the two wavelength channels are; the closer the Pearson correlation coefficient is to -1, the more opposite the trends are; and the closer the Pearson correlation coefficient is to 0, the more independent the spectral changes of the two channels are.

[0047] In this embodiment of the application, in order to identify the four-wave mixing crosstalk effect between non-adjacent channels, when calculating the inter-channel coupling relationship matrix, in addition to adjacent channels, the correlation coefficient between wavelength channels spaced one or two channel positions apart is also calculated, thereby generating an extended multi-order inter-channel coupling relationship matrix.

[0048] Furthermore, four-wave mixing refers to the interaction of light waves from multiple wavelength channels under the action of optical fiber nonlinear effects to generate new frequency components. These new frequency components may fall within the frequency range of other wavelength channels, forming crosstalk. The intensity of the four-wave mixing effect is related to the frequency spacing between channels. Therefore, there may also be significant crosstalk coupling between non-adjacent channels.

[0049] Step 4: Calculate the bidirectional polarization dynamic characteristics and generate the polarization symmetry difference vector;

[0050] For the evolution trajectory of the bidirectional polarization state, the drift rate, drift range, and drift periodicity characteristics of the polarization state on the Poincaré sphere are calculated respectively to generate a bidirectional polarization dynamic feature vector; the difference between the bidirectional polarization dynamic feature vectors is calculated to generate a polarization symmetry difference vector.

[0051] It should be noted that the drift rate The calculation method is as follows: for the sequence of polarization state trajectory points Calculate adjacent time intervals and The arc length displacement of the polarization point on the Poincaré sphere superscript Indicates time The value is then used to remove the arc length bits at each time interval. The instantaneous drift rate is obtained, and finally, the average of all instantaneous drift rates is taken to obtain the drift rate. ,in The total number of trajectory points. For the upper limit of summation; drift range The calculation method is as follows: First, calculate the centroid coordinates of the polarization state trajectory point sequence. ,in Given the total number of trajectory points, calculate the spherical angular distance between each trajectory point and the centroid. Finally, the maximum value of all spherical angular distances is taken as the drift range. ; Drift periodicity The calculation method is as follows: for the first component of the polarization state time series data Perform autocorrelation analysis on the sequence and calculate the autocorrelation function. ,in For time delay, To find the upper limit of the summation, we then search for the time delay corresponding to the first local maximum in the autocorrelation function; this time delay is the principal period component. The polarization dynamic eigenvector of forward propagation is represented as: The polarization dynamic eigenvector of reverse propagation is represented as ,in This indicates transpose.

[0052] Furthermore, arc length displacement The range of values ​​is Radius, drift range The range of values ​​is radians, time delay The search scope is One sampling point.

[0053] Furthermore, the drift rate reflects the speed of polarization state change; a larger drift rate indicates a more drastic change in polarization state. The drift range reflects the spatial span of polarization state change; a larger drift range indicates a more dispersed distribution of polarization state on the Poincaré sphere. The drift periodicity reflects whether there is a periodic pattern in the change of polarization state; the existence of drift periodicity is usually related to periodic changes in ambient temperature or mechanical vibration.

[0054] It should be noted that the polarization symmetry difference vector The calculation method is as follows:

[0055]

[0056] in This indicates the transpose. The polarization symmetry difference vector reflects the degree of asymmetric influence of polarization correlation effects on bidirectional transmission in the optical link.

[0057] Furthermore, in an ideal fiber optic link without polarization correlation effects, the polarization dynamic characteristics of forward and reverse transmission should be consistent, and each component of the polarization symmetry difference vector should be close to zero. When PDL or PMD effects exist in the optical link, the polarization state evolution paths experienced by bidirectional transmission are different, causing each component of the polarization symmetry difference vector to deviate significantly from zero.

[0058] In this embodiment of the application, in order to improve the ability of polarization state drift features to distinguish between PMD and PDL, the distribution ellipticity and principal axis direction angle of the polarization state trajectory on the Poincaré sphere are also extracted when calculating the polarization dynamic features, thereby generating an extended polarization dynamic feature vector. The distribution ellipticity is characterized by the ratio of the maximum eigenvalue to the minimum eigenvalue obtained by principal component analysis of the coordinates of the polarization state trajectory points. The principal axis direction angle is characterized by the azimuth angle of the first principal component vector obtained by principal component analysis in the Poincaré spherical coordinate system. The distribution ellipticity is related to PDL, and the principal axis direction angle is related to the principal polarization state of PMD.

[0059] Furthermore, principal component analysis decomposes the data matrix composed of the coordinates of the polarization trajectory points into eigenvalues ​​of the covariance matrix, obtaining the main direction of the data distribution and the variance in each direction. The eigenvector corresponding to the largest eigenvalue represents the principal axis direction of the data distribution, and the ratio of the largest eigenvalue to the smallest eigenvalue represents the ellipticity of the data distribution. The larger the ellipticity of the distribution, the more concentrated the polarization trajectory is in a certain direction.

[0060] Furthermore, the PDL effect causes optical signals of different polarization states to experience different losses, causing the polarization state trajectory to contract in a specific direction on the Poincaré sphere to form an elliptical distribution. The major axis of the elliptical distribution corresponds to the polarization state direction with the least loss. The PMD effect causes different propagation speeds of different polarization components, causing the polarization state to rotate on the Poincaré sphere along the principal polarization state axis. The principal axis direction angle reflects the direction of the principal polarization state.

[0061] Step 5: Calculate the coupling degree between spectral and polarization characteristics and generate a channel-level coupling degree matrix;

[0062] The cross-correlation coefficient between the temporal variation of the single-wavelength spectral fingerprint and the temporal variation of the polarization dynamic characteristics of each wavelength channel is calculated. Based on the threshold of the cross-correlation coefficient, the channels with strong spectral-polarization coupling and weak coupling are identified, and a channel-level spectral-polarization coupling degree matrix is ​​generated.

[0063] It should be noted that the temporal variation of a single-wavelength spectral fingerprint refers to the time series formed by the spectral fingerprint vectors of each wavelength channel within the monitoring period. The temporal variation of polarization dynamic characteristics refers to the time series of polarization dynamic characteristic vectors. Cross-correlation coefficients The calculation method is as follows: First, the spectral fingerprint vector time series Each component is related to the time series of polarization dynamic eigenvectors. The Pearson correlation coefficient was calculated for each component, and the results were obtained. The cross-correlation coefficient matrix is ​​obtained by taking the maximum absolute value of all elements in the matrix and then using it as the cross-correlation coefficient. ,in Indicates the first The first wavelength channel spectral fingerprint vector of the first wavelength channel Time series of each component The first eigenvector representing the polarization dynamic eigenvector Time series of each component This indicates the calculation of the Pearson correlation coefficient.

[0064] Furthermore, the cross-correlation coefficient reflects the correlation strength between the spectral characteristic changes and polarization dynamic characteristic changes of a certain wavelength channel. The larger the cross-correlation coefficient, the stronger the coupling between spectral degradation and polarization degradation of that wavelength channel.

[0065] It should be noted that when the number of cross-relationships... Greater than the preset threshold At that time, the judgment of the first Each wavelength channel is a strongly coupled spectral-polarization channel; when the cross-correlation coefficient is less than or equal to a preset threshold... At that time, it was determined to be a weakly coupled channel. Channel-level spectral-polarization coupling degree matrix. diagonal elements This represents the spectral-polarization coupling degree of each wavelength channel.

[0066] Furthermore, the cross-correlation threshold The value ranges from 0.6 to 0.8. When the channel spacing of the DWDM system is small or the link length is long, a lower cross-correlation coefficient threshold is selected to improve the detection rate of strongly coupled channels.

[0067] In this embodiment, to identify different types of coupling modes, the cross-correlation coefficients between each component of the spectral fingerprint and each component of the polarization dynamic feature are further calculated for the spectral-polarization strongly coupled channel, thereby generating a fine-grained spectral-polarization partial coupling degree matrix to distinguish three types of coupling modes: power coupling, wavelength drift coupling, and sideband distortion coupling. Among them, the power coupling mode is characterized by the strongest correlation between the power component of the spectral fingerprint and the polarization dynamic feature; the wavelength drift coupling mode is characterized by the strongest correlation between the center wavelength shift component of the spectral fingerprint and the polarization dynamic feature; and the sideband distortion coupling mode is characterized by the strongest correlation between the sideband steepness component of the spectral fingerprint and the polarization dynamic feature.

[0068] Furthermore, power-coupled faults are usually caused by the PDL effect, which causes optical signals in different polarization states to experience different losses. When the polarization state drifts, the channel power fluctuates accordingly. Wavelength-drift coupled faults are usually caused by temperature changes affecting both the filter center wavelength and fiber birefringence. Sideband distortion coupled faults are usually caused by the combined effect of filter aging and polarization-dependent filter response.

[0069] Step 6: Analyze the temporal causal relationships of coupled faults;

[0070] For strongly coupled spectral and polarization channels, the time-leading-lag relationship between changes in spectral characteristics and changes in polarization characteristics is analyzed. Based on the time-leading-lag relationship, the dominant fault factor is determined, and the causal inference results of coupled faults are generated.

[0071] It should be noted that the analytical method for the time series lead-lag relationship is as follows: for the first... Each spectral-polarization strongly coupled channel was used to generate its spectral fingerprint time series. Flattened as a scalar time series Principal component analysis was used to extract the first principal component as the time series of spectral comprehensive features, and the polarization dynamic feature time series was then analyzed. Similarly, the first principal component is extracted as the polarization-integrated characteristic time series through principal component analysis, and then the cross-correlation function of the two integrated characteristic time series is calculated. ,in For time delay, The length of the time series. To find the upper limit of the summation, determine the time delay corresponding to the peak value of the cross-correlation function. ;like This indicates that the spectral change leads the polarization change; if This indicates that the polarization change leads the spectral change; if Less than the preset synchronization judgment threshold This indicates that the two types of changes occur simultaneously.

[0072] Furthermore, time delay The search scope is Each sampling point is used to simultaneously determine the threshold. The value range is 2 to 5 sampling points. When the sampling frequency is high, a larger synchronization judgment threshold is selected to avoid misjudging small time deviations as timing lead-lag relationships.

[0073] Furthermore, the cross-correlation function Different time delays The value of the cross-correlation function reflects the correlation strength between the two time series at that time delay. The time delay corresponding to the peak of the cross-correlation function represents the time offset when the two time series reach the maximum correlation. The sign of this time offset reflects the direction of the causal relationship.

[0074] It should be noted that the fault-dominant factor determination rule based on the timing lead-lag relationship is as follows: if the spectral change leads, it is determined to be dominated by device degradation, which includes filter drift and amplifier gain distortion; if the polarization change leads, it is determined to be dominated by link polarization effect; if the change is synchronous, it is determined to be dominated by environmental factors, which include temperature change and mechanical vibration.

[0075] Furthermore, in failure scenarios dominated by device degradation, changes in device parameters first lead to abnormal spectral characteristics, and spectral distortion then affects polarization state measurement or influences polarization state evolution through nonlinear effects; in failure scenarios dominated by link polarization effects, changes in PMD or PDL in the optical fiber first lead to abnormal polarization state drift, and changes in polarization state affect spectral characteristics through polarization-related device responses or measurement errors; in failure scenarios dominated by environmental factors, temperature or vibration simultaneously act on the optical fiber link and optical devices, causing synchronous changes in spectral and polarization characteristics.

[0076] In this embodiment of the application, in order to improve the reliability of causal inference, the Granger causality test method is used when analyzing the lead-lag relationship of time series. The Granger causality test method constructs a vector autoregressive model of spectral features and polarization features to test the predictive ability of one type of feature time series to another type of feature time series, thereby obtaining a causal direction determination result with statistical significance.

[0077] Furthermore, the basic principle of the Granger causality test is: if the historical information of time series X can significantly improve the prediction accuracy of time series Y, then X is considered to be a Granger cause of Y. The Granger causality test is specifically implemented by constructing an autoregressive model that only contains the historical information of Y and a vector autoregressive model that contains the historical information of both X and Y, comparing the prediction residuals of the two models for Y, and using the F test to determine whether the improvement in prediction accuracy is statistically significant.

[0078] Step 7: Match the fault mode library and generate fault classification results;

[0079] By connecting the inter-channel coupling matrix with the polarization symmetry difference vector, a spectral-polarization multidimensional fault feature vector is constructed. After normalization preprocessing, the spectral-polarization multidimensional fault feature vector is matched with a preset composite fault mode library to distinguish single-channel device faults, crosstalk faults, PMD-dominated faults, and PDL-dominated faults, generating fault classification results.

[0080] It should be noted that the spectral-polarization multidimensional fault feature vector The construction method is to construct the inter-channel coupling relationship matrix. The upper triangular elements are expanded into a one-dimensional vector. , and the difference vector of polarization symmetry To splice:

[0081]

[0082] in This indicates transpose.

[0083] Furthermore, the coupling relationship matrix between channels is a symmetric matrix, and extracting only the upper triangular elements can avoid information redundancy. The upper triangular elements contain the coupling relationship information between all channel pairs.

[0084] It should be noted that the composite fault mode library contains feature templates for the following typical fault modes: single-channel device faults are characterized by abnormal spectral fingerprints of a single wavelength channel and low coupling coefficients with adjacent channels; crosstalk faults are characterized by synchronous changes in the spectral fingerprints of multiple adjacent wavelength channels and high coupling coefficients; PMD-dominated faults are characterized by significant drift rate differences and periodicity differences in the polarization symmetry difference vector; PDL-dominated faults are characterized by significant drift range differences in the polarization symmetry difference vector.

[0085] Furthermore, the criteria for determining single-channel device faults are that the deviation of the spectral fingerprint vector of a single wavelength channel from the historical reference value exceeds a preset abnormal threshold and the coupling coefficient between the wavelength channel and the adjacent channel is less than 0.3. The criteria for determining crosstalk faults are that the spectral fingerprint vectors of at least two adjacent wavelength channels change synchronously and the coupling coefficient between adjacent wavelength channels is greater than 0.7. The criteria for determining PMD-dominated faults are that the absolute value of the drift rate difference component in the polarization symmetry difference vector exceeds twice the standard deviation of the normal fluctuation range and the absolute value of the periodic difference component exceeds 1.5 times the standard deviation of the normal fluctuation range. The criteria for determining PDL-dominated faults are that the absolute value of the drift range difference component in the polarization symmetry difference vector exceeds twice the standard deviation of the normal fluctuation range.

[0086] Furthermore, the historical baseline value refers to the statistical mean of the spectral fingerprint vectors of each wavelength channel under normal system operation. The abnormal threshold is determined based on the statistical distribution of historical data, and the abnormal threshold is usually set to 3 times the standard deviation of the normal fluctuation range. The normal fluctuation range refers to the statistical distribution range of each component of the polarization symmetry difference vector under fault-free conditions. The normal fluctuation range is obtained through statistical analysis of historical monitoring data.

[0087] In this embodiment, to handle complex scenarios involving multiple types of concurrent faults, a multi-label classification neural network is used when matching the composite fault mode library. The input to the multi-label classification neural network is a normalized spectral-polarization multi-dimensional fault feature vector. The output of the multi-label classification neural network is the matching confidence of each fault mode, thereby identifying composite fault types.

[0088] Furthermore, the number of neurons in the input layer of a multi-label classification neural network is related to the spectral-polarization multidimensional fault feature vector. The dimensions are consistent. The hidden layer of the multi-label classification neural network adopts a two-layer fully connected layer structure. The number of neurons in the first hidden layer is twice the input dimension, and the number of neurons in the second hidden layer is once the input dimension. The activation function of the hidden layer is the ReLU function. The number of neurons in the output layer of the multi-label classification neural network is consistent with the number of fault mode types in the composite fault mode library. The output layer uses the Sigmoid activation function to make each output neuron independently output a confidence value between 0 and 1.

[0089] Furthermore, the ReLU function is expressed as follows: when the input value is greater than zero, the output is equal to the input value; when the input value is less than or equal to zero, the output is zero. The ReLU function can alleviate the gradient vanishing problem in deep neural networks. The Sigmoid activation function maps the output value to the interval between 0 and 1, so that the output of each output neuron can be interpreted as the probability or confidence of the existence of the corresponding fault mode.

[0090] Furthermore, the output of the multi-label classification neural network is the matching confidence value of each fault mode. The matching confidence value of each output neuron is compared with the preset fault judgment threshold. When the matching confidence value of a fault mode is greater than the fault judgment threshold, the fault mode is determined to exist. Thus, the continuous confidence value output by the multi-label classification neural network is decoded into discrete fault type judgment results.

[0091] Furthermore, the fault determination threshold ranges from 0.5 to 0.7. When a higher fault detection rate is required, a lower fault determination threshold is selected, and when a lower false alarm rate is required, a higher fault determination threshold is selected.

[0092] Step 8: Output a combined spectral-polarization diagnostic report;

[0093] Based on the comprehensive fault classification results, coupled fault causal inference results, and channel-level spectral-polarization coupling degree matrix, a joint spectral-polarization diagnostic report for the DWDM coherent optical system is output, indicating the confidence level of each fault type and the location of the degraded channel.

[0094] It should be noted that the spectral-polarization joint diagnostic report includes the following: a list of fault types and the matching confidence level of each type; an index of degraded channels and the degree of spectral fingerprint anomaly of each degraded channel; a list of spectral-polarization coupling channels and the coupling degree value of each spectral-polarization coupling channel; causal inference conclusions of coupling faults and determination of dominant factors.

[0095] Furthermore, the fault type list is sorted from high to low according to the matching confidence level. The deterioration channel index identifies the wavelength channel number where the anomaly occurs. The degree of spectral fingerprint anomaly is represented by the normalized value of the deviation between the spectral fingerprint vector and the historical benchmark value. The coupling degree value is the diagonal element of the channel-level spectral-polarization coupling degree matrix. The causal inference conclusion of coupling faults includes three types: spectral change leading, polarization change leading, or synchronous change. The dominant factor determination includes three types: device degradation leading, link polarization effect leading, or environmental factors leading.

[0096] In this embodiment of the application, in order to support fault trend prediction, the spectral-polarization joint diagnostic report also includes historical change curves of spectral fingerprints and polarization dynamic characteristics of each wavelength channel, as well as feature degradation trends predicted based on linear regression models, thereby providing a reference for preventive maintenance.

[0097] Furthermore, the input to the linear regression model is the historical time series of spectral fingerprints and polarization dynamic characteristics of each wavelength channel, and the output of the linear regression model is the predicted sequence of feature values ​​for future time periods.

[0098] Furthermore, the historical change curve is plotted as a line graph with time on the horizontal axis and feature values ​​on the vertical axis, which intuitively shows the evolution trend of each feature; the linear regression model assumes that the feature values ​​change linearly with time, and obtains the slope and intercept parameters by fitting historical data, and then extrapolates to the future time period to obtain the predicted value.

[0099] This implementation uses a narrowband filter window sequence to separate broadband spectral data and extract the spectral features of each wavelength channel independently. This overcomes the problem in DWDM systems where dense wavelength channel spectral overlap makes it impossible to distinguish the independent degradation of a single channel in broadband spectral fingerprints.

[0100] This implementation generates a channel coupling relationship matrix by calculating the correlation coefficient of the spectral fingerprint vectors of adjacent wavelength channels. Based on the distribution characteristics of the coupling coefficients of each wavelength channel in the channel coupling relationship matrix, it distinguishes between single-channel independent faults and multi-channel crosstalk faults. Therefore, it overcomes the problem of misjudging the fault source caused by similar characteristic changes in the spectrum due to nonlinear crosstalk between adjacent channels and device degradation.

[0101] This implementation method synchronously acquires the Stokes parameter timing data of the bidirectional transmission signals at both ends of the optical link, extracts the polarization dynamic feature vectors of the forward and reverse transmissions respectively, calculates the difference, and generates a polarization symmetry difference vector to characterize the degree of asymmetry of polarization correlation effect. Therefore, it overcomes the problem that traditional power or signal-to-noise ratio difference methods cannot detect bidirectional asymmetry of polarization state correlation.

[0102] This implementation method identifies strongly coupled and weakly coupled spectral-polarization channels by calculating the cross-correlation coefficient between the temporal changes of single-wavelength spectral fingerprints and the temporal changes of polarization dynamic characteristics. It then performs causal inference on the time-leading-lag relationship between spectral changes and polarization changes in strongly coupled spectral-polarization channels. This overcomes the problems of rapid polarization state drift interfering with spectral measurements and spectral distortion affecting polarization state estimation, leading to misjudgments in single-dimensional analysis. It achieves independent root cause diagnosis of coupling faults between spectral degradation and polarization degradation.

[0103] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A testing method for optical interconnect links for computing power clusters, characterized in that, Includes the following steps: Acquire wavelength configuration data and broadband spectral data of the DWDM system, and simultaneously acquire Stokes parameter timing data of bidirectional transmission signals from coherent transceiver modules at both ends of the optical link to generate a bidirectional polarization state evolution trajectory dataset. Based on the center wavelength and spacing parameters of each wavelength channel, the broadband spectral data is separated by a narrowband filter window sequence to generate independent spectral curve sets for each wavelength channel. The center wavelength shift, power value, and sideband steepness are extracted from the independent spectral curves of each wavelength channel to generate a single-wavelength spectral fingerprint vector set; the correlation coefficient of the spectral fingerprint vectors of adjacent wavelength channels is calculated to generate the coupling relationship matrix between channels; For the evolution trajectory of the bidirectional polarization state, the drift rate, drift range, and drift periodicity characteristics of the polarization state on the Poincaré sphere are calculated respectively to generate a bidirectional polarization dynamic feature vector; the difference between the bidirectional polarization dynamic feature vectors is calculated to generate a polarization symmetry difference vector. Calculate the cross-correlation coefficient between the temporal variation of the single-wavelength spectral fingerprint of each wavelength channel and the temporal variation of the polarization dynamic characteristics. Based on the cross-correlation coefficient threshold, identify the spectral-polarization strongly coupled channels and weakly coupled channels, and generate a channel-level spectral-polarization coupling degree matrix. By connecting the inter-channel coupling matrix with the polarization symmetry difference vector, a spectral-polarization multi-dimensional fault feature vector is constructed. This vector is then matched with a pre-defined composite fault mode library to generate fault classification results and output a spectral-polarization joint diagnostic report.

2. The optical interconnect link testing method for computing power clusters according to claim 1, characterized in that, The narrowband filtering window sequence refers to a set of bandpass filtering functions centered on the center wavelength of each wavelength channel and with the channel interval as the window width. The bandpass filtering function adopts a super-Gaussian window, and the window width parameter of the super-Gaussian window ranges from 0.3 to 0.5 times the channel interval, and the super-Gaussian order ranges from 2 to 6. Before applying a narrowband filtering window, the broadband spectral data is deconvolved and preprocessed. Then, Wiener filtering or the Richardson-Lucy iterative algorithm is used to perform deconvolution on the broadband spectral data.

3. The optical interconnect link testing method for computing power clusters according to claim 1, characterized in that, The sideband steepness is calculated as follows: determine the wavelength positions where the power drops to 50% of the peak power on both sides of the spectral peak position, calculate the slope on the left and the slope on the right respectively, and take the average of the absolute values ​​of the slopes on both sides as the sideband steepness; the elements of the inter-channel coupling relationship matrix are the Pearson correlation coefficients of the spectral fingerprint vectors of the two corresponding wavelength channels.

4. The optical interconnect link testing method for computing power clusters according to claim 1, characterized in that, The drift rate is calculated as follows: the arc length displacement of the polarization state point on the Poincaré sphere at adjacent time points is calculated, the arc length displacement at each time point is removed to obtain the instantaneous drift rate by time interval, and the average value of all instantaneous drift rates is calculated; the drift range is calculated as follows: the centroid coordinates of the polarization state trajectory point sequence are calculated, the spherical angular distance between each trajectory point and the centroid is calculated, and the maximum value of all spherical angular distances is taken; the drift periodicity characteristic is calculated as follows: autocorrelation analysis is performed on the polarization state time series data, and the time delay corresponding to the first local maximum value in the autocorrelation function is searched as the principal periodic component.

5. The optical interconnect link testing method for computing power clusters according to claim 1, characterized in that, When calculating the polarization dynamic characteristics, the distribution ellipticity and principal axis orientation angle of the polarization trajectory on the Poincaré sphere are also extracted. The distribution ellipticity is characterized by the ratio of the maximum eigenvalue to the minimum eigenvalue obtained by principal component analysis of the coordinates of the polarization trajectory points. The principal axis orientation angle is characterized by the azimuth angle of the first principal component vector obtained by principal component analysis in the Poincaré spherical coordinate system. The distribution ellipticity is related to the PDL, and the principal axis orientation angle is related to the principal polarization state of the PMD.

6. The optical interconnect link testing method for computing power clusters according to claim 1, characterized in that, The cross-correlation coefficient is calculated as follows: Pearson correlation coefficients are calculated between each component of the spectral fingerprint vector time series and each component of the polarization dynamic feature vector time series to obtain the sub-correlation coefficient matrix. The maximum absolute value of all elements in the sub-correlation coefficient matrix is ​​taken as the cross-correlation coefficient. When the cross-correlation coefficient is greater than a preset threshold, it is determined to be a strongly coupled spectral-polarization channel; when the cross-correlation coefficient is less than or equal to the preset threshold, it is determined to be a weakly coupled channel. The preset threshold ranges from 0.6 to 0.

8.

7. The optical interconnect link testing method for computing power clusters according to claim 1, characterized in that, It also includes the step of analyzing the temporal causal relationship of coupling failure: for the spectral-polarization strongly coupled channel, principal component analysis is used to extract the spectral comprehensive characteristic time series and the polarization comprehensive characteristic time series respectively, calculate the cross-correlation function of the two comprehensive characteristic time series, and determine the time delay corresponding to the peak of the cross-correlation function; if the time delay is greater than zero, it is determined that the device degradation is dominant; if the time delay is less than zero, it is determined that the link polarization effect is dominant. If the absolute value of the time delay is less than the synchronization determination threshold, it is determined that the environmental factors are dominant. The value range of the synchronization determination threshold is 2 to 5 sampling points.

8. The optical interconnect link testing method for computing power clusters according to claim 7, characterized in that, When analyzing the lead-lag relationship in time series, the Granger causality test method is used. The Granger causality test method constructs a vector autoregressive model of spectral features and polarization features, respectively constructing an autoregressive model that only contains historical information of target features and a vector autoregressive model that contains historical information of both types of features. The prediction residuals of the two models are compared, and the F test is used to determine whether the improvement in prediction accuracy is statistically significant.

9. The optical interconnect link testing method for computing power clusters according to claim 1, characterized in that, The composite fault mode library contains feature templates for the following fault modes: Single-channel device faults are characterized by abnormal spectral fingerprints of a single wavelength channel and coupling coefficients with adjacent channels below 0.3; Crosstalk faults are characterized by synchronous changes in the spectral fingerprints of at least two adjacent wavelength channels and coupling coefficients above 0.7; PMD-dominated faults are characterized by drift rate differences and periodicity differences in the polarization symmetry difference vector exceeding a preset multiple standard deviation of the normal fluctuation range; PDL-dominated faults are characterized by drift range differences in the polarization symmetry difference vector exceeding a preset multiple standard deviation of the normal fluctuation range. When matching a composite fault mode library, a multi-label classification neural network is used. The output layer of the multi-label classification neural network uses the Sigmoid activation function to output the matching confidence of each fault mode.

10. A testing system for optical interconnect links for computing power clusters, used to execute the testing method for optical interconnect links for computing power clusters as described in any one of claims 1-9, characterized in that, include: The multi-source data acquisition module is used to acquire wavelength configuration data and broadband spectral data of the DWDM system, and simultaneously acquire Stokes parameter timing data of bidirectional transmission signals from the coherent transceiver modules at both ends of the optical link to generate a bidirectional polarization state evolution trajectory dataset. The spectral separation module is used to separate broadband spectral data by applying a narrowband filter window sequence based on the center wavelength and spacing parameters of each wavelength channel, generating independent spectral curve sets for each wavelength channel; The spectral feature extraction module is used to extract the center wavelength shift, power value and sideband steepness from the independent spectral curves of each wavelength channel, generate a single-wavelength spectral fingerprint vector group, and calculate the correlation coefficient of the spectral fingerprint vectors of adjacent wavelength channels to generate the coupling relationship matrix between channels. The polarization feature extraction module is used to calculate the drift rate, drift range and drift periodicity of the polarization state on the Poincaré sphere for the evolution trajectory of the bidirectional polarization state, generate bidirectional polarization dynamic feature vectors, and calculate the difference between the bidirectional polarization dynamic feature vectors to generate polarization symmetry difference vectors. The coupling degree calculation module is used to calculate the cross-correlation coefficient between the temporal changes of the single-wavelength spectral fingerprint and the temporal changes of the polarization dynamic characteristics of each wavelength channel. Based on the cross-correlation coefficient threshold, it identifies spectral-polarization strongly coupled channels and weakly coupled channels, and generates a channel-level spectral-polarization coupling degree matrix. The fault diagnosis module is used to connect the inter-channel coupling relationship matrix with the polarization symmetry difference vector to construct a spectral-polarization multi-dimensional fault feature vector, match it with a preset composite fault mode library, generate fault classification results, and output a spectral-polarization joint diagnosis report.

Images