Photoelectric collaborative testing device of co-packaging module

Abstract

The invention relates to the technical field of photoelectric integration and testing, in particular to a photoelectric collaborative testing device for a co-packaging module, which effectively excites rare faults such as spectrum ghost and the like through synchronous excitation and acquisition reproduction of a dynamic coupling scene; a cross-domain coupling risk feature matrix is constructed through multivariable system identification, and accurate quantification and positioning of a resonance mode are realized; and through adaptive dynamic remodeling and iterative verification based on eigenvectors, a'test-diagnosis-regulation 'closed loop is formed, risks are actively eliminated in the test, and the product reliability is improved.

Description

Technical Field

[0001] This invention relates to the field of optoelectronic integration and testing technology, and more specifically, to an optoelectronic collaborative testing device for a co-packaged module. Background Technology

[0002] Co-packaged optics (CPO) technology significantly improves system bandwidth and energy efficiency by integrating silicon photonics engines with computing / switching chips at high density. However, this integration leads to tight closed-loop coupling of multiple physical fields such as electricity, optics, and heat at the microscale, introducing complex dynamic interaction problems not seen in traditional pluggable modules. Existing testing techniques mainly verify the static parameters of devices or the performance of a single channel under steady-state conditions, making it difficult to effectively reproduce and assess the systemic risks of modules in real-world dynamic operating scenarios.

[0003] A typical deep challenge stems from the parallel multi-wavelength channel structure within the CPO. Each channel typically employs an independent closed-loop thermal tuning mechanism (such as a microheater) to maintain wavelength stability. Under dynamic operating conditions, these parallel control loops can unexpectedly couple through shared substrates or thermal fields. For example, when a channel's thermal tuning control loop exhibits a small oscillation at a specific frequency (such as the limiting loop oscillation of a PID loop), and this frequency component happens to fall within the effective bandwidth of the photodetector of an adjacent channel, a rare "coherent crosstalk" can occur. This crosstalk is not broadband noise, but rather manifests as a frequency-selective resonant interference that can convert the control jitter of one channel into amplitude / phase modulation noise in adjacent channels. This effect is amplified dramatically near a specific temperature or operating point, causing a "cliff-like" degradation in the bit error rate of the affected channel, a phenomenon figuratively known as "spectral strangulation." Due to its frequency-locked and conditionally triggered characteristics, it is extremely difficult to detect and locate in conventional steady-state or single-factor frequency sweep tests, constituting a potentially high-risk failure mode for the CPO module.

[0004] Therefore, the existing technology system lacks a dedicated testing method and device capable of actively stimulating, synchronously observing, and accurately diagnosing such multi-domain dynamic coupling effects, especially frequency-selective resonance interference, under near-real-world operating dynamics. This has become a key technical bottleneck hindering the full reliability verification and performance marginal evaluation of CPO products. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide an optoelectronic collaborative testing device for co-packaged modules.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A photoelectric collaborative testing device for a co-packaged module includes: The synchronous excitation and acquisition module synchronously applies electrical domain test signals and optical domain test signals to the co-packaged module, and synchronously acquires its electrical domain response signals and optical domain response signals. The multi-domain signal analysis module analyzes the electrical domain response signal and the optical domain response signal, and separates and extracts the closed-loop thermal tuning control signal of each optical communication channel. The coupled system identification module performs multivariable system identification and intrinsic mode decomposition on the integrated system composed of optical communication channels based on the extracted closed-loop thermal tuning control signal, thereby extracting the cross-domain coupling risk feature matrix. The resonance risk assessment module analyzes the eigenvalues ​​and eigenvector distributions of the cross-domain coupling risk feature matrix to determine whether there are resonant coupling risk modes between optical communication channels. The control strategy dynamic reshaping module, when it is determined that there is a resonant coupling risk mode, performs adaptive dynamic reshaping of the closed-loop thermal tuning control strategy of the optical communication channel with the resonant coupling risk mode based on the eigenvector corresponding to the resonant coupling risk mode. The iterative convergence verification module repeats the synchronization excitation and acquisition steps up to the resonance risk determination step after the closed-loop thermal tuning control strategy is dynamically reshaped, until all risk modes in the cross-domain coupling risk feature matrix are suppressed.

[0007] Furthermore, the synchronous excitation and acquisition module includes a programmable arbitrary waveform generator and a tunable laser source. The programmable arbitrary waveform generator is configured to generate an electrical domain test signal containing a high-frequency power supply noise spectrum and a high-speed data code pattern, and the tunable laser source is configured to generate an optical domain test signal whose wavelength can be rapidly modulated. The synchronous excitation and acquisition module also includes a high-speed digital storage oscilloscope and an optical waveform analyzer, which are used to synchronously trigger and acquire the electrical domain response signal and the optical domain response signal, respectively.

[0008] Furthermore, the multi-domain signal analysis module includes a high-speed data acquisition unit and a digital signal processor. The digital signal processor is configured to perform joint time-frequency analysis on the acquired electrical domain response signal and optical domain response signal, and to separate and demodulate the micro heater drive current signal or thermoelectric cooler control voltage signal corresponding to each optical communication channel from the mixed response signal through blind source separation algorithm or digital filtering, as a closed-loop thermal tuning control signal.

[0009] Furthermore, the coupled system identification module is configured to perform the following numerical calculation process: abstract the integrated system composed of optical communication channels into a multivariable dynamic system model, wherein the closed-loop thermal tuning control signal of each channel is regarded as the system input, and the response signal after photoelectric conversion of each channel is regarded as the system output; An adaptive recursive parameter estimation algorithm is used to update the coupling parameter matrix in the multivariable dynamic system model online in real time. Modal perturbation analysis is performed on the updated coupling parameter matrix to solve for all its right eigenvectors and corresponding complex frequency eigenvalues. The cross-domain coupling risk feature matrix is ​​obtained by calculating the tensor shrinking product of the eigenvector matrix, the diagonal matrix of complex frequency eigenvalues, and a weight matrix determined by the physical spacing of each channel.

[0010] Furthermore, the steps for solving for all its right eigenvectors and their corresponding complex frequency eigenvalues ​​specifically include: An iterative numerical method of matrix spectral decomposition is used to transform the coupling parameter matrix into a Shanghai Senberg matrix; The Hessenberg matrix is ​​solved by diagonalization using an implicit QR algorithm with displacement, where the displacement is dynamically determined by calculating the eigenvalues ​​of the last second-order principal submatrix of the current iteration matrix. The algorithm iterates until the modulus of all off-diagonal elements is less than a preset small tolerance threshold. The iteration matrix converges into a quasi-upper triangular matrix, and the second-order blocks or first-order elements on its diagonal give the complex frequency eigenvalues. For each obtained complex frequency eigenvalue, the corresponding linear equation system is solved by back substitution to obtain a non-zero solution vector. After normalization, this solution vector constitutes the corresponding right eigenvector.

[0011] Furthermore, an iterative numerical method based on matrix spectral decomposition is used to transform the coupling parameter matrix into a Shanghai Senberg matrix, as follows: A numerical algorithm based on Haushold reflection transformation is adopted, which operates on each column of the matrix in turn: For the k-th column, k starts from 1, the numerical algorithm constructs the Haushold transformation matrix, which is designed to make all elements from the (k+2)-th row to the last row in the current column zero, while maintaining their orthogonality; Multiply this Haushold transformation matrix both on the left and on the right into the matrix to be transformed to complete an orthogonal similarity transformation. The orthogonal similarity transformation eliminates all elements from the (k+2)th row to the last row in the current column while keeping all eigenvalues ​​of the matrix unchanged. Repeat this process, processing columns one through three from the first to the third to last. After a series of such orthogonal similarity transformations, the original coupling parameter matrix is ​​transformed into a Shanghai Senberg matrix, which is a matrix in which all lower triangular elements except for the main diagonal and the first diagonal are zero.

[0012] Furthermore, the specific steps for constructing the Haushold transformation matrix include: Select a sub-vector from the (k+1)th element to the last element in the k-th column of the matrix to be transformed; Calculate the Euclidean norm of the subvector and construct a reflection vector based on the norm value and the first element of the subvector. Using this reflection vector, the corresponding projection matrix is ​​generated through outer product operation, and then the final Haushold transformation matrix is ​​calculated.

[0013] Furthermore, the specific steps for analyzing the eigenvalues ​​and eigenvector distributions of the cross-domain coupling risk feature matrix to determine the resonant coupling risk mode include: Calculate the real part of the eigenvalues, and set the real part greater than a preset positive real part threshold. The oscillation mode corresponding to the eigenvalues ​​is labeled as an unstable mode; where ; Calculate the component amplitudes of the eigenvectors corresponding to unstable modes, and identify those whose amplitudes exceed a preset channel participation threshold. The optical communication channel corresponding to the component is marked as a high participation channel; Calculate the phase difference between the eigenvector components corresponding to any two high-participation channels. If the absolute value of the phase difference is less than a preset phase synchronization threshold... If so, it is determined that there is a resonant coupling risk mode among these high-participation channels, and its risk level is jointly quantified by the real part of the corresponding eigenvalue and the participation level.

[0014] Furthermore, the specific steps for the control strategy dynamic reshaping module to perform adaptive dynamic reshaping include: Based on the amplitude and phase of each channel component in the eigenvector corresponding to the resonant coupling risk mode, the target adjustment amount of the thermal tuning control strategy for each high-risk channel is calculated. This adjustment amount is proportional to the amplitude of the component in the eigenvector of that channel and opposite in phase. The high-risk channel is determined by whether the amplitude of the corresponding component of the high-participation channel in the eigenvector exceeds the channel participation threshold. And whether the absolute value of its phase difference with the dominant channel is less than the phase synchronization threshold. Determined comprehensively; The integral time constant of the closed-loop thermal tuning control strategy for high-risk channels is adjusted by weighting, and the adjustment weight is determined by the real part of the corresponding eigenvalue. An adaptive notch filter is introduced into the closed-loop thermal tuning control strategy. The center frequency of the adaptive notch filter is calculated based on the imaginary component of the eigenvalue that generates the resonant coupling risk mode, and its suppression depth is adaptively adjusted according to the amplitude of the corresponding channel component in the eigenvector.

[0015] Compared with the prior art, the present invention has the following beneficial effects: This invention, by simultaneously applying electrical and optical test signals containing high-frequency noise and dynamic modulation components, and synchronously acquiring the response of the co-packaged module with high precision, effectively reproduces for the first time in a test environment the dynamic coupling scenario of the internal electrical, optical, and thermal multi-physics fields under real working conditions. This capability enables the active excitation and reliable capture of rare and severe coupling faults, such as spectral strangulation, caused by interference of multiple closed-loop control loops and manifested only under specific dynamic conditions. It overcomes the fundamental limitation of traditional steady-state or single-domain test methods in simulating complex dynamic interactions, providing a crucial test foundation for evaluating the reliability of co-packaged modules under extreme edge conditions. This invention innovatively introduces a multivariable system identification and intrinsic mode decomposition algorithm. It uses the closed-loop thermally tuned control signals of each channel, analyzed from the mixed response, as input to construct and solve a cross-domain coupling risk characteristic matrix characterizing the overall dynamic coupling relationship of the system. By analyzing the eigenvalues ​​and eigenvector distribution of this matrix, the resonant coupling risk modes present can be accurately quantified and determined, including their oscillation frequency, spatial distribution (specific channels involved), and instability growth rate. This enables mathematical modeling and accurate diagnosis of frequency selectivity and conditionally triggered interference, which are difficult to locate using traditional frequency domain testing. It delves deeper into the physical mechanism of resonant coupling in specific modes, moving from the superficial level of channel performance degradation to the underlying physical mechanism of resonant coupling. This invention goes beyond traditional testing, which merely identifies problems. By using the eigenvectors corresponding to resonant coupled risk modes, it adaptively and dynamically reshapes the parameters and structure of the closed-loop thermal tuning control strategy for relevant channels, and iteratively verifies the results through testing, forming a complete closed loop of perception-diagnosis-regulation-verification. This process proactively alters the dynamic characteristics of the system, disrupting resonance conditions and thus eliminating identified risk modes during testing. This is not only an innovation in testing methods but also a concept for building a dynamic immune system for co-packaged modules, significantly improving the final reliability level and the depth of robust design verification of the product. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the operating principle of an optoelectronic collaborative testing device with a co-packaged module. Figure 2 A schematic diagram illustrating the operating principle of the coupled system identification module; Figure 3 The flowchart shows the adaptive dynamic reshaping process performed by the control strategy dynamic reshaping module. Detailed Implementation

[0017] Reference Figures 1 to 3A photoelectric collaborative testing device for a co-packaged module (a co-packaged module is a highly integrated structure achieved through advanced packaging technology, the core feature of which is the heterogeneous integration of silicon photonic chips (or photonic engines) with application-specific integrated circuits at the physical level. This integration typically employs 2.5D / 3D packaging, silicon interposers, microbumps, and other processes, enabling the electrical and optical chips to be directly interconnected at the millimeter or even micrometer scale through internal packaging circuitry, forming a tightly coupled, indivisible single package), including the following: Synchronous Excitation and Acquisition Module: This module establishes the physical foundation for simulating the real, dynamic operating environment of the co-packaged module. Traditional separate testing cannot reproduce the tightly coupled temporal interaction of electro-optical-thermal multi-physics fields. This module ensures strict temporal alignment of excitation and response data by simultaneously applying electrical and optical test signals containing specific interference modes (such as high-frequency noise and rapid load transitions) and simultaneously acquiring the system's response. This provides a unique and correct original data source for subsequent analysis of cross-domain dynamic coupling, which can be used for causal correlation and is a prerequisite for capturing transient and coherent faults such as spectrum strangulation.

[0018] Multi-domain signal analysis module: This module acts as a decoder to extract key state variables from the original mixed signal. Within the co-packaged module, core information reflecting the operating state of each optical channel (such as wavelength stability) is encoded in the closed-loop thermal tuning control signal (such as microheater current), but this signal is physically mixed with high-speed data signals and power supply noise. This module uses advanced signal processing techniques (such as joint time-frequency analysis and blind source separation) to separate and extract the independent closed-loop thermal tuning control signal for each optical channel from the synchronously acquired mixed response signal in the electrical and optical domains. This step is crucial for treating the system as an observable and analyzable multivariable control system.

[0019] The Coupled System Identification Module elevates the co-encapsulated module from a set of parallel channels into a system with internal coupling dynamics for modeling and analysis. Based on the extracted closed-loop thermal tuning control signals (system input) and their photoelectric responses (system output) of each channel, this module, through multivariate system identification technology, quantitatively constructs a mathematical model of the dynamic coupling relationship between channels for the first time. Further, through eigenmode decomposition, this model is decoupled into a series of independent oscillation modes. The final output cross-domain coupling risk feature matrix mathematically encapsulates all potential resonant coupling modes of the system, their oscillation frequencies (imaginary parts of eigenvalues), growth rates (real parts of eigenvalues), and the degree of participation of each channel in this mode (eigenvectors). This represents a leap from phenomenological observation to mechanistic quantification.

[0020] The resonance risk assessment module provides decision rules for the automatic and intelligent diagnosis of systemic risks. It does not simply examine the performance indicators of individual channels, but rather analyzes the eigenvalues ​​and eigenvector distributions of the cross-domain coupling risk characteristic matrix. This is achieved by setting scientific thresholds (such as the positive real part threshold). Determining instability: Channel participation threshold Determine the scope of influence and the phase synchronization threshold. (For coherence assessment), this module can automatically identify, from numerous mathematical modes, which are actual resonant coupling risk modes that could lead to a precipitous drop in channel performance. This enables precise localization of latent faults with specific frequencies and spatial patterns, such as spectrum strangulation, caused by abnormal interactions of normal components.

[0021] The dynamic reshaping module for control strategies endows the test system (or the future product itself) with the ability to self-correct online, realizing a paradigm shift from passively discovering problems through testing to actively adjusting and eliminating them. Once resonance risk is determined, this module precisely performs adaptive dynamic reshaping of the closed-loop thermal tuning control strategy of the optical communication channel at risk, based on the eigenvectors corresponding to the resonant coupling risk mode (i.e., the spatial distribution map of the risk mode). Reshaping methods include targeted adjustments to control parameters and the introduction of adaptive notch filters. Its core purpose is to modify the dynamic characteristics of the relevant channel control loop, disrupting the phase, frequency, or gain conditions that lead to resonance, thereby suppressing the risk mode at its root, rather than merely masking the symptoms.

[0022] The iterative convergence verification module ensures the closed-loop nature, effectiveness, and convergence of the entire testing and control process. After the control strategy is reshaped, the system state has changed. This module repeats the entire process from synchronization stimulus to risk assessment, which is an active verification loop. Its goal is to confirm that after reshaping, all risk modes in the cross-domain coupled risk feature matrix have been suppressed (e.g., the real parts of the corresponding eigenvalues ​​turn negative). This iterative process guarantees the effectiveness of the control measures, and the test is only completed when all identified risks have been eliminated, thus ensuring the robustness of the final state. This embodies the complete systems engineering concept of "test-diagnosis-control-verification".

[0023] In one specific embodiment, a programmable arbitrary waveform generator and a tunable laser source are prepared and configured. The output of the programmable arbitrary waveform generator is connected to the power supply and high-speed electrical interface of the target co-package module via a high-frequency probe and an RF cable. The output light of the tunable laser source is precisely coupled to the optical input port of the target co-package module via a polarization controller and an optical fiber array. The programmable arbitrary waveform generator is programmed to generate a composite electrical domain test signal. This signal includes a high-frequency power supply noise spectrum covering 100 kHz to 1 GHz caused by the high-speed switching action of the analog chip, and is superimposed with a pseudo-random binary sequence code pattern with a rate of 112 Gbps used to simulate real data transmission. The tunable laser source is set to perform fast wavelength modulation in the form of a sine wave or a triangular wave at a rate higher than 10 kHz near the target wavelength. The aforementioned signal generating equipment is activated simultaneously, and the electrical domain test signal and the rapidly modulated optical domain test signal are applied to the target co-packaged module at the same time, such as an integrated device containing a 16-channel parallel silicon photonics engine and a switching chip, to simulate the load transients and temperature fluctuations it encounters in actual data center operation. A high-speed digital storage oscilloscope and an optical waveform analyzer, both triggered by an external clock, were used. The probe of the high-speed digital storage oscilloscope was connected to the monitoring pin of the co-packaged module, while the input of the optical waveform analyzer was connected to the optical output fiber of the co-packaged module. Under the same trigger signal, both instruments simultaneously captured and recorded the electrical and optical response signals output by the co-packaged module, including dynamic responses, at a rate exceeding 50 gigabits per second, ensuring strict time-domain alignment of the two signal data. Finally, the acquired electrical and optical response signals were transmitted to a powerful computing workstation. The workstation ran dedicated software to perform time-scale alignment verification and initial preprocessing of the data, verifying signal integrity and providing an accurate and synchronized raw data foundation for subsequent in-depth signal analysis and coupling feature extraction. This completed the first closed-loop excitation and response acquisition of the co-packaged module under dynamic and real-world operating conditions. Specifically, the time-scale alignment check uses the rising edge of the same external trigger signal received by the high-speed digital storage oscilloscope and the optical waveform analyzer as the absolute time zero reference, and applies precise time delay compensation to the electrical domain response signal data stream and the optical domain response signal data stream recorded by the two respectively; this compensation amount is calculated by the difference between the instrument's inherent transmission delay and the physical length of the link, ensuring that the characteristic transient points generated by the same excitation event in the two signals completely coincide on the time axis; The initial preprocessing performs gain normalization and baseline correction on the two aligned raw data streams. Specifically, based on the known injection signal amplitude and acquisition device settings, the voltage and optical power values ​​are converted to a unified per-unit system, and the DC bias is removed. A linear-phase digital filter with a passband matching the excitation signal spectrum is applied to suppress out-of-band noise while maintaining the waveform's phase information without distortion. Signal integrity verification is accomplished by calculating and evaluating three key quantitative indicators: first, calculating the signal-to-noise ratio (SNR) of both signals during the excitation duration, ensuring both are above a preset minimum threshold (e.g., 35 dB); second, checking whether the dynamic range of the signal waveform covers the expected response amplitude range, without clipping or saturation distortion; and third, analyzing the timing jitter at specific characteristic points between the two signals using a cross-correlation algorithm, confirming that its root mean square value is less than the system's allowed clock tolerance (e.g., 1 picosecond). Only when all the above verification and preprocessing steps are successfully passed is the synchronously acquired data marked as valid and allowed to flow into the subsequent signal analysis process. In one specific implementation, the time-aligned and preprocessed effective electrical domain response signal and optical domain response signal are loaded onto a computing platform equipped with a multi-core digital signal processor. The platform first performs joint time-frequency analysis on these two time-synchronized signals, specifically employing a synchronous compressed short-time Fourier transform algorithm to generate a time-spectrum diagram that accurately characterizes the distribution of signal energy in a two-dimensional time and frequency plane, thereby revealing, for example, periodic fluctuations with a frequency of approximately 85 kHz that occur at specific times, corresponding to the 1560 nm wavelength channel. Based on the prior knowledge of the signal components revealed by the time-frequency spectrum, a hybrid model containing multiple potential source signals is constructed. An independent component analysis algorithm based on maximum likelihood estimation is used as a blind source separation method. This algorithm iteratively optimizes to find an unmixing matrix that maximizes the statistical independence between the output signals, thereby decomposing the original hybrid response signal into several statistically independent source signal components. By analyzing the time-frequency characteristics, amplitude range, and correlation with known excitation modes of each separated source signal component, signal components whose characteristics match the electrical drive characteristics of silicon photonic chip microheaters or thermoelectric coolers are identified and screened. For example, a current signal with an amplitude fluctuating within the 10 mA range and whose dominant frequency component matches the bandwidth of the thermally tuned control loop. For the identified candidate signals, digital lock-in amplification technology is further applied, using the nominal operating wavelength or control clock of each optical communication channel as a reference frequency for precise demodulation. This process ultimately extracts independent microheater drive current signals or thermoelectric cooler control voltage signals carrying real-time thermal tuning status information for each channel. These signals serve as the closed-loop thermal tuning control signals required for subsequent coupling analysis. For example, from the mixed response of a four-wavelength channel co-packaged module, four independent thermoelectric cooler control voltage waveforms with different frequencies, corresponding to the 1545nm, 1550nm, 1555nm, and 1560nm channels respectively, may be successfully separated. In one specific implementation, the extracted closed-loop thermal tuning control signals of multiple optical communication channels, such as the time series of control voltages of the thermoelectric coolers for each of the four wavelength channels, are used. to , defined as the input vector U(t) of the multivariable dynamic system; simultaneously, the time series of the optical power response signal of the corresponding channel after conversion and preprocessing by the photodetector. to Let Y(t) be the output vector. Based on this, a discrete-time state-space model is established as a multivariable dynamic system model, with the following form: Where k is the sampling time, X is the internal state vector to be estimated, A, B, and C are the system matrix and input / output matrices to be identified, and W and V represent process noise and measurement noise. In this embodiment, the coupling parameter matrix to be identified is the system matrix A, which encapsulates the dynamic characteristics of each channel and their coupling relationships. A recursive least squares algorithm with a forgetting factor is employed to update the estimate of the coupling parameter matrix A online in real time. This recursive process is performed at each new sampling time k: based on the parameter estimates from the previous time step... and state estimation covariance matrix Combining the measured input U(k) and output Y(k) at the current moment, the gain vector K(k) is calculated. The gain vector K(k) is given by the formula... Given, among which The regression vector is composed of historical input and output data. A forgetting factor between 0.95 and 0.995 is used to assign higher weights to new data to adapt to time-varying characteristics. Update parameter estimates. And simultaneously update the covariance matrix. This prepares for the recursion at the next time step. Through continuous recursion, the estimated value of the coupling parameter matrix A can track and converge to the current true dynamic characteristics of the system; The coupling parameter matrix tends to stabilize after sufficient recursive updates Modal perturbation analysis is performed to solve for all right eigenvectors and their corresponding complex frequency eigenvalues. This process is accomplished through numerical computation: first, a series of orthogonal similarity transformations (such as the Householder transformation) are used to transform the matrix... The matrix is ​​transformed into a Shanghai Senberger matrix H to simplify calculations. An implicit QR algorithm with displacement is applied to H for diagonalization. In each iteration, the displacement μ is determined by calculating the eigenvalues ​​of the last second-order principal submatrix of the current iteration matrix to accelerate algorithm convergence. Iteration continues until the magnitudes of all off-diagonal elements are less than a preset small tolerance (e.g., 1e-10). The resulting quasi-upper triangular matrix T has all its complex frequency eigenvalues ​​given by its diagonal elements or second-order blocks. The real part The imaginary part reflects the decay or growth rate of a mode. The corresponding oscillation frequency. For each complex frequency eigenvalue... Solving the system of linear equations by back substitution To obtain the non-zero solution vector After normalization, it becomes the corresponding right eigenvector, which describes the amplitude and phase distribution of the oscillation mode in each channel; Based on the above solution results, a cross-domain coupling risk feature matrix is ​​constructed. All right eigenvectors are arranged column-wise to form an eigenvector matrix V, and the corresponding complex frequency eigenvalues ​​are used to construct a diagonal matrix S. Simultaneously, based on the physical layout coordinates of each optical communication channel on the chip, the pairwise Euclidean distances between them are calculated. and using the formula Construct a weight matrix W, where L is the feature length scaling constant, and the weights are... This reflects the possibility that the physical proximity between channels may lead to enhanced thermal or electrical coupling; The cross-domain coupling risk feature matrix R is obtained by calculating the tensor shrinkage product of these three matrices, and the specific operation is as follows: ,in Indicates the Kronecker product. This represents the Hadamard product (element-by-element multiplication). Let V be the conjugate transpose of V, and let ones be a matrix of all ones with the same dimension as S. The real operation is performed on the real part. The resulting matrix R is a symmetric matrix, and its (i,j)th element... The dynamic coupling risk intensity between the i-th oscillation mode and the j-th oscillation mode was quantified after considering spatial proximity weighting. The larger the value, the higher the possibility of resonance risks such as spectral strangulation caused by the cross-mode coupling, thus providing a direct and quantitative basis for subsequent risk assessment. In one specific implementation, the coupling parameter matrix to be processed is loaded into the computing environment and its dimensions are determined. For example, for a fourth-order real matrix identified from a four-channel system, its dimensions are 4 rows and 4 columns. The transformation process starts from the first column of the matrix, starting with the column indexed as 1. For this column, a sub-vector from the second row to the last row is selected, and its Euclidean norm is calculated based on this sub-vector. A specific reflection vector is then constructed based on this. The first element of the reflection vector is formed by combining the first element of the sub-vector with the norm and a specific sign rule, and the remaining elements are directly taken from the corresponding elements of the sub-vector. Using the constructed reflection vector, the corresponding Householder transformation matrix is ​​generated through mathematical operations. Specifically, the outer product of the reflection vector and its transpose is first calculated, then the outer product matrix is ​​multiplied by a scaling factor determined by the inner product of the reflection vectors, and finally the scaled outer product matrix is ​​subtracted from the identity matrix to obtain a symmetric and orthogonal square matrix, which is the Householder transformation matrix applied to the current column. This Householder transformation matrix is ​​then multiplied both left and right onto the current coupling parameter matrix to be transformed. The left multiplication operation operates on the rows of the matrix, and the right multiplication operation operates on the columns of the matrix. This pair of transformations constitutes a complete orthogonal similarity transformation. The core effect of this transformation is that it precisely reduces all elements in the current column, starting from the third row downwards, to zero. At the same time, due to the orthogonal similarity of the transformation, all eigenvalues ​​of the original matrix remain completely unchanged. Repeat the above process, incrementing the column indices sequentially, performing the exact same construction and transformation operations on the second to third-to-last columns of the matrix. In each step, the transformation is applied to the intermediate matrix obtained in the current step. After processing all specified columns, the original full matrix is ​​transformed into a Hessenberg matrix. Taking a 4x4 matrix as an example, the final result is a matrix where non-zero values ​​may exist only on the main diagonal, the previous diagonal, and the second diagonal immediately below the previous diagonal, while all elements in the lower triangular region are zero.

[0024] In one specific implementation, the matrix to be orthogonally similarized and the column index k being processed are determined. For example, for a fifth-order intermediate matrix, when processing its second column (i.e., k=2), a specific sub-vector needs to be selected from that column. This sub-vector consists of all elements in that column starting from the (k+1)th element (i.e., the third element) to the last element (the fifth element). Let this sub-vector be denoted as x, and its specific value might be... ; Based on this subvector x, a key reflection vector u is constructed. The calculation process is as follows: First, the Euclidean norm of the subvector x is calculated, denoted as nrm. Next, a sign factor sgn is determined, whose value is the sign of the first element x1 of the subvector x, but it is agreed that if x1 is 0, then sgn is 1. Then, the first element of the reflection vector u is calculated as... The remaining elements of the reflection vector u are directly taken from the second and subsequent elements of the corresponding sub-vector x. Thus, the reflection vector u is constructed, and its function is to specify the direction of the hyperplane normal around which the Householder reflection is performed; Using the constructed reflection vector u, a projection matrix is ​​generated through an outer product operation. Specifically, the reflection vector u and its transpose are first calculated. The outer product yields a matrix of rank one. Calculate the inner product of the reflection vector u, i.e. The key projection matrix, or reflection operator, is determined by the formula... The calculation shows that the matrix represents the projection onto the space spanned by u.

[0025] Calculate the final Householder transformation matrix H. This matrix is ​​obtained by subtracting twice the aforementioned projection matrix from the identity matrix I, i.e. This transformation matrix H possesses the important properties of symmetry and orthogonality, that is, it satisfies... and When applied to matrix transformations, it can accurately reflect the atomic vector x in a direction parallel to the first coordinate axis, thereby achieving the goal of clearing the elements in the target interval of a specific column of the original matrix to zero, while ensuring that the transformation is an orthogonal similarity transformation that preserves eigenvalues.

[0026] In one specific implementation, eigenvalue decomposition is performed on the constructed cross-domain coupling risk feature matrix to obtain all its eigenvalues ​​and corresponding eigenvectors. The real part of each eigenvalue is calculated, and the real part value is compared with a preset positive real part threshold. Compare; the threshold The setting is based on general engineering practices regarding system stability margin in classical control theory, and combined with the stringent requirements for device stability in the high-speed optical interconnect industry. Typically, a positive decimal much less than 1 is chosen, such as 0.05 or 0.1. Its physical meaning is that it allows for an extremely weak positive growth rate of the oscillating mode to cover model errors and noise, but once this threshold is exceeded, it means that the mode has a significant risk of divergence. All real parts greater than... The dynamic behavior characterized by the eigenvalues ​​is labeled as unstable modes, which are potential sources of resonance risk. For example, for a set of six eigenvalues, two of which have real parts of 0.002 and 0.15, let... Then only the modes corresponding to the eigenvalues ​​with a real part of 0.15 are marked as unstable; For each labeled unstable mode, analyze the components of its corresponding eigenvector. Calculate the absolute value of the amplitude of each component (corresponding to one optical communication channel) and compare these amplitudes with a preset channel participation threshold. Compare. Threshold The significance level is set based on the statistical concept, typically taking a value of 1.5 to 2 times the mean amplitude of all components in that mode, or empirically fixed at 0.2 (after normalization). Its purpose is to screen out channels that play a dominant role and make significant energy contributions in that unstable mode. Channels with amplitudes exceeding... The optical communication channel corresponding to the eigenvector of an unstable mode is labeled as the high-participation channel of that mode. For example, the component amplitudes of the eigenvector of an unstable mode on the four channels are as follows: ,set up Then the second and fourth channels are marked as high-participation channels; For each unstable mode, examine the phase relationship between all its high-involvement channels. Calculate the difference in phase angles between the corresponding components of the eigenvectors of any two high-involvement channels and take their absolute values. Then, compare this absolute phase difference with a preset phase synchronization threshold. Compare. Threshold Based on coherence theory and the synchronization acquisition range principle of phase-locked loops, in optical communication, π / 6 (30 degrees) to π / 4 (45 degrees) is typically selected as the boundary for determining whether two periodic signals are in a strongly coherent or phase-locked state. If the absolute value of the phase difference between any pair of highly involved channels is less than... If this is the case, then it is determined that there is a potential risk of phase synchronization among these channels, and thus the unstable mode is formally determined to be a specific resonant coupling risk mode. Continuing the previous example, if the phase difference between the second and fourth channel components is 15 degrees, and... If so, then the condition is met; The risk level of identified resonant coupling risk modes is comprehensively quantified. The risk level is characterized by a composite index, which is proportional to the real part of the corresponding eigenvalue of the mode (reflecting the rate of instability growth) and multiplied by the sum of the amplitudes of all high-participation channels in that mode (reflecting the breadth and intensity of the participating channels). For example, a risk mode with a real part of 0.15 and a high-participation channel amplitude sum of 1.1 will have a higher risk level than another risk mode with a real part of 0.08 and an amplitude sum of 0.9. This quantification result is output to guide the adjustment intensity and priority ranking of subsequent dynamic reshaping strategies, thereby achieving accurate, graded early warning and handling of spectrum strangulation-type faults.

[0027] In one specific implementation, high-risk channels requiring regulation are accurately identified based on the determined resonant coupling risk modes and their corresponding eigenvectors. The identification process incorporates two criteria: first, the amplitude of the corresponding component of the channel in its eigenvector must exceed a preset channel participation threshold. This indicates that it makes a significant energy contribution in this risk mode; secondly, the absolute value of the difference between the phase of this channel component and the phase of the dominant channel with the largest amplitude among all highly involved channels must be less than a preset phase synchronization threshold. This ensures they are in a strongly coherent state. For example, for a risk mode, its eigenvectors show that the amplitudes of channel 2 and channel 4 are 0.8 and 0.3, respectively. ), and their phase differences with the dominant channel 2 are 0 degrees and 15 degrees, respectively. If the channel is identified as high-risk, then both channel 2 and channel 4 are identified as high-risk channels in this restructuring. Based on the eigenvectors of the resonant coupling risk modes, the target adjustment amount for the thermally tuned control strategy (such as the drive current) is calculated for each high-risk channel. This adjustment amount... The calculation follows the formula: ,in and These are the magnitude and phase of the corresponding component in the eigenvector of the channel, respectively. It is the reference phase (usually the dominant channel phase or the overall average phase of the risk mode). It is a global convergence coefficient. The negative sign in the formula and the sine function ensure that the direction of the adjustment is opposite to the direction of the vibration component of the channel in the risk mode, thus playing an active cancellation role. For example, for channel 4, if , , , Then the calculation yields This means that a small, inverse adjustment needs to be applied to its drive current; For identified high-risk channels, the integral control parameters of their closed-loop thermal tuning control strategy are adjusted using weighted adjustments. The real part of the eigenvalues ​​corresponding to the resonant coupling risk modes in which this channel participates. The size is directly related, specifically it can be... Where κ is the proportionality coefficient. This is the aforementioned threshold for the positive real part. The original integration time constant of the channel is... Adjusted to The physical meaning of this adjustment lies in targeting faster growth ( For risk modes that are larger (increase in risk), the integral time constant should be increased accordingly to reduce the gain of the control loop near the corresponding frequency, thereby weakening its support for oscillations. A second-order adaptive notch filter is inserted into the feed path of the closed-loop thermally tuned control strategy for the affected high-risk channel. The center frequency of this filter is... The imaginary part of the eigenvalues ​​that directly cause the risk (Angular frequency) is calculated, that is Its suppression depth parameter Then, based on the amplitude of the corresponding component of that channel in the eigenvector... To perform adaptive adjustment, for example , To set the coefficients. In this way, the filter can selectively and deeply attenuate specific frequency interference signals characterized by resonant risk modes, fundamentally disrupting the conditions for positive feedback at that frequency. For example, for The risk mode, with the filter center frequency set to 85kHz, if This generates a notch with a relatively large depth, thereby effectively suppressing the frequency component.

[0028] In one specific implementation, after the adaptive dynamic reshaping of the closed-loop thermally tuned control strategy for the high-risk channel is completed, a completely new and comprehensive test and diagnostic cycle is immediately initiated. The operator or automatic control program will again apply the same or similar spectral characteristics of the electrical and optical test signals to the co-packaged module via a programmable arbitrary waveform generator and a tunable laser source. Using a high-speed digital storage oscilloscope and an optical waveform analyzer, the electrical and optical response signals output by the co-packaged module under this new control strategy will be simultaneously triggered and acquired, thereby obtaining raw data pairs characterizing the dynamic behavior of the system after reshaping. Secondly, the newly acquired electrical domain response signal and optical domain response signal are sent to the multi-domain signal analysis process for processing. The computing platform performs joint time-frequency analysis and blind source separation on these signals, thereby separating and demodulating the new closed-loop thermal tuning control signal that actually operates under the current reshaped control strategy for each optical communication channel from the mixed response, such as the real-time waveform of the microheater drive current. Based on the newly extracted closed-loop thermal tuning control signal and the corresponding photoelectric response signal, the coupled system identification process is executed again. This process employs an adaptive recursive parameter estimation algorithm to update the estimation of the system dynamic model and calculate a new cross-domain coupling risk characteristic matrix reflecting the current system state. This matrix encapsulates the residual dynamic coupling relationships and potential oscillation mode information between all optical communication channels after the control strategy adjustment. Resonance risk assessment analysis is performed on the newly obtained cross-domain coupling risk feature matrix. The real parts of all eigenvalues ​​of the matrix are calculated, and it is checked whether any real parts still exceed a preset positive real part threshold. The eigenvalues ​​are determined, and the channel participation and phase synchronization of their corresponding eigenvectors are analyzed. If the analysis results show that the real parts of all eigenvalues ​​are below a threshold... If any residual unstable mode fails to meet the dual criteria of high participation and phase synchronization, then all previously identified resonant coupling risk modes are deemed to have been effectively suppressed, and the verification loop terminates. Conversely, if new or residual risk modes are still found, the dynamic reshaping step of the control strategy is restarted based on information such as its eigenvectors, and this verification loop is repeated until the system finally meets the convergence condition, ensuring that the co-packaged module operates at a dynamically stable operating point. For example, after two rounds of reshaping and verification, if the real parts of all eigenvalues ​​in the risk matrix are less than 0.01, the system declares test convergence.

[0029] The above formulas are all dimensionless calculations, and the preset parameters in the formulas should be set by those skilled in the art according to the actual situation.

[0030] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A photoelectric collaborative testing device for a co-packaged module, characterized in that, include: The synchronous excitation and acquisition module synchronously applies electrical domain test signals and optical domain test signals to the co-packaged module, and synchronously acquires its electrical domain response signals and optical domain response signals. The multi-domain signal analysis module analyzes the electrical domain response signal and the optical domain response signal, and separates and extracts the closed-loop thermal tuning control signal of each optical communication channel. The coupled system identification module, based on the extracted closed-loop thermal tuning control signal, performs multivariable system identification and intrinsic mode decomposition on the integrated system composed of optical communication channels, thereby extracting the cross-domain coupling risk feature matrix; The resonance risk assessment module analyzes the eigenvalues ​​and eigenvector distributions of the cross-domain coupling risk feature matrix to determine whether there are resonant coupling risk modes between optical communication channels. The control strategy dynamic reshaping module, when it is determined that there is a resonant coupling risk mode, performs adaptive dynamic reshaping of the closed-loop thermal tuning control strategy of the optical communication channel with the resonant coupling risk mode based on the eigenvector corresponding to the resonant coupling risk mode. The iterative convergence verification module repeats the synchronization excitation and acquisition steps up to the resonance risk determination step after the closed-loop thermal tuning control strategy is dynamically reshaped, until all risk modes in the cross-domain coupling risk feature matrix are suppressed.

2. The optoelectronic collaborative testing device for a co-packaged module according to claim 1, characterized in that, The synchronous excitation and acquisition module includes a programmable arbitrary waveform generator and a tunable laser source. The programmable arbitrary waveform generator is configured to generate an electrical domain test signal containing a high-frequency power supply noise spectrum and a high-speed data code pattern, and the tunable laser source is configured to generate an optical domain test signal whose wavelength can be rapidly modulated. The synchronous excitation and acquisition module also includes a high-speed digital storage oscilloscope and an optical waveform analyzer, which are used to synchronously trigger and acquire electrical domain response signals and optical domain response signals, respectively.

3. The optoelectronic collaborative testing device for a co-packaged module according to claim 1, characterized in that, The multi-domain signal analysis module includes a high-speed data acquisition unit and a digital signal processor. The digital signal processor is configured to perform joint time-frequency analysis on the acquired electrical domain response signal and optical domain response signal, and to separate and demodulate the micro heater drive current signal or thermoelectric cooler control voltage signal corresponding to each optical communication channel from the mixed response signal through blind source separation algorithm or digital filtering, as a closed-loop thermal tuning control signal.

4. The optoelectronic collaborative testing device for a co-packaged module according to claim 1, characterized in that, The coupled system identification module is configured to perform the following numerical calculation process: abstract the integrated system composed of optical communication channels into a multivariable dynamic system model, in which the closed-loop thermal tuning control signal of each channel is regarded as the system input, and the response signal after photoelectric conversion of each channel is regarded as the system output; An adaptive recursive parameter estimation algorithm is used to update the coupling parameter matrix in the multivariable dynamic system model online in real time. Modal perturbation analysis is performed on the updated coupling parameter matrix to solve for all its right eigenvectors and corresponding complex frequency eigenvalues. The cross-domain coupling risk feature matrix is ​​obtained by calculating the tensor shrinking product of the eigenvector matrix, the diagonal matrix of complex frequency eigenvalues, and a weight matrix determined by the physical spacing of each channel.

5. The optoelectronic collaborative testing device for a co-packaged module according to claim 4, characterized in that, The specific steps for finding all its right eigenvectors and corresponding complex frequency eigenvalues ​​include: An iterative numerical method of matrix spectral decomposition is used to transform the coupling parameter matrix into a Shanghai Senberg matrix; The Sjösenberg matrix is ​​solved by diagonalization using an implicit QR algorithm with displacement, where the displacement is dynamically determined by calculating the eigenvalues ​​of the last second-order principal submatrix of the current iteration matrix. The algorithm iterates until the modulus of all off-diagonal elements is less than a preset small tolerance threshold. The iteration matrix converges into a quasi-upper triangular matrix, and the second-order blocks or first-order elements on its diagonal give the complex frequency eigenvalues. For each obtained complex frequency eigenvalue, the corresponding linear equation system is solved by back substitution to obtain a non-zero solution vector. After normalization, this solution vector constitutes the corresponding right eigenvector.

6. The optoelectronic collaborative testing device for a co-packaged module according to claim 5, characterized in that, The coupling parameter matrix is ​​transformed into a Shanghai Senberg matrix using an iterative numerical method of matrix spectral decomposition, as follows: A numerical algorithm based on Haushold reflection transformation is adopted, which operates on each column of the matrix in turn: For the k-th column, k starts from 1, the numerical algorithm constructs the Haushold transformation matrix, which is designed to make all elements from the (k+2)-th row to the last row in the current column zero, while maintaining their orthogonality. Multiply this Haushold transformation matrix both on the left and on the right into the matrix to be transformed to complete an orthogonal similarity transformation. The orthogonal similarity transformation eliminates all elements from the (k+2)th row to the last row in the current column while keeping all eigenvalues ​​of the matrix unchanged. Repeat this process, processing columns one through three from the first to the third to last. After a series of such orthogonal similarity transformations, the original coupling parameter matrix is ​​transformed into a Shanghai Senberg matrix, which is a matrix in which all lower triangular elements except for the main diagonal and the first diagonal are zero.

7. The optoelectronic collaborative testing device for a co-packaged module according to claim 6, characterized in that, The specific steps for constructing the Haushold transformation matrix include: Select a sub-vector from the (k+1)th element to the last element in the k-th column of the matrix to be transformed; Calculate the Euclidean norm of the subvector and construct a reflection vector based on the norm value and the first element of the subvector. Using this reflection vector, the corresponding projection matrix is ​​generated through outer product operation, and then the final Haushold transformation matrix is ​​calculated.

8. The optoelectronic collaborative testing device for a co-packaged module according to claim 1, characterized in that, The specific steps for analyzing the eigenvalues ​​and eigenvector distributions of the cross-domain coupling risk feature matrix to determine the resonant coupling risk mode include: Calculate the real part of the eigenvalues, and set the real part greater than a preset positive real part threshold. The oscillation mode corresponding to the eigenvalues ​​is labeled as an unstable mode; where ; Calculate the component amplitudes of the eigenvectors corresponding to unstable modes, and identify those whose amplitudes exceed a preset channel participation threshold. The optical communication channel corresponding to the component is marked as a high participation channel; Calculate the phase difference between the eigenvector components corresponding to any two high-participation channels. If the absolute value of the phase difference is less than a preset phase synchronization threshold... If so, it is determined that there is a resonant coupling risk mode among these high-participation channels, and its risk level is jointly quantified by the real part of the corresponding eigenvalue and the participation level.

9. The optoelectronic collaborative testing device for a co-packaged module according to claim 8, characterized in that, The specific steps for the control strategy dynamic reshaping module to perform adaptive dynamic reshaping include: Based on the amplitude and phase of each channel component in the eigenvector corresponding to the resonant coupling risk mode, the target adjustment amount of the thermal tuning control strategy for each high-risk channel is calculated. This adjustment amount is proportional to the amplitude of the component in the eigenvector of that channel and opposite in phase. The high-risk channel is determined by whether the amplitude of the corresponding component of the high-participation channel in the eigenvector exceeds the channel participation threshold. And whether the absolute value of its phase difference with the dominant channel is less than the phase synchronization threshold. Determined comprehensively; The integral time constant of the closed-loop thermal tuning control strategy for high-risk channels is adjusted by weighting, and the adjustment weight is determined by the real part of the corresponding eigenvalue. An adaptive notch filter is introduced into the closed-loop thermal tuning control strategy. The center frequency of the adaptive notch filter is calculated based on the imaginary component of the eigenvalue that generates the resonant coupling risk mode, and its suppression depth is adaptively adjusted according to the amplitude of the corresponding channel component in the eigenvector.

Images