Optical test function high and low temperature test box control system for electronic components

Abstract

The invention provides an optical test function high and low temperature test box control system for electronic components, relates to the technical field of optical test, and provides an optical test function high and low temperature test box control system for electronic components through innovative technologies such as a double-target real-time sensing module, core temperature feed-forward compensation, window distortion active compensation, data validity discrimination and adaptive scheduling. The real junction temperature of an electronic component in a high and low temperature test box is highly consistent with the ambient temperature, the wavefront aberration of an observation window is counteracted in real time, abnormal data are automatically discriminated and remeasured, the temperature control precision is improved to + / -0.1 DEG C and the wavefront distortion residual error is reduced to lambda / 20 through dynamic temperature point selection based on gauss process regression and thermal inertia model game, and the temperature control precision is improved to + / -0.1 DEG C and the thermal inertia model game. The test period is shortened by more than 30%, the data misjudgment rate is lower than 1%, and the test precision, fidelity and efficiency are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of optical testing, in particular to a high-low temperature test box control system for optical testing of electronic components. BACKGROUND

[0002] In recent years, with the increasing demand for performance of optoelectronic devices in the fields of optical communication, autonomous driving and consumer electronics, high-precision optical testing technology in a wide temperature range has developed rapidly. On the one hand, high-low temperature test box technology has evolved from the initial "black box" temperature cycle to "visualization" and "intelligentization", introducing advanced algorithms such as real-time PID and model predictive control (MPC) to meet the fine adjustment of device junction temperature; on the other hand, optical testing technology has gradually expanded from traditional two-dimensional image field measurement to multi-dimensional characterization based on wavefront sensing and coherent imaging, which can capture spot morphology and wavefront aberration at the micron or even sub-micron scale. In addition, with the maturity of hardware platforms such as transparent micro-heating arrays and MEMS micro-mirrors, researchers have begun to try to build programmable temperature gradients in the observation window to achieve online compensation for optical distortion. Although the above technologies have made significant progress, most of them still separate temperature control, optical compensation and test scheduling into independent modules, making it difficult to meet the demand for "temperature-optical-data" full-link collaborative optimization.

[0003] In the prior art, a split test chamber with thermal imaging function is disclosed in CN118533751B. The test chamber includes a test chamber body and a control device. The test chamber body is internally provided with a test cavity, a high-pressure test device, a temperature test device, and a transmission device. The high-pressure test device and the temperature test device are used in combination to conduct comprehensive testing on the measured sample. The test chamber body is provided with an infrared imager at the top. The test device separates the test chamber body and the control mechanism, improving the safety of the entire test process. At the same time, the test chamber body can perform comprehensive testing combining high pressure and environmental temperature, and the temperature of the product is tracked in real time by the infrared imager throughout the test, generating a temperature camera image, which more directly reflects the temperature change of the test product during the overall test process, providing more accurate epoxy resin test data, and more conducive to evaluating the health status of the epoxy resin.

[0004] The prior art has significant deficiencies in wide temperature range testing:

[0005] 1. First, the test chamber relies only on the shell temperature feedback, and cannot sense the internal junction temperature in real time. The deviation of the junction temperature caused by temperature self-heating and environmental fluctuations directly leads to testing errors;

[0006] 2. Second, the traditional observation window cannot correct the wavefront aberration caused by temperature difference online, and the spot distortion seriously affects the accuracy of optical performance evaluation;

[0007] 3. Furthermore, multi-temperature point testing often uses static preset sequences, which lacks a dynamic balance between testing value and cost, resulting in wasted resources and difficulty in capturing curve inflection points; at the same time, test data is mostly "screened after the fact," lacking real-time validity identification and automatic retesting mechanisms.

[0008] The information disclosed in the background section above is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0009] The purpose of this invention is to provide a high and low temperature test chamber control system for optical testing of electronic components, so as to solve the problems mentioned in the background art.

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] The high and low temperature test chamber control system for optical testing of electronic components specifically includes:

[0012] Dual-target real-time sensing module: used to simultaneously acquire device status parameters reflecting the core operating temperature of the electronic component under test, and window status parameters reflecting the optical transmission characteristics of the observation window of the test chamber; the window status parameters include wavefront aberration information;

[0013] Core temperature feedforward compensation module: Based on the device state parameters and the preset device thermal resistance model, it calculates the estimated core temperature of the electronic component under test, and dynamically generates an environmental temperature adjustment command for the high and low temperature test chamber according to the deviation between the estimated core temperature and the target core temperature.

[0014] Window distortion active compensation module: Based on the window state parameters, it generates and executes heating power control commands for each heating unit in the transparent micro heating array to generate a compensating temperature gradient to offset the wavefront aberration.

[0015] Test execution and data synchronization module: When the deviation between the core temperature estimate and the target core temperature and the window state parameters both meet the preset stable threshold, the optical test instrument is triggered to perform multi-dimensional optical data acquisition, and the acquired optical data is synchronously associated with the corresponding device state parameters and core temperature estimate to generate a single-point optical performance dataset.

[0016] Data validity identification module: used to perform real-time statistical analysis on the single-point optical performance dataset. When the statistical characteristic value of the optical data in the single-point optical performance dataset exceeds the preset control limit, the single-point optical performance dataset is determined to be invalid, and a retest instruction is generated to the test execution and data synchronization module.

[0017] Test flow adaptive scheduling module: used for calculating the "test value index" of each to-be-tested temperature point by the game model based on the "expected information gain" predicted based on the Gaussian process regression and the "time cost" predicted based on the system thermal inertia model, and dynamically selecting the next optimal test temperature point in the test sequence.

[0018] Compared with the prior art, the beneficial effects of the present application are: through the introduction of a plurality of innovative technologies such as double-target real-time sensing, core temperature feedforward compensation, window distortion active compensation, data validity discrimination and adaptive scheduling in the high-low temperature test chamber control system, the key short board of traditional wide temperature range optical testing is overcome:

[0019] Accurate junction temperature restoration eliminates self-heating deviation - the double-target real-time sensing module synchronously acquires the device junction temperature state and the external observation window state, the core temperature feedforward compensation module dynamically issues environmental temperature adjustment instructions based on the device thermal resistance model, so that the internal temperature of the test chamber is highly consistent with the real junction temperature of the electronic component, and the temperature measurement error caused by self-heating or environmental fluctuations is reduced;

[0020] Precise optical path correction eliminates window distortion - the window distortion active compensation module uses a transparent miniature heating array to generate a programmable temperature gradient on the window to real-time offset wavefront aberration, ensuring that the light beam passing through the window maintains the original wavefront appearance, and significantly improves the fidelity of spot and wavefront measurement.

[0021] Intelligent data filtering ensures reliable results - the data validity discrimination module analyzes the statistical characteristics of the collected data in real time during the test execution process, and automatically determines the data exceeding the control limit as invalid and triggers retesting, eliminating the low efficiency and risk of missed judgment of manual screening in the later stage, and ensuring the high reliability of each set of optical performance data set.

[0022] Dynamic scheduling accelerates testing and optimizes resource utilization - the test flow adaptive scheduling module combines the uncertainty predicted by the Gaussian process regression and the switching time cost predicted by the system thermal inertia model, calculates the test value index through game fusion, dynamically selects the optimal next to-be-tested temperature point, can focus on capturing the performance curve inflection point, and can skip the redundant area, and realizes the substantial shortening of test time and the maximization of information acquisition. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 The present application is a whole method flowchart;

[0024] Figure 2 The present application is a high-low temperature test chamber schematic diagram. DETAILED DESCRIPTION

[0025] In order to make the above objectives, characteristics and advantages of the present application more apparent, a detailed description of the specific embodiments of the present application will be given below with reference to the accompanying drawings.

[0026] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. However, it will be apparent to one skilled in the art that the present application can be practiced without the specific details given herewith. In other instances, well-known methods have not been described in detail in order to avoid unnecessarily obscuring the present application. Therefore, the specific embodiments given herewith are not to be construed as limiting, but merely as being given by way of illustration.

[0027] Embodiment one:

[0028] Please refer to Figure 1 and Figure 2 , the present application provides a technical solution:

[0029] The optical test function high and low temperature test box control system for electronic components is applied to a high and low temperature test box containing an observation window and a transparent micro-heating array, and specifically comprises:

[0030] Dual-target real-time sensing module: used for synchronously acquiring a device state parameter reflecting the core working temperature of the electronic component to be tested and a window state parameter reflecting the optical transmission characteristics of the observation window of the test box; the window state parameter includes wavefront aberration information;

[0031] Further explanation: 1.1) The dual-target real-time sensing module of this embodiment: its function is to synchronously and non-invasively acquire two types of core parameters that serve as the basis for decision-making of the control system. This module is physically integrated on the control cabinet and the box body of the high and low temperature test box, and its specific composition and operation method are as follows:

[0032] The "electronic component to be tested" is abbreviated as DUT, or "device"; the dual-target real-time sensing module includes a device state sensing unit and a window state sensing unit;

[0033] The device state sensing unit is used for accurately quantifying the self-heating level of the electronic component to be tested when it is powered on and working, thereby obtaining the device state parameter;

[0034] Hardware deployment: a high-precision four-quadrant source measurement unit (SMU, Keithley-2400 series) is connected in series on the power supply circuit of the DUT. The SMU is connected to the DUT test fixture installed inside the test box through a shielded cable.

[0035] The SMU is configured in Source-Voltage, Measure-Current (SMU) or Source-Current, Measure-Voltage (SMU) mode. In each excitation step of the test sequence, the SMU synchronously acquires and outputs the driving current flowing through the DUT and the voltage applied across the DUT at a sampling rate no less than 1 kHz. The processor in the control system receives these data streams.

[0036] The processor calculates the acquired data in real time to obtain the instantaneous input electric power of the DUT; specifically represented as P DUT,in (t) = I DUT (t) x V DUT (t); it is to be noted that the calculation formula conforms to Ohm's law and is not described in detail;

[0037] Wherein, P DUT,in (t) is the instantaneous input electric power of the DUT at time t, and the unit is watt (W); I DUT (t) is the driving current measured by the SMU at time t, and the unit is ampere (A); V DUT (t) is the voltage measured by the SMU at time t, and the unit is volt (V);

[0038] The window state perception unit is used to quantitatively measure the dynamic optical wavefront distortion of the observation window of the test chamber caused by the internal and external temperature difference in real time and comprehensively, so as to obtain the window state parameters, including the Zernike coefficient; hereinafter, the "observation window of the test chamber" is referred to as "observation window";

[0039] Hardware deployment: outside the test chamber, along the main test optical axis, a Shack-Hartmann wavefront sensor (SHWFS) is fixedly installed. Inside the test chamber, a laser diode module is installed at the corresponding position, which emits a collimated and wavelength-stable reference light beam. In this embodiment, the reference light beam is selected as a HeNe laser with a wavelength of 632.8 nm, which is perpendicular to the observation window and enters the external SHWFS;

[0040] Specific implementation: the microlens array inside the SHWFS divides the received reference light beam into multiple sub-beams and focuses them on the CCD / CMOS detector to form a spot array. When the observation window produces thermal distortion, the wavefront will be distorted, causing the positions of the spots in the spot array to deviate from their reference positions. The built-in processor of the SHWFS reconstructs the phase distribution of the incident wavefront in real time according to these deviations.

[0041] The reconstructed wavefront phase distribution φ(x, y) is decomposed into a set of normalized Zernike polynomial coefficients; the control system of the embodiment focuses on the Zernike polynomial coefficients related to the main optical aberrations;

[0042] Zernike decomposition: wherein c i is the coefficient of the i-th Zernike polynomial, Z i (x, y) is the polynomial base function corresponding to the phase (x, y) of the i-th Zernike polynomial; N is the total number of Zernike polynomial coefficients;

[0043] The control system extracts the defocus coefficient Z defocus and the 0° / 45° astigmatism coefficient Z astigmatism,0 and Z astigmatism,45 ; it should be noted that in the embodiment, Z defocus = c4; Z astigmatism,0 = c5;

[0044] Z astigmatism,0 = c6; the wavefront aberration information is characterized as a comprehensive optical distortion index; specifically, the coefficients of the Zernike polynomials are combined into a dimensionless comprehensive optical distortion index Δd optic ; the calculation formula is characterized as wherein,

[0045] Z defocus , Z astigmatism,0 , and Z astigmatism,45 : the Zernike coefficients obtained from the wavefront sensor in real time, with units of multiples of wavelength or microns; α1, β1 are preset weighting coefficients for adjusting the relative importance of defocus and astigmatism in the comprehensive index, which are set according to the specific requirements for focus point stability and spot shape; for fiber-coupled testing, the weight of α1 is higher.

[0046] In another embodiment: the dual-target real-time perception module is only responsible for "providing data"; in extreme temperature and complex electromagnetic environments, the sensor itself may appear noise, drift or transient interference; the embodiment scheme evaluates the data quality in real time, so that the control system can identify the information quality of "providing data"; the following is an expansion scheme of the embodiment:

[0047] 1.2) The dual-target real-time perception module further comprises an evaluation unit for perceiving data quality, which is used to calculate and output a perception data quality index representing the overall reliability of the current perception data in real time based on the statistical fluctuation characteristics of the device state parameters and the window state parameters within a short time interval through a cross-weighted fusion algorithm. The specific implementation content is as follows:

[0048] The statistical fluctuation characteristics of the device state parameters and the window state parameters in a short time interval are defined as the device state instability index and the window state instability index, respectively.

[0049] The evaluation unit for sensing data quality performs statistical analysis on the collected raw data in a sliding time window of 100 ms, and performs the following calculation steps:

[0050] 1.21) Calculate the short-term instability index of each parameter: device state instability index: this index reflects the noise level of the DUT electrical parameters; by collecting multiple samples of the driving current and the voltage across the terminals in a sliding time window, the standard deviation is calculated, and an exponential function with saturation effect is used to map the standard deviation to a normalized (0, 1) interval, thereby obtaining the device state instability index; the calculation formula is represented as:

[0051] Q DUT =1-exp(-k I ·σ I -k V ·σ V );

[0052] where Q DUT is the device state instability index, with a value range of (0, 1), reflecting the noise level of the DUT electrical parameters; σ I ,σ V are the standard deviations of the driving current I DUT and the voltage V DUT sample sets collected in the sliding time window Δt, which are directly calculated from the collected data. k I ,k V are the current and voltage instability conversion coefficients, respectively. These two coefficients are preset according to the baseline noise level of the SMU and the test accuracy requirements, and are used to adjust the sensitivity of the exponential function.

[0053] 1.22) Calculate the window state instability index: by collecting multiple Zernike coefficient samples in a sliding time window, the standard deviation is calculated, and a similar exponential function is used to map the standard deviation to a normalized (0, 1) interval, thereby obtaining the window state instability index. The calculation formula is represented as:

[0054] where Q optic is the window state instability index, reflecting the jitter degree of the window optical distortion parameters, with a value range of (0, 1).

[0055] σ Zdefocus ,σ Zastigmatism,0 ,σ Zastigmatism,45: Standard deviation of the corresponding Zernike coefficient sample set collected in the sliding time window Δt, unit consistent with Zernike coefficient. These values are calculated directly from the wavefront sensor output data.k Z : Zernike coefficient instability conversion coefficient, unit is the reciprocal of Zernike coefficient unit; preset according to the measurement accuracy of the wavefront sensor and the stability requirement of the optical path; exp is the abbreviation of exponential function.

[0056] 1.23) Through the cross-weighted fusion mechanism, the device state instability index and the window state instability index are fused; wherein the weight of each instability index not only depends on its own reference importance, but also is dynamically affected by the size of the other instability index; finally, the weighted fusion total instability is subtracted from 1 to obtain the perception data quality index representing the overall credibility of the current perception data; the specific implementation content is as follows:

[0057] Through principal component analysis (PCA) method, a large amount of historical data is analyzed offline to determine the contribution degree of electrical noise and optical jitter to the error of the final measurement result in a typical test, and the contribution degree is taken as the basis weight W base,DUT and W base,optic of the corresponding parameter respectively; in this embodiment, the PCA analysis shows that the contribution of electrical noise is 30% and the contribution of optical jitter is 70%, so W base,DUT = 0.3, W base,optic = 0.7.

[0058] The dynamic cross-influence weight is calculated as follows: W dyn,DUT = W base,DUT + γ1·Q optic ; W dyn,optic = W base,optic + γ1·Q DUT ;

[0059] The maximum weight is obtained by normalization:

[0060] The perception data quality index PDQI is calculated as follows: PDQI = 1-(w DUT ·Q DUT +w optic ·Q optic ); wherein the value range of PDQI is designed to be (0, 1). W base,DUT ,W base,optic : the basis weight determined by offline analysis, and the sum is 1. γ1 is the cross-influence factor, which is a preset hyperparameter, and in this embodiment, it is set to 0.5, which is used to control the strength of cross-influence; W dyn,DUT ,W dyn,optic are the dynamic original weights considering cross-influence. w DUT ,w opticis the normalized final dynamic weight, and the sum is 1;

[0061] This embodiment greatly improves the intelligence and robustness of the system. Instead of passively accepting all data, it actively and in real time labels each frame of data with a "confidence" perception data quality index.

[0062] The perception data quality index PDQI provides a decision basis for the upper control system. The control system initially sets a PDQI threshold of 0.9. When the PDQI is lower than this value, the control system can automatically take measures, specifically: 1) suspend the test process and wait for the sensor data to stabilize; 2) start an alarm to prompt the operator to check for external vibration sources or electromagnetic interference; 3) temporarily increase the order or time constant of the digital filter to smooth the data. This avoids temperature compensation or optical compensation based on unreliable "garbage data", ensuring the stability of the entire closed-loop control and the accuracy of the final test results from the source;

[0063] The design of the cross-weighted fusion algorithm has a deep physical meaning. For example, when the window optical distortion is severely dithered, Q optic increases, meaning that the test light path through the window itself is unreliable. At this time, even if the DUT's electrical signal appears stable, the reliability of the measured optical performance data is also reduced. The algorithm amplifies any electrical instability Q DUT by increasing the weight of w DUT , which negatively affects the final perception data quality index.

[0064] The effective value range of PDQI is limited to the interval (0, 1), and the closer to 0 represents lower quality, and the closer to 1 represents higher quality. This provides a unified evaluation scale for data quality under different time and test conditions, making it easy to set a unified threshold and control logic, and also facilitating data visualization and long-term quality trend monitoring.

[0065] This calculation is based on the basic idea of statistical process control (SPC), that is, the output parameters of a stable process should show small random fluctuations around their mean. When the standard deviation represented by the fluctuations exceeds the normal range, it indicates that the process is out of control or disturbed. The formula uses the saturation property of the exponential function to map the unbounded standard deviation to a bounded instability index, and then through the innovative method of cross-weighted fusion, it simulates the coupling effect of multiple risk sources.

[0066] When PDQI approaches 1: the working state of the current perception system is better; the "perception data reliability" is higher; according to the PDQI calculation formula, Q DUT → 0 means that the standard deviation σ I and σ VThe smaller, the more stable the reading of the SMU, and the less the noise interference of the DUT working state. Similarly, Q optic →0 means that the standard deviation σ Z of each Zernike coefficient is smaller, indicating that the reading of the wavefront sensor is more stable, and the jitter of the optical path of the observation window is smaller. At this time, the device state parameters and window state parameters obtained by the dual-target real-time perception module have higher signal-to-noise ratio and credibility;

[0067] When PDQI tends to be closer to 0: indicates that the abnormal or untrusted state of the current perception system is more serious; if the test box is subjected to severe mechanical vibration, the relative displacement of the reference light source and the wavefront sensor will be caused, so that the standard deviation σ Z of the Zernike coefficient will increase sharply, so that Q optic →1. According to the cross weighting formula, the weight of w optic will increase sharply, and the weight of w DUT will also increase due to the increase of Q optic ; this causes w optic ·Q optic to increase, and any slight electrical noise Q DUT will be further amplified by the amplified weight w DUT , which will jointly pull PDQI to 0. Technical effect: this indicates that the more serious the perception data is polluted, the more untrusted it is. At this time, if the test continues, the probability of obtaining an error result is higher. The system determines whether to abort the test process according to the PDQI value and the preset experience value, so as to avoid wasting time and prevent error data from polluting the database, which fully embodies the robustness of the design.

[0068] The core temperature feedforward compensation module is used to calculate the core temperature estimation value of the electronic component to be tested based on the device state parameters and a preset device thermal resistance model, and to dynamically generate an environmental temperature adjustment instruction for the high-low temperature test box according to the deviation between the core temperature estimation value and a target core temperature;

[0069] Further explanation: For the existing "high-low temperature test box environmental temperature adjustment instruction", it relies on single electrical measurement results, which is disturbed by instantaneous noise; this embodiment intelligently fuses them by running two preset device thermal resistance models in parallel, a slow-changing model representing long-term physical trends and a fast-changing measurement reflecting immediate electrical characteristics, and introducing a "confidence" indicator, thereby quickly responding to real changes and effectively suppressing measurement noise, achieving a perfect balance between precision and stability; this embodiment not only "uses" the model, but also continuously "improves" the model. Through confidence weighting, the system can determine when to "believe" the measurement data and use it to correct the physical model, and when to "doubt" the measurement data and rely more on the historical model. This forms a complete "estimation-measurement-comparison-evaluation-fusion-learning" closed loop, enabling the thermal resistance model to evolve autonomously as the device ages and operating conditions change, which is not possessed by the prior art.

[0070] The following dynamic thermal resistance fusion and adaptive correction mechanism based on dual model consistency and confidence weighting is constructed;

[0071] 2.1) The core temperature feedforward compensation module includes a dynamic thermal resistance fusion unit; the dynamic thermal resistance fusion unit is used to obtain a physical thermal resistance estimation value based on a long-term physical model and an electrical thermal resistance measurement value based on an instantaneous electrical characteristic measurement in parallel, and generate a physical model confidence index by calculating the consistency deviation between the two, and finally fuse the two thermal resistance values according to the confidence index to generate a dynamic thermal resistance value for calculating the core temperature estimation value;

[0072] 2.2) The specific implementation content is that the dynamic thermal resistance fusion unit updates the dynamic thermal resistance value for core temperature estimation in each temperature control cycle by the following steps:

[0073] First, the dynamic thermal resistance fusion unit maintains a physical thermal resistance estimation value representing the long-term and stable thermal characteristics of the device. At the same time, at each temperature stable point, the dynamic thermal resistance fusion unit performs a small electrical scan to calculate an electrical thermal resistance measurement value reflecting the current instantaneous heat dissipation capability using the pre-calibrated device electrical parameter-temperature relationship.

[0074] Subsequently, the dynamic thermal resistance fusion unit calculates the model consistency deviation between the two thermal resistance values. Based on the deviation, the dynamic thermal resistance fusion unit calculates a physical model confidence index through a preset nonlinear mapping function, which represents the credibility of the current physical thermal resistance estimation value.

[0075] Then, the dynamic thermal resistance fusion unit uses the confidence index as a weight to perform confidence weighted fusion on the physical thermal resistance estimation value and the electrical thermal resistance measurement value, to obtain a final dynamic thermal resistance value that is both stable and accurate.

[0076] Finally, the final dynamic thermal resistance value is used on one hand for the current core temperature estimation value calculation, and on the other hand, through a low-pass filtering algorithm, for online correction of the physical thermal resistance estimation value, so that it can slowly and stably track the long-term changes in the thermal characteristics of the device.

[0077] This mechanism of step 1.1) improves the accuracy and reliability of the core temperature estimation value. When the electrical measurement is disturbed by transient noise, causing abnormal results, the model consistency deviation will increase, and the confidence index will decrease, so the system will automatically reduce the dependence on this abnormal measurement, and rely more on the stable and long-term physical model, thereby showing strong anti-interference ability (robustness). Conversely, when the device ages and its real thermal resistance changes slowly and continuously, the electrical measurement value will continuously and stably deviate from the old physical model value, and the system will use this continuous deviation to gradually and smoothly update the physical model through the online correction mechanism, so that it can follow the changes in the real state of the device, and show excellent adaptive ability.

[0078] The physical model of the embodiment represents "experience and rules", the electrical measurement represents "instantaneous reading", the confidence index is a quantitative evaluation by the system of whether the "instantaneous reading" agrees with the "experience and rules", and the weighted fusion and online correction are the most reasonable comprehensive decisions based on this evaluation. This logic closed loop ensures that the system can output the thermal resistance value closest to the real situation in any case. For a detailed description of step 1.1), first, the following parameters are predefined:

[0079] Physical thermal resistance estimation value: R th,phy (unit: ℃ / W); Electrical thermal resistance measurement value: R th,elec (unit: ℃ / W); Model consistency deviation: δ R ; Confidence index of physical model: C phy ; Dynamic thermal resistance value: R th,dyn (unit: ℃ / W);

[0080] The corrected physical thermal resistance estimation value is represented as: R th,phy (k1+1) (unit: ℃ / W); The electrical thermal resistance measurement value R th,elec is calculated; The formula is derived from the "temperature sensitive electrical parameter (TSEP)" method in semiconductor physics. The basic principle is that the electrical characteristics (such as the forward voltage drop of the PN junction) of a semiconductor device have a nearly linear relationship with the junction temperature.

[0081] The formula composition principle is that a voltage drop difference can be obtained by measuring the forward voltage drop of the DUT under the current working state and subtracting the calibrated voltage drop of the DUT under the known reference temperature and corresponding to the same current. The temperature rise of the core relative to the reference point can be obtained by dividing the voltage drop difference by the pre-calibrated voltage-temperature coefficient. Finally, the current electrical thermal resistance measurement value can be obtained by dividing the temperature rise by the heat power that generates the temperature rise.

[0082] R th,elec The calculation formula is represented as:

[0083] Wherein, V DUT , I DUT , P DUT，in : all come from real-time acquisition data or calculation results of the "dual-target real-time perception module"; P opt is the output optical power (W) of the DUT; V calib (I DUT ) is a calibration function or lookup table, which represents the terminal voltage of the DUT when the current is I DUT at the reference temperature. This function needs to be calibrated once for the DUT before testing; K TSEP is the voltage-temperature sensitivity coefficient of the DUT, which is a device characteristic constant that needs to be pre-calibrated, indicating the change amount of the forward voltage drop when the junction temperature rises by 1℃;

[0084] The confidence index C phy of the physical model is calculated; the formula is based on the confidence modeling idea of error decay, and an exponential decay function is adopted. First, the normalized relative deviation between the physical model estimated value and the electrical model measured value is calculated, which directly reflects the consistency of the two models. Then, the deviation value is substituted into an exponential decay function. When the deviation is zero, the function output is 1 (complete confidence); as the deviation increases, the function output rapidly decreases in an exponential form and tends to 0 (complete unconfidence). A small normal number is added in the denominator to prevent division by zero error in the theoretical case of zero physical thermal resistance.

[0085] C phy The calculation formula is represented as: Wherein, R th,phy is the estimated value of the physical thermal resistance at the current time, and its initial value comes from the device manual or empirical value; λ2 is the confidence decay coefficient, which is dimensionless and is an adjustable hyperparameter; the initial value of the embodiment is 5; it controls the decline speed of the confidence as the deviation increases. The larger λ2 is, the heavier the "penalty" of the system to the model deviation, and the faster the confidence decreases. R is a small normal number, and the value of the embodiment is 10 -6 , which is used to prevent the denominator from being zero and ensure the numerical stability of the calculation;

[0086] Computing dynamic thermal resistance value R th,dyn ; this formula is the "weighted average method", which is one of the most basic and commonly used algorithms in data fusion field; specifically, the physical thermal resistance estimated value and the electrical thermal resistance measured value are linearly weighted and summed. Among them, the weight of the physical thermal resistance estimated value is directly determined by its confidence index C phy , while the weight of the electrical thermal resistance measured value is set to 1-C phy . In this way, the higher the confidence of the model, the greater the influence on the final fusion result.

[0087] R th,dyn Computing formula representation: R th，dyn = C phy ·R th，phy +(1-C phy )·R th，elec ;

[0088] Online correction of physical thermal resistance estimated value R th,phy (k1+1); this update rule is a "first-order low-pass filter" or "exponential moving average (EMA)" algorithm in control theory and signal processing.

[0089] For the new physical thermal resistance estimated value, the old value at the last time and the dynamic thermal resistance value obtained by the current fusion are weighted and averaged. This makes the change of the physical model smooth and inertial, and can filter out the high-frequency noise in the dynamic thermal resistance, only responding to its long-term and continuous change trend.

[0090] R th,phy (k1+1)Computing formula representation: R th,phy (k1+1) = (1-η1)·R th,phy (k1)+η1·R th,dyn ;

[0091] Where R th,phy (k1+1), R th,phy (k1) represent the physical thermal resistance estimated value at the k1+1th and k1th control period respectively; η1 is the learning rate, dimensionless, which is a small constant between (0, 1). It determines the speed of the physical model "learning" from the current measurement result. The smaller η1, the slower the model updates, the more stable; the larger η1, the faster the model updates, the more sensitive to new data. The value range of C phy is naturally limited to (0, 1] by the exponential function. Parameter "0" or "1" case: when R th,phy =R th,elec , C phy takes 1, but theoretically cannot take 0, only approaches to 0 infinitely.

[0092] When C phyThe closer the output is to 1, the higher the consistency between the physical model and the electrical measurement results; C phy The closer it is to 1, the closer its exponent power is to 0, i.e., |R|. th,phy -R th,elec |→0. This indicates that the physical model R obtained through device manuals or long-term learning is correct. th,phy R measured in real time by the TSEP method th,elec The closer the values ​​are to the same, the more accurate the system's modeling of the DUT's thermal properties. At this point, the fusion formula R... th,dyn The result will approach R th,phy Or R th,elec The estimated core temperature has extremely high accuracy.

[0093] When C phy The closer the output is to 0, the greater the deviation between the physical model and the electrical measurement results, and the lower the system's "confidence" in the current physical model. Parameter change logic reasoning: C phy →0 means that the power of its exponent tends to negative infinity, which requires |R| to be zero. th,phy -R th,elec The sum of terms relative to the denominator is very large. This occurs in two cases: 1) the instantaneous electrical thermal resistance measurement R. th,elec 1) Outliers with severe anomalies due to noise interference; 2) The actual thermal resistance of the device has changed significantly due to aging and other reasons, causing the old physical model R to be invalidated. th,phy This deviates significantly from the facts. In this case, the fusion formula R... th,dyn It will be mainly composed of (1-C) phy )·R th,elec The dominant mechanism is that the system will decisively discard unreliable old models and instead trust the current real-time measurements. Simultaneously, this significant deviation will be corrected online by adjusting the estimated physical thermal resistance R. th,phy The "(k1+1)" update mechanism powerfully "pulls" the physical model to make corrections. This demonstrates the system's strong error correction and adaptive capabilities when faced with model failures.

[0094] Further explanation: The core temperature feedforward compensation module calculates the estimated core temperature based on a real-time measured ambient temperature value, plus the product of the net heat generation power of the device and the aforementioned fused dynamic thermal resistance value; the specific implementation details are as follows:

[0095] The formula for setting the core temperature estimate is derived from the steady-state thermal resistance law in heat conduction, which describes the relationship between the temperature difference between the two ends of an object and the heat flow power and thermal resistance of the object when heat flows through it; its form corresponds to Ohm's law.

[0096] The core temperature estimate is obtained by adding a real-time measured substrate temperature, which best represents the immediate heat dissipation environment of the DUT, to a temperature rise obtained by multiplying the net heat generation power of the DUT (i.e. the input electrical power minus the optical output power) by its online corrected dynamic thermal resistance. The calculation formula of the core temperature estimate is characterized as:

[0097] T j,est (k1) = T case (k1) + (P DUT,in (k1) - P opt (k1)) x R th,dyn (k1); wherein T j,est (k1) is the core temperature estimate calculated in the k1th control cycle, and the unit is Celsius (°C). T case (k1) is the real-time measured DUT shell or substrate temperature in the k1th control cycle, and the unit is Celsius (°C).

[0098] The embodiment needs to additionally install a high-precision temperature sensor (such as PT100 thermal resistance or NTC thermistor) on the test fixture, close to the shell of the DUT or the PCB board below it. The data provided by this sensor is the accurate basis for calculating the temperature rise. DUT,in (k1) is the real-time input electrical power of the DUT measured by the device state sensing unit in the k1th control cycle, and the unit is watt (W). opt (k1) is the estimated optical output power of the DUT in the k1th control cycle, and the unit is watt (W). th,dyn (k1) is the latest dynamic thermal resistance value calculated by the "dynamic thermal resistance fusion and adaptive correction mechanism based on dual-model consistency and confidence weighting" in the k1th control cycle, and the unit is Celsius per watt (°C / W).

[0099] It should be noted that (P DUT,in (k1) - P opt (k1)) is the total input electrical power of the device minus the power output in the form of light, and the difference is the power converted into heat, i.e. the heat generation power. Multiplying the heat generation power by the dynamic thermal resistance, the final dimension of the second term on the right side of the equation is temperature (°C). Therefore, all terms on both sides of the equation are ultimately unified to the dimension of "temperature", and through reasonable physical operation combination, it is in accordance with the basic equation of heat conduction effect in mathematics and physical logic.

[0100] By using the substrate temperature T case(k1) As a reference, instead of the air temperature inside the chamber, the accuracy and dynamic response speed of the core temperature estimation are improved. There is thermal resistance and thermal capacity between the air inside the chamber and the DUT substrate, which causes the change of the air temperature to always lag behind the substrate temperature. Taking the substrate temperature as a reference, the calculation model only needs to deal with the thermal resistance from the "substrate to the core", and the model is simpler and more accurate.

[0101] Further explanation: The high-low temperature test chamber has the typical industrial control problems of large inertia, pure lag, and nonlinearity. The traditional PID controller has fixed parameters, and it is difficult to achieve optimal control effect in the entire wide temperature range (-40°C to +105°C), often leading to slow response, large overshoot, and long stabilization time. This scheme uses the known DUT heat generation information to directly give an environmental temperature reference close to the ideal value through predictive feedforward, which fundamentally shortens the response time of the system.

[0102] On this basis, an adaptive fuzzy PID feedback is introduced, which simulates the control experience of human experts and dynamically adjusts the parameters of the PID controller in real time and online according to the current core temperature error and error change rate. This makes the controller "aggressive" when far from the target point to quickly approach, and "conservative" when close to the target point to prevent overshoot, perfectly overcoming the nonlinearity and time-varying characteristics of the system.

[0103] 2.3) The core temperature feedforward compensation module further comprises an environmental temperature instruction generation unit, which adjusts the proportional, integral, and differential parameters of the PID controller in real time through a fuzzy inference system based on the core temperature error and its core temperature error change rate of the core temperature estimation value and the target core temperature, and calculates and outputs the final environmental temperature set value applied to the high-low temperature test chamber in combination with a predictive feedforward quantity based on a device thermal resistance model;

[0104] The specific implementation is that the generation process of the environmental temperature instruction is divided into two parallel calculation steps: predictive feedforward quantity calculation and adaptive fuzzy PID feedback adjustment;

[0105] The predictive feedforward quantity is represented as T env,base : This component is based on a physical model, which calculates the net heat generation power of the DUT by using the online corrected DUT dynamic thermal resistance value and the measured DUT input electric power, and subtracting the estimated light output power, and then combining the target core temperature to inversely calculate the theoretical environmental temperature reference value required to achieve the target core temperature;

[0106] The calculation formula is represented as: T env,base = T j,set -(P DUT,in -P opt )×Rth,dyn ; wherein, T env,base is the predicted ambient temperature reference value in Celsius (°C); T j,set is the target core temperature set by the user in Celsius (°C). P DUT,in is the instantaneous input electrical power of the DUT measured in real time by the device state perception unit in Watts (W). This is a collectible data. P opt is the estimated optical output power of the DUT at the current operating point in Watts (W). This value is obtained by consulting the DUT specification book or pre-calibration as a known parameter input. R th,dyn is the online corrected dynamic thermal resistance value of the DUT in Celsius per Watt (°C / W).

[0107] The component of adaptive fuzzy PID feedback adjustment is denoted as ΔT adjust ; the following input variables are specifically set:

[0108] Core temperature error: E(k1) = T j,set - T j,est (k1); core temperature error rate of change:

[0109] EC(k1) = E(k1) - E(k1-1); wherein, k1 is the index of the current control period, T j,est (k1) is the core temperature estimation value at the current control period k1; the input variables are mapped to fuzzy language variables through membership functions; the language variable set is defined as {NB, NS, ZE, PS, PB}, representing {negative big, negative small, zero, positive small, positive big} respectively.

[0110] A triangular membership function is used for mapping, for the core temperature error E(k1), the domain is [-3, 3] °C, then the membership function definition of "zero" is:

[0111] Wherein, μ is the symbol of the membership function; ZE is the identification of the fuzzy set "zero (Zero)";

[0112] A fuzzy rule base based on expert experience is established, which is used to adjust the increments of the proportional gain, integral gain and derivative gain in the PID parameters respectively as ΔK p , ΔK i1 , ΔK d . And the fuzzy rule base of this embodiment uses the Mamdani fuzzy reasoning method. The Center of Gravity (Center of Gravity) is used for defuzzification, and the accurate increment values of the PID parameters are calculated, and then ΔK p , ΔK i , ΔK d are obtained respectively;

[0113] Real-time update of PID parameters: K p =K p,base +ΔK p ;K i1 =K i1,base +ΔK i1 ;K d =K d,base +ΔK d ;K p,base ;K i1,base and K d,base These are the basic proportional gain, basic integral gain, and basic differential gain; K p ;K i1 and K d These are the updated proportional gain, integral gain, and derivative gain; K p The "p" in K stands for Proportional; i1 In this context, i1 is an abbreviation for integral; K d The “d” in the text is an abbreviation for Derivative.

[0114] K p This determines the strength of the controller's response to the current core temperature error; K i1 This determines the strength of the controller's response to accumulated past errors. Its main function is to eliminate steady-state errors; K d This determines the controller's response strength to future error changes; the adjustment component is calculated using an incremental PID algorithm.

[0115] ΔT adjust (k1)=K p ·[E(k1)-E(k1-1)]+K i1 ·E(k1)+K d ·[E(k1)-2E(k1-1)+E(k1-2)];

[0116] The final environmental temperature command for price is characterized by T. env,new,set (k1)=T env,base (k1)+ΔT adjust (k1); where T env,new,set (k1) is the final ambient temperature setpoint output to the actuator of the high and low temperature test chamber in the k1th control cycle, in degrees Celsius (°C). ΔT adjust (k1) is the PID parameter adjustment in the k1th control cycle.

[0117] Compared with the traditional PID, the temperature stabilization time can be shortened by 30% to 50%, and the overshoot is suppressed within 0.1℃; when the working power of the DUT changes, the system can quickly re-lock the core temperature, ensuring the continuity and effectiveness of the test. The experimental embodiment of the above is as follows:

[0118] The embodiment is intended to verify the performance advantage of the proposed hybrid control strategy based on predictive feedforward and adaptive fuzzy PID feedback in accurately controlling the core junction temperature of a high-power laser diode (LD). A 10W butterfly packaged fiber-coupled laser diode with the model "Lumentum-S26-7602-100" is selected as the DUT. The typical working voltage of the DUT is 1.9V, and the working current is 7.5A. Its thermal characteristics are highly sensitive to environmental temperature and its own heat, making it an ideal object for testing the performance of a temperature control system. The test platform consists of a custom high-low temperature test chamber (temperature control range -55℃ to +125℃), the dual-target real-time sensing module, the core temperature feedforward compensation module, and the upper computer control software. To form a control, two control groups are set: the control group uses the industry standard and carefully adjusted traditional PID controller to regulate the environmental temperature of the test chamber; the invention group uses the hybrid control strategy described in the invention.

[0119] In the test preparation stage, the DUT is first pre-calibrated. The DUT is placed in a constant temperature environment of 25.0℃, and a high-precision four-quadrant source measurement unit (Keithley-2450-SMU) is used to scan its voltage-current (V-I) characteristic curve in the range of 0A to 8.0A with a step of 0.1A, and the relationship between the forward voltage drop and the temperature at very low power consumption (such as 0.1A) is recorded, so as to accurately calibrate the voltage-temperature sensitivity coefficient K TSEP of the DUT, which is -1.85mV / ℃, and a lookup table of the calibration function V calib (I DUT ) is generated. At the same time, through steady-state thermal testing, the initial physical thermal resistance estimate of the DUT is obtained as 6.5℃ / W.

[0120] First, start the control group test. The upper computer sets the target environmental temperature of the test chamber to an empirical value, and the PID controller K p =25; K i1 =0.8 and K d =3.5 starts working, with its input being the deviation of T j,est (k1) estimated by the traditional thermal resistance model from the target value T j,set . The test chamber starts to cool down, and the SMU drives the DUT to work at a constant current of 7.5A.

[0121] Subsequently, the test of the present application was carried out. After the system started, the core temperature feedforward compensation module first performed the predictive feedforward amount calculation. Based on the target core temperature T j，set = 85.0℃, the initial thermal resistance and the input power P DUT，in ≈ 14.25 of the DUT at 7.5A, the predictive ambient temperature reference value was calculated. At the same time, the adaptive fuzzy PID feedback adjustment module started to work. In the process of the test chamber temperature changing dramatically, the dynamic thermal resistance fusion and adaptive correction mechanism based on the consistency of the two models and the confidence began to work. Through the TSEP method, the electrical thermal resistance measurement value was calculated, which deviated from the slightly corrected physical thermal resistance estimation value. The system calculated the physical model confidence index accordingly, and fused to obtain a more accurate dynamic thermal resistance value, which was immediately used to update the core temperature estimation value. The fuzzy PID controller adjusted its PID parameters in real time according to the error and error change rate between the target core temperature and 85.0℃, and output the fine adjustment amount ΔT adjust (k1). The final ambient temperature instruction T env,new,set (k1) was synthesized from T env,base (k1) and ΔT adjust (k1). In the whole process, the system executed data acquisition, model correction and instruction generation at a frequency of 10Hz until the core temperature was stable. The test recorded the whole process data of the two control groups from the start to the core temperature entering and maintaining in the target interval of 85.0±0.1℃.

[0122] Table 1 Study of the mixed control strategy of predictive feedforward amount and adaptive fuzzy PID feedback:

[0123]

[0124]

[0125] The test data of the above examples clearly show the great advantages of the control strategy proposed by the present application.

[0126] 1. Significantly shorten the stabilization time: In the test of the core temperature rising from 25℃ to 85℃, the present application group only took 152 seconds, compared with the control group of 285 seconds, the stabilization time was shortened by 46.7%. This is due to the predictive feedforward amount which can quickly place the test chamber in a low temperature environment close to the final equilibrium point, accelerating the warming process of the DUT core. In the power step test simulating the DUT working mode switching, the advantage of the present application group is also obvious, and the time to restore stability is shortened by 49.5%.

[0127] 2. Extreme overshoot suppression capability: The traditional PID control of the control group produced a temperature overshoot as high as +1.8℃, which can be fatal to precision optical devices. While the self-adaptive fuzzy PID of the application group successfully suppressed the overshoot to +0.08℃, completely within the required 0.1℃. This outstanding stability is verified in all test scenarios.

[0128] 3. Excellent steady-state accuracy and robustness: After stabilization, the core temperature fluctuation range of the application group is only ±0.07℃, far superior to the control group's ±0.45℃, providing a solid and reliable temperature reference for subsequent high-precision optical performance testing. More importantly, in the "after aging" test after simulating the long-term work of the DUT, its true thermal resistance has changed. The control group's control performance deteriorates severely (stabilization time increases to 350 seconds, overshoot reaches 2.5℃) due to the use of a fixed thermal resistance model. While the application group benefits from the online correction of the dynamic thermal resistance model, it can quickly adapt to the changes of the DUT and still maintain excellent control performance, demonstrating strong robustness and self-adaptive ability.

[0129] 4. Significant energy saving effect: Due to faster stabilization and avoidance of unnecessary temperature repeated adjustment, the total energy consumption of the application group is significantly lower than that of the control group in completing the same test task, saving 37.8% in the main test.

[0130] 2.4) Further explanation: Define a controller adjustment intensity index, whose value range is designed in the interval (0, 1);

[0131] Set the controller adjustment intensity index, which combines the absolute value of the current core temperature error and the absolute value of the core temperature error rate through a normalized weighted norm, and then maps it to the interval (0, 1) through an inverse S-shaped function; to reflect the "effort" of the controller to correct the deviation.

[0132] The calculation formula is represented as: Where E ctrl is the controller adjustment intensity index, dimensionless, with a value range of (0, 1); ω E ,ω EC are the weight coefficients of the core temperature error and the core temperature error rate, respectively, dimensionless, set according to experience; ω E = 0.7; ω EC = 0.3; T norm , R norm are the normalized scale factors of the corresponding parameters, respectively, used to scale the core temperature error and the core temperature change rate to a comparable order of magnitude.

[0133] When E ctrlThe closer to 0, the more the control system is in the ideal steady-state locking state; the higher the trend of the "core temperature control accuracy", the fewer the number of output adjustment instructions of the control system. ctrl → 0 means that the denominator tends to 1. Only when E(k1)→ 0 and EC(k1)→ 0, the weighted term in the denominator tends to 0, making the entire denominator tend to 1. This precisely corresponds to the DUT core temperature has stabilized at the target value, and the trend of fluctuation is small. At this time, the controller tends to zero in its adjustment strength.

[0134] When E ctrl The closer to 1, the more the control system is in the dynamic adjustment process; the lower the trend of the "core temperature control accuracy", the greater the deviation being corrected by the system; E ctrl → 1 means that the denominator in the formula is larger; the weighted term in the denominator will become larger, leading to the entire denominator tending to infinity. The initial setting of this embodiment is when E ctrl When the value is above 0.85, it corresponds to two typical working conditions: 1) the test just starts, or a new target temperature point is just set, at which time the core temperature error is large; 2) the power of the DUT suddenly changes, causing the core temperature to change dramatically, at which time the core temperature error rate is large. In these two cases, the controller should adjust with the maximum strength. This index can accurately reflect this "emergency state", verifying its effectiveness as a dynamic performance index of the system. The setting of "0.85" in this embodiment is based on adaptive adjustment by the expert group according to actual experimental data.

[0135] The window distortion active compensation module is configured to generate and execute a heating power control instruction for each heating unit in the transparent micro-heating array based on the window state parameter, so as to generate a compensatory temperature gradient for offsetting the wavefront aberration.

[0136] Further explanation: This embodiment creatively introduces a compensation energy efficiency index as a second optimization target. Instead of "eliminating distortion at any cost", the system works under the more advanced principle of "eliminating distortion to the maximum extent with the minimum energy cost". This makes the compensation process more stable and efficient, and also prolongs the life of the transparent micro-heating array, which is a key point that has not been considered in the prior art.

[0137] It should be noted that: the transparent micro-heating array is an Indium Tin Oxide (ITO) film, which has high transmittance in the visible and near-infrared wavebands; the transparent micro-heating array is integrated on the inner surface of the observation window; the ITO film is precisely etched into the shape of an 8x8 independent heating unit array; each heating unit is provided with a micro-heater. The conductive lead connecting each heating unit leads the electrode to the edge of the window for connection to the external driving control circuit;

[0138] The compensatory temperature gradient distribution is generated by the window optical distortion active compensation actuator; on the inner surface of the observation window, there is deposited an array of ITO transparent heating films divided into 8x8 independent units; each unit is controlled by an independent PWM driving channel.

[0139] The window distortion active compensation module receives the comprehensive optical distortion index and the Zernike coefficient through the controller; to construct a pre-trained inverse neural network model by machine learning; the function of the inverse neural network model is: input the target Zernike coefficient to be offset, and output the power distribution map required by the entire 8x8 transparent micro-heating array; that is, P ij = f NN (-Z defocus ,-Z astigmatism ,...), where P ij is the target heating power of the i,j heating unit, P ij The heating power value required by the independent heating unit in the i row and j column of the value constitutes a complete "power distribution map" or "heat pattern", which generates a specific temperature gradient on the observation window, and then changes the refractive index distribution of the window glass, to generate an optical effect for compensating aberration; Z defocus represents the defocus coefficient in the Zernike coefficient; Z astigmatism represents the astigmatism coefficient in the Zernike coefficient; "..." represents that the input vector also contains other Zernike coefficients that need to be corrected;

[0140] The inverse neural network model f NN adopts the structure of multi-layer perception (Multi-Layer-Perceptron, MLP); the input layer is the number of Zernike coefficients used to describe and compensate wavefront aberration. The output layer is 64 neurons. Since the heating array is 8x8, 64 independent power values need to be output, and each neuron corresponds to a heating unit P ij .

[0141] Function: output a 64-dimensional vector, which is reshaped into an 8x8 power distribution map.

[0142] The output layer uses a linear activation function (Linear-Activation), because it allows the output to be any range of power values. If it is necessary to ensure that the output power is non-negative, ReLU or Softplus activation functions can also be used, or the output values can be processed outside the model.f NNis an inverse neural network model; an "inverse neural network model" is a model that takes a desired wavefront aberration as input and directly calculates the power distribution that would produce that aberration. The data used to train the network comes from finite element simulations and experimental calibration, i.e. applying a large number of different power distributions to the array and measuring the resulting aberration with a wavefront sensor.

[0143] 3.1) the window distortion active compensation module further comprises an iterative optimization control unit; this unit measures the residual wavefront aberration after the initial heating power instruction generated by the inverse model is executed, and calculates a compensation energy efficiency index based on the residual wavefront aberration and the total heating power currently applied; the iterative optimization control unit iteratively corrects the output of the inverse model according to the residual wavefront aberration and the compensation energy efficiency index, so as to minimize the residual wavefront aberration and maximize the compensation energy efficiency, thereby generating the final and optimal heating power control instruction;

[0144] 3.2) the specific implementation is as follows:

[0145] The iterative optimization control unit generates and optimizes the control instruction for the transparent micro-heating array in a closed-loop control cycle by performing the following steps: first, the iterative optimization control unit receives the initial window state parameters measured by the wavefront sensor, i.e. the original wavefront aberration information. Based on this information, a pre-trained inverse neural network model calculates a set of initial heating power distributions and applies them to the transparent heating film array on the inner surface of the observation window;

[0146] Next, after the initial heating power is applied, the wavefront sensor immediately measures the compensated optical wavefront to obtain a residual wavefront aberration; at the same time, the system calculates the total heating power currently applied to the entire transparent micro-heating array.

[0147] Then, the iterative optimization control unit uses a pre-set fusion formula to combine the amplitude of the residual wavefront aberration and the total heating power, which are two physical quantities that affect each other, into a single, standardized compensation energy efficiency index; the compensation energy efficiency index is used to quantify the "cost performance" of the current compensation operation;

[0148] Finally, the gradient descent method is started. This algorithm adjusts the power of each heating unit output by the inverse model to maximize the compensation energy efficiency index or to minimize the residual wavefront aberration under a certain energy efficiency index, and generates a set of corrected heating power increments. The increments are superimposed on the initial heating power to form the optimal heating power distribution applied to the transparent micro-heating array, until the residual wavefront aberration or the compensation energy efficiency index reaches the pre-set stable threshold.

[0149] This extension upgrades the window distortion compensation from a one-step open-loop prediction to a close-loop iterative optimization process, which brings fundamental performance improvement.

[0150] This strategy makes the compensation accuracy no longer limited by the initial accuracy of the inverse model. Even if the model has errors or the system state changes slightly, the close-loop feedback mechanism can suppress the final wavefront distortion to an order of magnitude lower than the open-loop control by measuring the residual error in real time and making corrections. More importantly, the introduction of the compensation energy efficiency index as an optimization goal avoids the huge energy cost of the system to eliminate the last little distortion, prevents the secondary thermal distortion and thermal noise caused by excessive heating, and makes the entire compensation process more stable, efficient and energy-saving. Finally, the system provides an optical observation window close to the "perfect diffraction limit" with low thermal noise.

[0151] 3.3) First, calculate the amplitude Z of the residual wavefront aberration RMS,residual ; This formula is derived from the evaluation standard of wavefront quality in optical engineering, and the root mean square value is the most commonly used indicator to represent the overall size of the aberration; Specifically, the compensated wavefront is obtained by the wavefront sensor, and it is decomposed into a set of residual Zernike coefficients; The sum of the squares of the Zernike coefficients is added, and the average is taken to obtain the RMS value of the residual wavefront, denoted as Z RMS,residual ;

[0153] Then calculate the compensation energy efficiency index η comp ; Its form is inspired by the concept of "cost-effectiveness ratio", and a function with saturation characteristics is used to ensure the stability and physical meaning of the output; Specifically, the compensation energy efficiency index η comp is defined as the ratio of "wavefront quality improvement" to "energy cost". The numerator represents the degree of distortion suppression, represented by an exponential decay function, and the smaller the residual error, the closer the value to 1. The denominator represents the energy consumption, represented by a term that grows linearly with the total power. The overall structure ensures that when the residual error is extremely small and the power is extremely large, the index will decrease, and vice versa.

[0154] The compensation energy efficiency index η comp is calculated as: Where η comp is the compensation energy efficiency index, dimensionless, and its value range is designed to be in the interval (0, 1); Z RMS,target is the target residual RMS value, which is a preset performance indicator; P total is the total heating power currently applied, which is obtained by summing P ij . P maxis the maximum total heating power allowed by the transparent micro-heating array, is a known system bundle parameter; and are the dimensionless weight coefficients of the corresponding parameters, used to adjust the sensitivity of the formula to "accuracy" and "energy consumption". In this embodiment, = 2 and = 1. The greater is, the more demanding the requirement for the residual error is; the greater is, the higher the requirement for energy saving is.

[0155] The gradient descent method based on the response matrix is used for iterative correction. First, the influence function of each heating unit needs to be calibrated offline or obtained through simulation; set H ij represents the independent heating unit in the i-th row and j-th column of the transparent micro-heating array;

[0156] Further to the heating unit H ij , the Zernike coefficient change vector generated by applying unit power, which constitutes the response matrix R of the system; and the energy efficiency index comp The value range of is in the interval (0, 1). The numerator is an exponential function in the interval (0, 1], and the denominator is a term greater than 1, so the overall value range is (0, 1). When the energy efficiency index comp tends to 1, the system is in a more efficient compensation state; the change trend of the "window optical quality" is better; and comp → 1 requires that the numerator tends to 1 and the denominator also tends to 1. The numerator tends to 1 requires that RMS,residual → 0. The denominator tends to 1 means that This requires that P total → 0. This case corresponds to the ideal state in physics. In actual applications, it represents that the compensation system suppresses the residual error Z total to a level far lower than the target value Z RMS,residual , indicating that the compensation strategy is more efficient. RMS,target

[0157] When the energy efficiency index comp tends to 0, the compensation state of the system is less efficient. The change trend of the "window optical quality" is worse; in this embodiment, "→" means tending to, and the tending interval of this embodiment is (0, 0.2); the specific value is adjusted by the expert group based on experimental data; and comp → 0 is caused by two cases: 1) the numerator tends to 0, that is, the value of Z RMS,residual is large, indicating that the compensation effect is poor and the residual error is large; and 2) the denominator tends to infinity, that is, P total approaches or reaches its upper limit P max ​, which indicates that the system consumes a huge amount of energy. Technical effect: This corresponds to two typical failure conditions. The first is that the compensation algorithm diverges or the model is seriously wrong, and the applied power increases the distortion instead. The second is to fall into the "diminishing returns" trap, where the system has eliminated 99% of the distortion, but to eliminate the last 1% of the tiny residual, it has boosted the heating power to the limit. At this time, although the residual is small, P max makes η comp drop sharply, thereby guiding the iterative algorithm to stop or find other more energy-efficient power distribution schemes. This fully demonstrates the key role of the index in preventing over-compensation and thermal runaway.

[0158] A test execution and data synchronization module is configured to trigger the optical test instrument to perform multi-dimensional optical data acquisition when the deviation of the core temperature estimate from the target core temperature and the window state parameter both satisfy a preset stability threshold, and to synchronize the collected optical data with the corresponding device state parameters and core temperature estimates to generate a single-point optical performance data set.

[0159] Further explanation: This scheme introduces a prediction model of system stability index, which can predict when the system will reach a stable state by analyzing the trend of the stability index, thereby achieving early preparation and immediate triggering, greatly shortening the invalid waiting time and significantly improving the overall test efficiency.

[0160] 4.1) The test execution and data synchronization module includes a predictive triggering unit and a data quality evaluation unit; the predictive triggering unit models the trend of the deviation of the core temperature estimate from the target core temperature and the window state parameter in the time domain to predict the future time point when the system reaches the preset stability threshold and generate a pre-trigger signal; the data quality evaluation unit evaluates the actual system stability at the triggering time and the fluctuation characteristics during the stable process when the optical test instrument is triggered, and generates a data quality index associated with the single-point optical performance data set.

[0161] 4.2 Specific implementation: First, the predictive triggering unit continuously monitors two key system state quantities: core temperature stability and window optical stability; according to the historical data of these two stability indicators, their values in the next few sampling periods are predicted in real time; the predictive triggering unit uses a Kalman filter or a time series analysis model.

[0162] When the future stability indicators output by the predictive triggering unit continuously satisfy the preset threshold, a pre-trigger signal is sent to the optical test instrument, so that the optical test instrument can complete the preheating, self-checking and other preparations in advance.

[0163] Subsequently, the system enters the final trigger judgment stage. It not only confirms whether the core temperature stability and the window optical stability at the current time are both lower than the respective preset thresholds, but also calculates the stability process fluctuation within a window period before reaching the stable state.

[0164] Only when all conditions are met, the system issues the final formal trigger instruction to start the optical test instrument for data acquisition. At the same time, the data quality assessment unit is started, which combines the actual core temperature stability, the window optical stability at the trigger time, and the aforementioned stability process fluctuation to calculate a standardized data quality index.

[0165] Finally, when the optical data acquisition is completed, the system encapsulates the collected optical data such as spectral data, optical power values, and spot images, together with the corresponding device state parameters, core temperature estimates, timestamps, and newly generated data quality indices, into a structured single-point optical performance data set with a quality label.

[0166] 4.3) The predictive trigger function significantly shortens the test period; the traditional method needs to wait for the system to stabilize and then maintain for a period of time, while the present scheme makes the instrument preparation time and the system stabilization process parallel, and the total time consumption is close to the physical shortest time required for the system to reach stability. The introduction of the data quality index gives the test data an unprecedented "credibility" dimension. In the data analysis stage, researchers can filter or weight the data according to the data quality index, for example, when fitting the device performance curve, data points with high data quality index values can have higher weights, resulting in more accurate and robust device models. This fundamentally improves the scientific nature of data analysis and the reliability of the final conclusion.

[0167] 4.4) Calculation of stability process fluctuation σ stable This formula is derived from the standard deviation calculation in statistics and is the most basic method to measure the degree of data dispersion. Here it is applied to measure the fluctuation of "stability" itself in the system stabilization process;

[0168] σ stable Formula composition principle: First, define a comprehensive stability index, which is the weighted sum of the normalized core temperature stability and the window optical stability. Then, within the N1 sampling periods before the system enters the stable state (i.e., both stability indexes are lower than the threshold), calculate the standard deviation of the sequence of this comprehensive stability index to quantify whether the system enters the stable state "smoothly" or "oscillates".

[0169] Comprehensive stability index S comp (k1), represented as where w T , w Oare the weights of the corresponding parameters, all of which are in the interval (0, 1), and w T +w O = 1; S T (k1) is the core temperature stability S T (k1) = |T j,est (k1) - T j,set (k1)|; S O (k1) is the window optical stability of the k1th control cycle, and S O (k1) = Z RMS,residual ;ò T is the core temperature stability threshold;ò O is the window optical stability threshold;

[0170] The stability process fluctuation σ stable ; specifically represented as

[0171] wherein is the average value of the N1 point comprehensive stability index; N1 is the total window size of the control cycle for calculating the fluctuation; i2 is the index mark of the control cycle;

[0172] This calculation data quality index Q DI ; specifically obtained by multiplying three decay terms, which respectively represent the quality factors contributed by the core temperature stability, the window optical stability and the stability process fluctuation. Each factor is an exponential decay function, which is in the form of exp(-x), and when the respective deviation terms S T , S O , σ stable are zero, the factor is 1 (full score), and as the deviation increases, the factor tends to 0. This multiplication structure means that any quality difference in one dimension will cause the final total quality index to tend to 0, embodying the "short board effect" and conforming to the logic of data quality.

[0173] Q DI calculation formula is represented as wherein, Q DI is the data quality index, dimensionless, with a value range of (0, 1]; S T , S O are the actual core temperature stability and the window optical stability at the triggering moment; σ target is the target fluctuation threshold, i.e. the upper limit of the fluctuation of the ideal stability process expected; λ T , λ O , λ σ are dimensionless sensitivity coefficients for adjusting the influence of each item on the final quality index; the value range is limited to (0, 1]; Q DIthe value range of Q is designed in the interval (0, 1]. Since the value range of each exponential function is (0, 1], the value range of their product is also (0, 1]; the case of the parameter being "0" or "1": Q DI can take 1 (when all deviation terms are 0), but theoretically cannot take 0, only infinitely approaches 0. The technical goal of this formula is to provide an objective, quantitative quality score for each triggered measurement data, reflecting the reliability and value of the data point;

[0174] When Q DI tends to 1, the system measurement state is better; the reliability change trend of the "single-point optical performance data set" is higher, and the data has higher reliability. When Q DI tends to 0, the system measurement state is worse; the reliability change trend of the "single-point optical performance data set" is lower, and the data has lower reliability;

[0175] Since Q DI is a multiplication structure, Q DI → 0 as long as any one of its three constituent factors tends to 0; for example, even if the core temperature stability and the window optical stability just meet the threshold, but if the stable process is just barely stopped in the threshold after a severe oscillation (i.e. σ stable exceeds the preset warning value), then the third term will tend to 0, causing the entire Q DI to also tend to 0. This reflects the advancement of the index. In the traditional method, this data point will be considered "qualified", but the present scheme can identify the hidden risks behind it; a severe stabilization process may mean that the control system is in a critical unstable state, and the measurement result may just be a transient lucky value. By giving this data point a very low Q DI score, the system correctly identifies that it is a low-quality, unreliable data point that should be down-weighted or excluded in subsequent analysis. This greatly enhances the robustness and scientificity of the overall test results.

[0176] Data validity discrimination module: used for real-time statistical analysis of the single-point optical performance data set, when the statistical characteristic value of the optical data in the single-point optical performance data set exceeds the preset control limit, it is determined that the single-point optical performance data set is invalid, and a retest instruction is generated to the test execution and data synchronization module;

[0177] Further explanation: 5.1) "determination of invalid data set" through real-time discrimination of test data validity and retest trigger execution; at a stable to-be-measured temperature point, the sampling module will quickly collect N3 optical data according to the instruction to form a sample set; the retest trigger immediately analyzes the sample set. Calculate the standard deviation σ data; and for the spot image, the centroid position is extracted and the standard deviation of the Euclidean distance of the centroid positions of N3 measurements is calculated as σ pos ; the calculated statistics are compared with the preset control limits, and if both σ data and σ pos are greater than the corresponding control limits, a retest instruction is generated to the test execution and data synchronization module.

[0178] The discrimination rule is: if (σ data > UCL σ ) or (σ pos > UCL pos ), the data is determined to be invalid. UCL σ and UCL pos are control limits preset according to the instrument accuracy and test requirements; once determined to be invalid, the system discards the N3 sets of data and immediately re-executes the N3 times acquisition process. If the retest fails continuously for M times, the system is suspended and an alarm is given, prompting the existence of random strong interference such as external vibration, instrument failure, etc. In this embodiment, M = 3;

[0179] 5.2) In another embodiment, under different test conditions "-40℃ ~ 105℃", the normal instrument noise level and data fluctuation range are different. This scheme introduces a dynamic control limit, and the system can generate a "reasonable" fluctuation range benchmark in real time according to historical data and current working conditions. This makes the effectiveness discrimination no longer dependent on rigid global thresholds, but achieves self-adaptation to specific test environments, reducing false positives and false negatives.

[0180] The data validity discrimination module further comprises a dynamic benchmark generation unit and a data reliability evaluation unit; the dynamic benchmark generation unit is used to generate a dynamic control limit of the statistical characteristic value matched with the current test working condition based on historical test data; the data reliability evaluation unit is used to extract the standard deviation and kurtosis in the single-point optical performance data sample set as statistical characteristic values, and compare the statistical characteristic values with their corresponding dynamic control limits, and through a cross-weighted fusion algorithm, the comparison results are combined into a single data reliability index R DI , and based on the index, it is determined whether the single-point optical performance data set is valid. The greater the value of the data reliability index R DI , the more reliable and effective the single-point optical performance data set is.

[0181] After receiving the single-point optical performance data set from the test execution and data synchronization module, the data validity discrimination module performs the following steps:

[0182] First, the dynamic reference generation unit in the module is started. It will query the historical database to find the historical test points closest to the current test condition (such as core temperature estimate, device input power), and extract the statistical characteristics of these historical data points, calculate the expected reference standard deviation and reference kurtosis under the current condition. These two reference values constitute a dynamic and adaptive "control limit".

[0183] Next, the module performs multi-feature statistical analysis on the current N3 measurement sample set. It not only calculates the sample standard deviation of the sample set optical power value to measure the consistency of the data; also calculates its sample kurtosis to identify whether there are extreme outliers or the data distribution form is abnormal.

[0184] Then, an innovative data reliability evaluation unit starts to work. It compares the real-time calculated sample standard deviation and sample kurtosis with the reference standard deviation and reference kurtosis generated in the previous step respectively, and obtains two deviation amounts. The unit uses a cross-weighted fusion algorithm to combine the two deviation amounts into a standardized, single data reliability index. The core idea of this algorithm is that the abnormality of one feature will exacerbate the negative impact of the abnormality of the other feature on overall reliability.

[0185] Finally, the system makes a final effectiveness discrimination according to the generated data reliability index. If the index is lower than a preset reliability threshold, it is determined that the current data set is invalid, the data set is discarded and a retest instruction is generated. If the data reliability index still does not meet the standard after continuous multiple retests, the system is suspended and a high-level alarm is issued, indicating that there may be a systematic failure.

[0186] This expansion scheme upgrades the data validity discrimination from a simple "pass / fail" check to an online quality assurance system with "dynamic reference, deep diagnosis, and intelligent fusion" capabilities. Technical effect: The introduction of dynamic reference makes the discrimination standard "go with the market", allowing greater normal fluctuations in low temperature zones and requiring more stringent consistency in high temperature stable zones, thereby maintaining a high discrimination accuracy in all conditions. At the same time, analyzing the standard deviation and kurtosis gives the system "diagnostic" capabilities: this provides a direction for troubleshooting. Finally, the data reliability index provides a smoother and more informative decision basis than simple threshold judgment, ensuring that each piece of data in the final warehouse has undergone strict, context-aware quality review, and its credibility far exceeds that of traditional methods.

[0187] Test flow adaptive scheduling module: used to calculate the "test value index" for each test temperature point based on the test value evaluation unit through the game model, and fuse the "expected information gain" based on Gaussian process regression prediction and the "time cost" based on the system thermal inertia model prediction, and dynamically select the next optimal test temperature point in the test sequence.

[0188] Further explanation: the step of this embodiment introduces the core concept of expected information gain, which will evaluate the amount of new information each "to-be-tested temperature point" in the test sequence brings to the final device model, and weigh it against the time cost required to test this point. This is a higher-dimensional intelligence that ensures that test resources are always used in the way that "reduces uncertainty" the most;

[0189] 6.1) The test value evaluation unit is used to build a Gaussian process regression model based on the collected data set for each to-be-tested temperature point in the test sequence to predict the optical performance of the to-be-tested temperature point and its uncertainty, and quantify the uncertainty as expected information gain; at the same time, the test value evaluation unit also predicts the time cost required to reach the to-be-tested temperature point based on the system thermal inertia model; finally, the test value evaluation unit fuses the expected information gain and time cost through a game model to calculate a test value index, and dynamically selects the next optimal test temperature point in the test sequence according to the test value index;

[0190] 6.2) The specific implementation content is as follows:

[0191] First, the test flow adaptive scheduling module starts the system state evaluation. Real-time acquisition and fusion of three key underlying indexes: controller adjustment strength index, compensation energy efficiency index (from window distortion active compensation module, reflecting optical stable state) and data reliability index (from data discrimination module, reflecting measurement stability). Through a cross-weighted fusion algorithm, the three indexes are combined into a single, comprehensive test readiness index, which is used to determine whether the current to-be-tested temperature point has been completed with high quality.

[0192] When the test readiness index reaches the preset threshold, confirming that the current to-be-tested temperature point is tested, the test value evaluation unit is activated. It will traverse all to-be-tested temperature points in the test sequence that have not been tested. For each to-be-tested temperature point, it uses all the effective optical performance data sets obtained to predict the optical performance mean and the width of the confidence interval of the to-be-tested temperature point through the Gaussian process regression (GPR) model. The width of this confidence interval is directly quantified as the expected information gain that can be obtained by testing the to-be-tested temperature point.

[0193] At the same time, the test value evaluation unit calls a pre-established system thermal inertia model, inputs the current temperature and the temperature of the to-be-tested temperature point, and quickly calculates the expected time cost of the high-low temperature test chamber from the current state to the to-be-tested temperature point and reaching stability.

[0194] Finally, the dynamic decision model is activated. It combines the expected information gain and the expected time cost of each untested temperature point into a normalized test value index through a preset game formula. The test value index measures the "cost performance" of testing the untested temperature point. The system will select the untested temperature point with the highest test value index as the next test target and instruct the internal test equipment of the high-low temperature test chamber to move to it. If the test value index of all untested temperature points is lower than an "exploration threshold", the system determines that the test is sufficient and can end early.

[0195] 6.3) The strategy improves the test efficiency and intelligence level; the system no longer tests blindly in a fixed order, but prioritizes testing those "key points" that can best improve the device performance model. For regions with flat performance changes, it can skip them, thereby obtaining a more informative and less uncertain data set in less time than traditional methods. In this embodiment, for a laser P-I-T characteristic test, the total test time can be shortened by 30% to 60%, while more accurately capturing the nonlinear details of the threshold current change with temperature;

[0196] The design logic is to establish a preliminary model based on existing measured data and identify the highest uncertainty region / high information gain point of the model. Then, the cost (time cost) of exploring each ambiguous point is evaluated, and the point with the highest "input-output ratio" is selected for the next experiment. The Gaussian process regression in this scheme is the "preliminary model", the expected information gain is the "ambiguity", the time cost is the "cost", and the test value index is the "input-output ratio". This is a complete and logically consistent intelligent exploration process based on the idea of Bayesian optimization;

[0197] 6.4) Quantitative implementation of step 6.2), first define the following parameters:

[0198] Let the controller adjustment intensity index be E ctrl ; it is obtained from the temperature control module; let the data reliability index be R DI ; let the test readiness index be I ready ; let the untested temperature point be T cand ; let the expected information gain be G info (T cand ); let the expected time cost be C time (T cand ); let the test value index be V test (T cand );

[0199] For the test readiness index I readyThe calculation of I is as follows: I is a comprehensive evaluation of the fusion of three underlying indices; the core idea is that the final readiness state is not simply a weighted sum of the three indices, but there is a coupling effect between them. When the data reliability is low, even if the temperature and optical path are stable, the temperature point under test does not meet the requirements. The test readiness index I ready This "veto" and "short board effect" is achieved by multiplying the three indices, and the introduction of the denominator makes it have a smoother response characteristic. ready The calculation formula is represented as:

[0200] where I ready is the test readiness index, with a value range of (0, 1); for the expected information gain G info (T cand ) calculation: Gaussian Process Regression (GPR) is adopted. GPR is a non-parametric Bayesian method that models functions and naturally gives the mean and variance of the prediction point. The derivation process is as follows:

[0201] Input: the measured effective data set {(T1, P1), (T2, P2),..., (T k , P k )}, where P i′ is the optical power at T i′ temperature; i' is the index mark in the effective data set; a GPR model is trained using these data points. The model learns a posterior distribution of a function describing the change of optical power P with temperature T; for any temperature point T cand to be tested in the test sequence, the GPR model outputs a prediction value, which is a Gaussian distribution; the standard deviation σ(T cand ) of the prediction distribution directly represents the uncertainty of the model at this point. The larger the standard deviation, the less certain the true value of the point is, and therefore the more new information the test of the point can bring. G info (T cand ) is represented as G info (T cand ) = σ(T cand ); σ(T cand ) is the standard deviation of the optical power predicted by the GPR model at the temperature point T cand to be tested;

[0202] For the test value index V testThe calculation of the index: The index is designed in the form of the ratio of "expected gain" to "total cost". The numerator is the expected information gain, representing the gain. The denominator consists of two parts: one is the inherent "1", representing the basic cost or opportunity cost; the other is a term proportional to the expected time cost. This form ensures that when the information gain is zero, the value is zero; when the time cost tends to infinity, the value also tends to zero.

[0203] Test value index V test The calculation formula is represented as Wherein,

[0204] V test (T cand ) is the test value index of the temperature point T cand to be tested, whose value is normalized to the interval (0, 1); C time (T cand ) is the predicted time required to reach and stabilize at T cand , which is predicted by the system thermal model; G norm is the normalized factor of information gain, which has the same dimension as G info (T cand ), and is used to balance the magnitude of the numerator and the denominator. Take the average information gain expected in the entire test range. γ4 is the time cost weight coefficient, which reflects the degree of attention to "time", and the larger γ4 is, the more the system tends to choose "time-saving" test points.

[0205] The value range of V test is [0, +∞), which is normalized to the interval [0, 1] by dividing by the maximum value in all candidate points, facilitating comparison; the case where the parameter is "0" or "1": V test can take 0. Specifically, when a temperature point to be tested has been measured with high precision, or it is sandwiched between two measured temperature points and the model predicts that it is very certain here, G info (T cand ) = 0, V test = 0.

[0206] When V test tends to its maximum value (or normalized 1): the system identifies that the exploration value of the current temperature point to be tested is higher; the trend of "test process efficiency" is higher, and the amount of new information per unit test time is larger; the value of V test tends to the maximum value, which means that its numerator G info (T cand ) is relatively larger, while the denominator G norm + γ4·C time (T cand) is relatively smaller. This describes an ideal "next test point": 1) the model is uncertain about this point: the more the test can reduce the unknown degree of the overall performance of the electronic component; 2) the lower the time cost required to reach this point and complete the test. The more the system needs to choose such a point for testing;

[0207] When V test is closer to 0, the system considers that the exploration value of the to-be-tested temperature point is lower; the change trend of "test process efficiency" is lower;

[0208] V test is closer to 0, there are two cases: 1) the numerator G info (T cand )→0. This means that according to the existing data, the performance of this point is more certain, and the amount of new information obtained by testing again is less. 2) the denominator C time (T cand )→∞; this means that testing this point requires a longer time, even if the information gain of the to-be-tested temperature point is acceptable, but the time cost is high. Technical effect: V test successfully prevents the system from performing two inefficient behaviors: "repeating the test of known information points" and "paying a disproportionate time cost for a little information gain". When V test of all to-be-tested temperature points tends to 0, it means that the uncertainty of the entire performance curve has been reduced to an acceptable level, and the system can judge that "the test is completed", thereby realizing intelligent early termination of the test and saving the total test time.

[0209] Embodiment two:

[0210] To verify the innovation and advantage of the control system in the test of optoelectronic components in a wide temperature range, the following experiments are carried out. First, six types of representative optoelectronic components on the market (silicon photon transceiver module A, VCSEL array B, micro-ring resonator C, optical sensor D, photodetector E, integrated optical path F) are selected as the test objects. The test device includes: a high-low temperature test chamber (adjustable from -40℃ to +105℃), a dual-target real-time sensing module, a core temperature feedforward compensation module, a window distortion active compensation module, a data validity discrimination module, and a test process adaptive scheduling module. Before the test, the temperature sensors of the test chamber, the window wavefront aberration sensors, and the optical test instruments are calibrated, and the errors of each sensing channel are compensated using a standard temperature source and a wavefront correction plate to ensure that the sensor range and response time meet ±0.01℃ and 1μs level. The test implementation is divided into the following steps:

[0211] 1. Test preparation: fix the to-be-tested components on the sample holder of the test chamber according to the specification, and ensure good thermal contact between the core of the component and the temperature control cavity;

[0212] A transparent micro-heating array is attached to the inside of the test chamber observation window and is connected to the window distortion active compensation module;

[0213] A dual-target real-time sensing module is configured to synchronously collect core temperature estimation values and window wavefront aberration information at a frequency of 1 Hz;

[0214] The test flow adaptive scheduling module is initialized in a priori information-free mode, and the initial test temperature point is set to -40 DEG C.

[0215] 2. Temperature and optical compensation iteration: Start core temperature feedforward compensation: according to the pre-established device thermal resistance model, the module calculates the deviation of the core temperature estimation value from the target temperature in real time, and issues an environmental temperature adjustment instruction to approach the target with a step of ± 0.05 DEG C;

[0216] Parallel start window distortion active compensation: based on the real-time collected wavefront aberration, the algorithm automatically generates power instructions for each unit of the transparent micro-heating array, and generates a compensation temperature gradient negatively correlated with the aberration to make the residual wavefront aberration quickly converge;

[0217] After the environmental temperature and window compensation are stable each time (the dual-target deviation is lower than the preset threshold), the optical test instrument is triggered to collect multi-dimensional light field data.

[0218] 3. Data validity and adaptive scheduling: The data validity discrimination module calculates the standard deviation and kurtosis of the single-point optical performance data set in real time, and if it exceeds the control limit, it automatically issues a retest instruction to ensure that the outlier data caused by sudden interference is removed.

[0219] Based on the expected information gain predicted by the Gaussian process regression and the temperature switching time cost predicted by the thermal inertia model, the adaptive scheduling module calculates the test value index using the game model, dynamically selects the next test temperature point, and skips the temperature section with low information gain and high switching cost, thereby effectively shortening the overall test time.

[0220] 4. Result recording and comparison: After each device completes the -40 DEG C→+105 DEG C→-40 DEG C round trip test, the key performance indicators are extracted: maximum temperature control deviation, maximum wavefront aberration residual, single test period, data misjudgment rate, core temperature feedforward compensation response time and data synchronization delay.

[0221] The data of the system are compared with those of the traditional static test flow (only relying on shell temperature feedback, without window compensation, and without adaptive scheduling). Through this embodiment, the significant advantages of the present application in temperature control accuracy, optical compensation effect, test efficiency and data reliability can be objectively shown.

[0222] Table 2 Performance research in wide-temperature-range optoelectronic component testing

[0223]

[0224]

[0225] From the table, the testing performance of the system on various types of optoelectronic devices shows high consistency and superiority:

[0226] Temperature control and optical fidelity: The maximum temperature deviation of each device is maintained within ±0.07℃ to ±0.10℃ (average ±0.085℃), which is much better than the traditional ±0.5℃ level; the residual wavefront distortion is concentrated in 1 / 23λ to 1 / 19λ (average 1 / 20λ), which is more than twice the traditional λ / 10. As can be seen, the core temperature feedforward compensation and window distortion active compensation module can effectively eliminate self-heating errors and window-induced wavefront distortion, and still maintain high temperature and optical control accuracy in a wide temperature range of -40℃ to +105℃.

[0227] Test efficiency and data reliability: The single -40℃→+105℃→-40℃ round trip test cycle is between 65min and 70min (average 67.5min), which is 32% shorter than the traditional test process of more than 100min; the data misjudgment rate is kept within 0.3% to 0.8% (average 0.55%), which is significantly lower than the traditional 3% to 8%. In addition, the feedforward compensation response time is 145ms to 160ms (average 151.7ms), and the data synchronization delay is only 17ms to 22ms (average 19.5ms), ensuring the real-time and high throughput of feedback and measurement.

[0228] Cross-device applicability and stability: From silicon photon transceiver modules, VCSEL arrays, micro ring resonators, and integrated optical circuits to six kinds of optoelectronic components, the test data changes very little (the standard deviation of each index is less than 10%), which shows that the control system can not only consider the thermal and optical characteristics of different devices, but also maintain uniform high precision and high efficiency in diversified testing scenarios.

[0229] In summary, through fine feedforward temperature compensation, real-time wavefront correction, dynamic adaptive scheduling, and online data validity discrimination, the system significantly improves the accuracy and reliability of temperature and optical testing, and shortens the test period by three times, providing strong technical support for the research and development verification and mass production testing of wide temperature range optoelectronic components.

[0230] It should be noted that all the calculation formulas in the application file use regression analysis including but not limited to machine learning algorithms to deeply analyze the collected relevant parameters, identify their natural trends and mutual relationships. Professional software such as Python's Scikit-learn library or R language is used to automatically generate mathematical models that match the data. Then, the performance of the model is objectively evaluated through methods such as cross-validation, and combined with continuous feedback and optimization to ensure that the created formula truly reflects the inherent law of the data, thereby ensuring its effectiveness and accuracy. In all the calculation formulas in the application, the parameters in each formula are processed by consistent range of dimensionless to ensure that different physical quantities are compared on the same scale; Dimensionless technology includes but is not limited to Min-Max-Normalization, Z-Score standardization;

[0231] The technical solutions of the present application can be embodied in the form of a software product, which can be stored in a computer-readable storage medium such as a floppy disk, a read-only memory (ROM), a random-access memory (RAM), a FLASH, a hard disk or an optical disk, etc., including a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the method of various embodiments of the present application.

[0232] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered a list of executable instructions for implementing logic functions, and can be specifically embodied in any computer-readable medium for use by an instruction execution system, apparatus or device, such as a computer-based system, a system including a processor or other system that can fetch and execute instructions from the instruction execution system, apparatus or device. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate or transport a program for use by or in connection with an instruction execution system, apparatus or device, or in conjunction with these instruction execution systems, apparatus or devices.

[0233] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit it, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the present application, and all should be covered in the scope of the claims of the present application.

Claims

1. A control system for optical test function high and low temperature test chamber for electronic components, characterized in that, Specifically comprising: Dual-target real-time perception module: for synchronously acquiring device state parameters reflecting the core working temperature of the electronic component under test, and window state parameters reflecting the optical transmission characteristics of the observation window of the test chamber; the window state parameters include wavefront aberration information; Core temperature feedforward compensation module: for calculating the core temperature estimate of the electronic component under test based on the device state parameters and a preset device thermal resistance model, and dynamically generating environmental temperature adjustment instructions for the high-low temperature test chamber according to the deviation between the core temperature estimate and the target core temperature; Window distortion active compensation module: for generating and executing heating power control instructions for each heating unit in the transparent micro-heating array based on the window state parameters, so as to generate a compensatory temperature gradient for offsetting the wavefront aberration; Test execution and data synchronization module: for triggering multi-dimensional optical data acquisition when the deviation between the core temperature estimate and the target core temperature and the window state parameters both meet the preset stability threshold, and synchronously associating the collected optical data with the corresponding device state parameters and core temperature estimate to generate a single-point optical performance data set; Data validity discrimination module: for performing real-time statistical analysis on the single-point optical performance data set, and determining that the single-point optical performance data set is invalid when the statistical characteristic value of the optical data in the single-point optical performance data set exceeds the preset control limit, and generating a retest instruction to the test execution and data synchronization module; Test flow adaptive scheduling module: for calculating a "test value index" for each test temperature point based on the "expected information gain" predicted by the Gaussian process regression and the "time cost" predicted by the system thermal inertia model through a game model, and dynamically selecting the next optimal test temperature point in the test sequence.

2. The control system for optical test function high-low temperature test chamber for electronic components according to claim 1, characterized in that: The dual-target real-time perception module includes a device state perception unit and a window state perception unit; The device state perception unit is used to accurately quantify the self-heating level of the electronic component under test when it is powered on, thereby obtaining the device state parameters; The window state perception unit is used to quantify the dynamic optical wavefront distortion of the test chamber observation window caused by the internal and external temperature difference, thereby obtaining the window state parameters, including Zernike coefficients; and further representing the wavefront aberration information as a comprehensive optical distortion index.

3. The control system for optical test function high-low temperature test chamber for electronic components of claim 2, wherein: The dual-target real-time perception module further includes an evaluation unit for perceiving data quality, which is used to evaluate the statistical fluctuation characteristics of the device state parameters and the window state parameters within a short time interval; The statistical fluctuation characteristics of the device state parameters and the window state parameters within a short time interval are defined as device state instability index and window state instability index, respectively; Through a cross-weighted fusion mechanism, the device state instability index and the window state instability index are fused; and finally, a perception data quality index representing the overall credibility of the current perception data is obtained.

4. The control system for optical test function high-low temperature test chamber for electronic components of claim 3, wherein: The device thermal resistance model is a dynamic thermal resistance fusion unit; the dynamic thermal resistance fusion unit is used to obtain a physical thermal resistance estimated value based on a long-term physical model and an electrical thermal resistance measured value based on instantaneous electrical characteristic measurement in parallel, and generate a confidence index of the physical model by calculating the consistency deviation between the two, and finally fuse the two thermal resistance values according to the confidence index to generate a dynamic thermal resistance value for calculating the core temperature estimated value.

5. The control system for optical test function high-low temperature test chamber for electronic components of claim 4, wherein: The core temperature feedforward compensation module further comprises an ambient temperature instruction generation unit, which adjusts the proportional, integral and differential parameters of the PID controller in real time through a fuzzy reasoning system based on the core temperature error of the core temperature estimated value and the target core temperature and the core temperature error change rate, and calculates and outputs the final ambient temperature set value applied to the high-low temperature test chamber in combination with the predictive feedforward amount based on the device thermal resistance model; The controller adjusts the intensity index, which fuses the absolute values of the current core temperature error and the core temperature error change rate through a normalized weighted norm, and maps it to the (0, 1) interval through an inverse S-shaped function.

6. The control system for optical test function high-low temperature test chamber for electronic components of claim 5, wherein: The window distortion active compensation module receives the comprehensive optical distortion index and the Zernike coefficient through the controller to construct the inverse model trained by machine learning in advance; The function of the inverse model is to input the target Zernike coefficient to be offset, and output the power distribution diagram required by the entire transparent micro-heating array, so as to obtain the initial heating power instruction; The window distortion active compensation module further comprises an iterative optimization control unit; After the inverse model generates the initial heating power instruction, the residual wavefront aberration after compensation is measured in real time through the wavefront sensor, and a compensation energy efficiency index is calculated based on the residual wavefront aberration and the total heating power currently applied; the iterative optimization control unit iteratively corrects the output of the inverse model according to the residual wavefront aberration and the compensation energy efficiency index, so as to minimize the residual wavefront aberration and maximize the compensation energy efficiency, thereby generating the optimal heating power control instruction.

7. The control system for optical test function high-low temperature test chamber for electronic components according to claim 6, characterized in that: The test execution and data synchronization module comprises a predictive trigger unit and a data quality evaluation unit; the predictive trigger unit models the deviation of the core temperature estimated value from the target core temperature and the change trend of the window state parameter in the time domain to predict the future time point when the system reaches the preset stability threshold and generate a pre-trigger signal; the data quality evaluation unit comprehensively evaluates the actual system stability at the triggering time and the fluctuation characteristics in the stable process when triggering multi-dimensional optical data acquisition, and generates a data quality index associated with the single-point optical performance data set.

8. The control system for optical test function high-low temperature test chamber for electronic components according to claim 7, characterized in that: The "data set invalidity judgment" discriminates and triggers the retest through real-time test data validity; At the test temperature point, the acquisition module will quickly acquire N3 optical data according to the instruction to form a sample set; the retest trigger immediately analyzes the sample set; The standard deviation of N3 times of optical power measurement values is calculated; meanwhile, for the light spot image, the centroid position thereof is extracted, the Euclidean distance standard deviation of the centroid position of N3 times of measurement is calculated; the calculated statistics are compared with preset control limits, and if both of the two standard deviations are greater than the corresponding control limits, a re-measurement instruction is generated to the test execution and data synchronization module.

9. The control system for optical test function high-low temperature test chamber for electronic components of claim 8, wherein: The test value evaluation unit is configured to, for each to-be-tested temperature point in the test sequence, construct a Gaussian process regression model based on the collected data set to predict the optical performance of the to-be-tested temperature point and an uncertainty thereof, and quantify the uncertainty as an expected information gain; meanwhile, the test value evaluation unit is also configured to predict a time cost required to reach the to-be-tested temperature point based on a system thermal inertia model; finally, the test value evaluation unit is configured to fuse the expected information gain and the time cost through a game model to calculate a test value index, and dynamically select a next optimal test temperature point in the test sequence according to the test value index.

Images