Probe calibration method and device and probe control equipment

Abstract

The invention relates to the technical field of probe calibration, and discloses a probe calibration method and device and probe control equipment, and the method comprises the steps: calculating a vibration wavelength, and constructing a multi-probe measurement structure; synchronously acquiring a wafer surface vibration signal by using a multi-probe measurement structure, and calculating probe vibration characteristic data; acquiring a multi-measurement path electrical signal of the multi-probe measurement structure, and processing the multi-measurement path electrical signal through a baseline automatic encoder to obtain an environment temperature characteristic; calculating a signal deviation value between the measured signal and the baseline signal, and correcting the signal deviation value to obtain an electrical signal after temperature compensation; layered calibration is carried out on the probe vibration characteristic data and the electrical signal after temperature compensation to obtain a wafer electrical characteristic measurement result, so that a more comprehensive and accurate temperature compensation effect is realized, the calibration weight of each layer can be dynamically adjusted according to the probe vibration state and the temperature change degree, the calibration effect is optimized, and the measurement accuracy is improved. And the calibration method has higher environmental adaptability and robustness.

Description

Technical Field

[0001] The present application relates to the technical field of probe calibration, and particularly to a probe calibration method, device, and probe control equipment. Background Art

[0002] In the traditional wafer probe station during the testing process, due to the high-frequency vibration problem of the probes themselves, the contact between the probes and the wafer surface is often unstable, thus affecting the accuracy of electrical characteristic measurement. Especially when there is a significant mismatch between the high vibration frequency of the probes and the actually achievable low sampling frequency, the undersampling phenomenon makes it extremely difficult to accurately extract vibration parameters. This technical bottleneck severely restricts the reliability and stability of wafer testing.

[0003] At the same time, the temperature change in the operating environment of the probe station is also an important factor affecting the measurement accuracy. In the actual testing environment, temperature fluctuations will cause changes in the physical properties of the measurement circuit and probe materials, thereby causing an offset in the electrical signal characteristics. If these temperature effects are not effectively compensated, it will lead to systematic errors in the electrical characteristic measurement results. Especially during long-term testing, the accumulation of temperature gradient effects will cause the measurement error to gradually increase over time. Summary of the Invention

[0004] The present application provides a probe calibration method, device, and probe control equipment, thereby achieving a more comprehensive and accurate temperature compensation effect. It can dynamically adjust the weights of each layer of calibration according to the probe vibration state and the degree of temperature change, optimize the calibration effect, and make the calibration method have stronger environmental adaptability and robustness.

[0005] In the first aspect of the present application, a probe calibration method is provided. The probe calibration method includes: Calculating the vibration wavelength according to a preset vibration frequency and sampling frequency and constructing a multi-probe measurement structure; Using the multi-probe measurement structure to synchronously collect the vibration signals on the wafer surface, calculating the covariance values between each pair of probes, and obtaining the probe vibration characteristic data; Collecting the multi-measurement path electrical signals of the multi-probe measurement structure under different temperature conditions and processing them through a baseline autoencoder to obtain the environmental temperature characteristics; Calculating the signal deviation value between the measured signal and the baseline signal according to the environmental temperature characteristics and correcting the signal deviation value to obtain the temperature-compensated electrical signal; Performing hierarchical calibration on the probe vibration characteristic data and the temperature-compensated electrical signal to obtain the wafer electrical characteristic measurement result.

[0006] In the second aspect of the present application, a probe calibration device is provided. The probe calibration device includes: A building block for calculating a vibration wavelength according to a preset vibration frequency and a sampling frequency and constructing a multi-probe measurement structure; A calculation module for synchronously collecting vibration signals on the surface of a wafer by using the multi-probe measurement structure, calculating covariance values between each pair of probes, and obtaining probe vibration characteristic data; A processing module for collecting multi-measurement path electrical signals of the multi-probe measurement structure under different temperature conditions and processing them through a baseline autoencoder to obtain environmental temperature characteristics; A correction module for calculating a signal deviation value between a measured signal and a baseline signal according to the environmental temperature characteristics and correcting the signal deviation value to obtain temperature-compensated electrical signals; A hierarchical calibration module for hierarchically calibrating the probe vibration characteristic data and the temperature-compensated electrical signals to obtain a measurement result of the electrical characteristics of the wafer.

[0007] The third aspect of the present application provides a probe control device for executing the above-mentioned probe calibration method.

[0008] Compared with the prior art, the probe calibration method proposed by the present invention physically ensures the measurability of continuous covariance samples by constructing a multi-probe measurement structure with a general covariance layout, fundamentally solving the under-sampling problem caused by the mismatch between the high-frequency vibration of the probe and the low sampling frequency, and significantly improving the accuracy of vibration parameter identification. The baseline autoencoder calibration technology of the present invention does not require any biased samples for training, can adaptively extract environmental temperature characteristics from measurement data, realize real-time monitoring and compensation of temperature changes, and greatly improve the stability of electrical characteristic measurement in a variable-temperature environment. The present invention innovatively designs a probability compensation algorithm based on delay, which not only considers the influence of temperature changes on the signal amplitude, but also captures the influence of the temperature change rate on the signal phase through a time-delay model, achieving a more comprehensive and accurate temperature compensation effect. The present invention adopts a hierarchical calibration strategy, dividing the calibration process into three levels: the physical layer, the signal layer, and the parameter layer, and accurately calibrating the contact instability caused by probe vibration, the signal offset caused by temperature changes, and the final electrical characteristic parameters respectively, avoiding error accumulation and mutual interference. The present invention introduces an adaptive weight mechanism, which can dynamically adjust the weights of each level of calibration according to the probe vibration state and the degree of temperature change, optimize the calibration effect, and make the calibration method have stronger environmental adaptability and robustness. Description of the Drawings

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0010] The structures, proportions, sizes, etc. shown in the drawings of this specification are only used to cooperate with the content disclosed in the specification for those who are familiar with this technology to understand and read, and are not used to limit the conditions under which the present invention can be implemented. Therefore, they do not have a substantial technical meaning. Any modification of the structure, change in the proportional relationship, or adjustment of the size, without affecting the effects that the present invention can produce and the purposes that can be achieved, should still fall within the scope that can be covered by the technical content disclosed by the present invention.

[0011] Figure 1 is a schematic flowchart of the probe calibration method provided by an embodiment of the present invention; Figure 2 is a schematic block diagram of the structure of the probe calibration device provided by an embodiment of the present invention. Detailed implementation manners

[0012] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0013] The flowcharts shown in the drawings are only illustrative examples, and do not necessarily include all the content and operations / steps, nor do they necessarily need to be executed in the described order. For example, some operations / steps can be decomposed, combined, or partially merged. Therefore, the actual execution order may be changed according to the actual situation.

[0014] It should also be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit this application. As used in this application specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an", and "the" are intended to include the plural forms.

[0015] It should be further understood that the term "and / or" used in this application specification and the appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations. Please refer to Figure 1 , an embodiment of the probe calibration method in the embodiments of this application includes: Step 100: Calculate the vibration wavelength according to the preset vibration frequency and sampling frequency, and construct a multi-probe measurement structure; It can be understood that the execution subject of this application can be a probe calibration device, or a terminal or a server. Specifically, it is not limited here. In the embodiments of this application, the server is taken as the execution subject for illustration.

[0016] Specifically, taking the typical vibration frequency and actual sampling frequency of the wafer under test as input parameters, through the relationship between the vibration wave propagation speed and frequency, the vibration wavelength λ is calculated using the formula λ = V / f, where V is the propagation speed of the vibration wave in the wafer material and f is the known preset vibration frequency. Based on the vibration wavelength λ, the spatial sampling point spacing parameters d1 and d2 of the probe arrangement are set in combination with the covariance sampling theory. Among them, d1 is taken as λ / 4 to obtain the lateral spacing corresponding to one-fourth of the vibration period, and d2 is selected as λ / 4×(1 + 2k), where k is a non-negative integer, to ensure that the longitudinal spacing has the ability to sample with a 90-degree phase difference in space, forming a probe spacing design scheme. Under the guidance of this design scheme, based on a high-precision nanoscale probe moving platform, the probe stage is positioned and arranged. The first probe is installed at the coordinate origin (0, 0), the second probe is installed at the position with a lateral spacing of d1, that is, (d1, 0), the third probe is vertically placed at the position with a longitudinal coordinate of d2, that is, (0, d2), and the fourth probe is arranged at (d1, d2), forming a four-probe basic layout structure with a rectangular boundary. This four-point arrangement ensures the extraction of orthogonal covariance samples in four directions for constructing the spatial covariance matrix. To improve the covariance sample density and enhance the adaptability to the spatial structure, the four-probe structure is symmetrically extended by adding probes at the four key midpoint positions of (d1 / 2, d2 / 2), (d1 / 2, 0), (0, d2 / 2), and (d1, d2 / 2) in the original coordinate structure to construct an eight-probe layout structure with higher spatial resolution. To adapt to the diversity of the actual shapes of the wafer test areas, including rectangles, circles, or irregular areas, based on the initial eight-probe layout, according to factors such as the wafer boundary size, effective contact area, and path coverage range, the overall layout of the spatial positions of each probe is optimized and fine coordinate adjustments are made to form a non-uniformly distributed multi-probe measurement structure that adapts to the characteristics of the target area on the wafer probe stage.

[0017] Step 200: Use the multi-probe measurement structure to synchronously collect the vibration signals on the wafer surface, calculate the covariance values between each pair of probes, and obtain the probe vibration characteristic data; Specifically, a fixed contact pressure set by the force control system is applied to each probe to ensure that the physical states of all probes in contact with the wafer surface are consistent, excluding signal interference introduced by local pressure differences. The probe measurement system is controlled by a synchronous triggering mechanism, enabling eight probes to simultaneously collect high-frequency micro-vibration displacement signals on the wafer surface at a set unified sampling frequency. The vibration signals of the wafer surface for each probe are sorted out and preprocessed. Band-pass filtering is performed on the original sampling data corresponding to each probe to filter out low-frequency thermal noise and high-frequency random interference, and then a normalization operation is carried out. By removing the DC offset of each signal and standardizing its amplitude range, a standardized vibration signal set on a unified scale in a statistical sense is obtained. According to the pre-designed eight-probe spatial layout, four pairs of probe signals with typical phase difference characteristics are selected, namely (0,0) and (d1,0), (0,0) and (0,d2), (d1,0) and (d1,d2), and (0,d2) and (d1,d2). The vibration signals at these probe positions are extracted from the standardized signal set as input pairs, and the covariance is calculated for these four groups of signal pairs respectively. Since the selected probe pairs correspond to typical phase differences of 0°, 90°, 180°, and 270° in spatial distribution, the covariance results can cover four key points within one cycle. The calculation of covariance adopts an expected form to accurately reflect the joint change trend of each signal pair in the sense of time average. These four covariance values are constructed into a continuous time series in chronological order, and multiple overlapping sample intervals are formed on this time series through a sliding window processing technique to capture the dynamic change trend and constitute a covariance sample sequence. Each sliding window corresponds to a local covariance snapshot within a time segment, thus expanding the covariance at a single moment into an analyzable time-series change structure. To improve the stability of the covariance sample sequence and the frequency-domain analysis effect, detrending processing is performed on it to eliminate the slow-varying baseline term, and smoothing algorithms such as Savitzky-Golay filtering are combined to suppress high-frequency jitter noise, forming a stable covariance sample set. The Fourier transform is used to map the covariance time series to the frequency-domain space, and the dominant frequency component of the signal is extracted to initially obtain the main vibration frequency of the probe. The iterative least-squares fitting technique is used to globally fit the model A·cos(2πft + φ) = C(t) containing amplitude, frequency, and initial phase, and iterated repeatedly until the sum of the squared residuals is minimized, obtaining a probe vibration characteristic data set including key parameters such as the main frequency f, amplitude A, and phase φ.

[0018] The covariance sample set is regarded as a dynamic signal sequence with time series characteristics, and discrete Fourier transform processing is performed on this signal sequence to convert the covariance change in the time domain to the frequency domain representation, obtaining the preliminary characteristics of the energy distribution in the signal spectrum. By analyzing the frequency components with higher amplitudes in the spectrum, the interval where the dominant vibration frequency appears is effectively extracted, and based on this, a preliminary vibration frequency range is constructed. This frequency range includes a main peak and its adjacent sidebands. Based on the preliminary frequency range, a high-precision search interval is set, and dense frequency point scanning is performed within this interval. For example, the entire frequency axis is subdivided at intervals of 0.1 kHz or even smaller to construct a target frequency candidate set containing multiple high-probability candidate frequency points. The target frequency candidate set is introduced into the mathematical modeling process to construct a vibration parameter solution model. This model is based on the cosine function structure. By mapping the frequency points to the cosine space, that is, establishing a fitting residual function between the cosine waveform and the covariance sequence at each candidate frequency, a multi-parameter solution system to be optimized is formed. After the vibration model is constructed, the iterative least squares method is applied to fit and solve it. The optimal starting frequency point within the candidate frequency set range is selected as the initial value, and the initial amplitude and initial phase angle values are set based on this frequency. At the same time, the sum of the squared residuals between the current model output and the actual covariance sample is calculated using the least squares principle to obtain the first-round fitting output results, including the preliminarily estimated vibration frequency, amplitude, and phase parameters. Based on the current fitting results, the residuals between the model prediction values and the true covariance data are calculated, and a loss function is constructed with the sum of the squared residuals as the cost function to perform iterative optimization processing on the vibration parameters using the gradient descent method. In each iteration process, the parameter values are corrected according to the partial derivative information of the cost function with respect to frequency, amplitude, and phase, and the model output is updated in real time until the change rate of the sum of the squared residuals between two consecutive iterations is lower than the set threshold (such as 0.001), indicating that the model has converged to the optimal solution. The output vibration parameter set constitutes the probe vibration characteristic data, including the high-precision dominant frequency, equivalent amplitude, and initial phase.

[0019] Step 300: Collect the electrical signals of multiple measurement paths of the multi-probe measurement structure under different temperature conditions, and process them through a baseline autoencoder to obtain the environmental temperature characteristics; It should be noted that multiple temperature test points are preset on the wafer probe station, and these test points cover the target temperature range. The local temperature environment of each test point is adjusted by a constant temperature control device or a hot air heating module to ensure that each round of testing in the subsequent acquisition process is carried out under clear and stable temperature conditions. In this temperature control environment, the multi-path sampling of electrical signals is implemented on the wafer under test using a pre-built multi-probe measurement structure. E independent measurement paths are selected at each temperature test point. Each path includes an electrical signal source, a signal injection point that contacts the wafer through a probe, and a signal receiver connected to the back-end acquisition system to form a closed test channel to capture the non-consistent response behavior of electrical characteristics between different paths under temperature changes. After obtaining multiple groups of measurement path electrical signals, in order to ensure that they have a consistent input format in the neural network model, all path signals are uniformly preprocessed, including normalizing the amplitude to eliminate the numerical offset caused by the path gain difference, and performing time domain alignment operations at the same time. Through time interpolation or synchronization point reconstruction, the offsets of different paths in the physical sampling timing are unified to the same time scale system, and a multi-measurement path signal data set with unified structure, consistent amplitude, and time alignment is constructed. The preprocessed path signal is input into the baseline autoencoder model for environmental feature extraction. The path fusion module in the baseline autoencoder structure uses an attention weight mechanism or an adaptive weighting algorithm to fuse all E paths, calculate the relative importance weights of each path, and linearly combine them according to the weights to output a fusion vector representing the joint characteristics of all measured paths. The fused feature input vector is passed to the first hidden layer through the input layer of the baseline autoencoder for nonlinear transformation. The transformation acts on the input data by setting the weight matrix and the bias vector to map it to a high-dimensional feature space to explore the deep relationship between the electrical signals of each path and reveal its implicit trend driven by temperature changes. The intermediate result output by the first hidden layer is the first layer feature map, which is further passed to the second hidden layer for deeper activation processing. In this layer, ReLU (rectifier linear unit) is used as the activation function to compress negative values to zero and retain only the positive activation area, thereby enhancing the sparsity and robustness of the feature representation, and outputting a set of implicit representation vectors with clear temperature distinction capabilities to obtain the environmental temperature characteristics.

[0020] Step 400, calculating the signal deviation value between the measured signal and the baseline signal according to the ambient temperature characteristics, and correcting the signal deviation value to obtain a temperature-compensated electrical signal; Specifically, through the decoder part of the baseline autoencoder, the baseline signals of each measurement path are reconstructed according to the environmental temperature characteristics. The decoder restores the multi-measurement path baseline signals corresponding to the current thermal environment by performing a non-linear inverse mapping on the environmental temperature characteristics. This reconstruction process is regarded as an approximate prediction of the electrical response in the ideal state without temperature perturbation by the model. The output reconstructed baseline signals represent the theoretical unbiased response references of each path at different temperatures. On this basis, the corresponding difference calculation is performed between the currently actually collected electrical signals of each measurement path and the reconstructed baseline signals to obtain the signal offset caused by temperature change under each path, forming a set of signal deviation values, which reflects the amplitude drift and phase shift effects of temperature perturbation on the electrical signal. Time series analysis processing is performed on the above signal deviation values. By comparing the change trends of the deviation values of each path in a continuous time period, the delay relationship of the signal response between the measurement paths is established, and a time-delay model is constructed. This model depicts the asynchrony between the propagation and response of temperature change on different paths, and the signal time difference characteristics caused by temperature conduction are obtained by fitting the offset delay between each path. In order to improve the accuracy of the time-delay modeling, the cubic spline interpolation method is introduced to perform smooth fitting on the non-linear change, so that the model has good prediction ability under different temperature gradient changes. Based on this time-delay model, time alignment and phase correction are performed on the original signal. The deviation values of each path are time-shifted according to the corresponding delay to obtain a preliminary corrected signal, eliminating the system-level non-uniformity caused by the difference in response speed. Based on the preliminary corrected signal, the posterior probability distribution of temperature under the condition of observing the signal deviation is calculated. According to the posterior probability density distribution of temperature corresponding to each measurement point under the current signal deviation, the confidence weight coefficient of each path in this state is calculated. This process calculates the most likely temperature influence degree in the current state by taking the known signal deviation value as the observation input and combining the prior distribution and temperature response mapping constructed during the training of the baseline autoencoder model, and thus determines the weight value of each path. Based on the probability weight, the electrical characteristic values of each measurement point are weighted and corrected, and the temperature influence degrees suffered by different paths are incorporated into the final compensation model in a weighted form to form the output of the temperature-compensated electrical signal.

[0021] Cross-correlation analysis is performed on the signal deviation values of each measurement path to establish the time-domain correlation structure of the dynamic response between paths. Cross-correlation analysis is performed on the deviation signals between each pair of measurement paths to identify the degree of signal similarity between the two paths at different time delays, forming a delay relationship matrix that includes the cross-correlation functions of all path pairs. Each element in this matrix represents the correlation index at different delay times between a given pair of paths. Through this matrix, the relative timing of the path response is evaluated, and the non-uniformity and hysteresis during the propagation of temperature perturbations in each path are quantified. The maximum value of the cross-correlation function between each pair of paths in the delay relationship matrix is detected to determine the time delay point corresponding to the maximum cross-correlation coefficient. This delay point represents the time offset at which the signals in that pair of paths reach the maximum similarity. The maximum correlation delays of all path pairs are systematically extracted and statistically summarized to obtain a set of original time delay datasets covering all path combinations. Numerically, this dataset reflects the different degrees of response time differences induced by temperature in the entire multi-path system. The original time delay dataset is input into a hierarchical clustering analysis model. Based on similarity metrics such as Euclidean distance or dynamic time warping, the time delay values are divided into multiple hierarchical groups with consistent behavioral characteristics according to their response patterns, forming a top-down hierarchical time delay data structure in the clustering tree structure. This structure reflects the hierarchical propagation characteristics between different paths under temperature perturbations, facilitating separate modeling for different dynamic response intervals. Representative control points are selected in each layer, i.e., time delay samples at typical or boundary positions in that clustering layer, and these control points are used as fitting nodes. The coefficients of each segment of the cubic spline interpolation function are determined using the least squares method to construct a time delay characteristic curve equation that continuously expresses the response trend of each layer. Boundary constraint processing is performed on the time delay characteristic curve equation, i.e., derivative or curvature limits are set at both ends of the function to ensure its smooth continuity within the overall domain, resulting in a smooth time delay model function. A mapping relationship is established between the time delay model function and the temperature change rate, i.e., by statistically analyzing the relationship between different temperature change rates (such as the temperature rise rate per unit time) and the corresponding model outputs, a time delay model is obtained.

[0022] Step 500: Perform hierarchical calibration on the probe vibration characteristic data and the electrically compensated signal after temperature compensation to obtain the measurement results of the wafer electrical characteristics.

[0023] Specifically, the vibration frequency and amplitude are extracted from the probe vibration characteristic data to reflect the mechanical perturbation characteristics during the contact process between the probe and the wafer. In the physical layer calibration stage, by analyzing the contact changes caused by vibration, the electrical contact instability caused by minute jitters is compensated, thereby correcting the resistance offset phenomenon caused by vibration to obtain the calibration result of the first layer. The physical layer calibration result is fused with the temperature-compensated electrical signal. During the fusion process, numerical errors are superimposed and corrected, and through the time consistency processing of the signal, the phase misalignment or signal drift problems caused by the superposition of temperature changes and mechanical perturbations are solved to form the calibration result of the second layer. According to the signal layer calibration result and the probe position parameters, the mapping relationship between the measurement signal and the spatial coordinates is constructed. This relationship takes into account the spatial non-uniformity effects caused by probe arrangement, detection area differences, and wiring layout. Through spatial correlation modeling, the original signal is transformed into an electrical characteristic expression in a unified coordinate system. An adaptive weight mechanism is introduced to dynamically evaluate the importance of the physical layer, signal layer, and parameter layer. This mechanism automatically adjusts the influence degree of each layer's calibration result in the final output according to the changes in the dominant interference factors in the current test environment, such as vibration amplitude, temperature fluctuation, or probe position density. The dynamic weighting based on scenario driving enhances the flexible adaptation ability of the model and also avoids the error amplification effect when a certain type of interference is too strong. By weighted fusion of the calibration results of the three layers, a calibrated electrical characteristic dataset is generated. Consistency verification and outlier detection are performed on the calibrated electrical characteristic data. Consistency verification is used to confirm whether the data meets the physical constraints of wafer manufacturing in terms of statistical distribution, while outlier detection is used to remove mutation points, drift points, or noise anomalies. The wafer electrical characteristic measurement result is obtained.

[0024] In the embodiments of the present application, the probe calibration method proposed by the present invention constructs a multi-probe measurement structure with a general covariance layout, physically ensuring the measurability of continuous covariance samples, fundamentally solving the under-sampling problem caused by the mismatch between the high-frequency vibration of the probe and the low sampling frequency, and significantly improving the accuracy of vibration parameter identification. The baseline auto-encoder calibration technology of the present invention does not require any biased samples for training, can adaptively extract environmental temperature features from measurement data, realize real-time monitoring and compensation of temperature changes, and greatly improve the stability of electrical characteristic measurements in a variable temperature environment. The present invention innovatively designs a delay-based probability compensation algorithm, which not only considers the influence of temperature changes on the signal amplitude, but also captures the influence of the temperature change rate on the signal phase through a time-delay model, achieving a more comprehensive and accurate temperature compensation effect. The present invention adopts a hierarchical calibration strategy, dividing the calibration process into three levels: the physical layer, the signal layer, and the parameter layer, and precisely calibrating the contact instability caused by probe vibration, the signal offset caused by temperature changes, and the final electrical characteristic parameters respectively, avoiding error accumulation and mutual interference. The present invention introduces an adaptive weight mechanism, which can dynamically adjust the weights of each layer calibration according to the probe vibration state and the degree of temperature change, optimize the calibration effect, and make the calibration method have stronger environmental adaptability and robustness.

[0025] In a specific embodiment, the process of executing step 100 may specifically include the following steps: Calculate the vibration wavelength λ according to the preset vibration frequency and sampling frequency, and determine the probe spacing parameters d1 and d2 based on the vibration wavelength λ to obtain a probe spacing design scheme; Perform probe positioning on the wafer probe table according to the probe spacing design scheme, place the first probe at the coordinate (0,0), the second probe at the coordinate (d1,0), the third probe at the coordinate (0,d2), and the fourth probe at the coordinate (d1,d2) to obtain a four-probe layout structure; Expand the four-probe layout structure, and set four probes at the coordinates (d1 / 2,d2 / 2), (d1 / 2,0), (0,d2 / 2), and (d1,d2 / 2) to obtain an eight-probe layout structure; Place each probe in the eight-probe layout structure according to the shape characteristics of the wafer test area to obtain a multi-probe measurement structure.

[0026] Specifically, according to the preset vibration frequency and sampling frequency, the key physical parameter for spatial structure design, namely the vibration wavelength λ, is derived. In the process of probe signal sampling, the vibration wavelength represents the spatial length corresponding to a complete cycle of the vibration signal propagating on the surface of the wafer material. Under actual test conditions, since the vibration frequency is much higher than the system sampling frequency, the vibration characteristics are analyzed by a layout method based on spatial covariance, thereby avoiding the undersampling problem and effectively restoring the vibration frequency structure at a low sampling frequency, thereby improving the response capability and stability of the entire measurement system to high-frequency dynamic disturbances. The probe spacing is proportionally designed based on the vibration wavelength, and the key parameters d1 and d2 in two spatial directions are constructed. d1 is used to describe the arrangement spacing of the probes in the x-axis direction, while d2 is used to describe the arrangement spacing in the y-axis direction. In design, d1 is taken as one-fourth of the vibration wavelength, and d2 is set to introduce a differential spacing length on this basis, forming a distribution unequal to d1, so as to ensure that the phase difference distribution between probes in different directions is not repeated, so that the probe array has sufficient spatial resolution and phase feature separation capabilities. According to the designed spacing scheme, the actual probe positioning operation is carried out on the wafer probe station. Set the first probe at the origin, that is, the coordinate is (0,0), as the reference point of the entire array structure. This position is selected near the center or symmetry axis of the wafer test area so that the subsequent probe arrangement can be symmetrically expanded. In the x-axis direction to the right of the origin, the second probe is arranged at the set spacing d1, and its position is (d1,0) to obtain the vibration covariance change in the lateral direction. In the y-axis direction above the origin, the third probe is arranged at the set spacing d2, and its position is (0,d2) to obtain the covariance signal in the longitudinal direction. The fourth probe is arranged at the intersection of the lateral and longitudinal offsets, that is, the coordinate position (d1,d2), thereby completing a rectangular four-probe layout structure covering four quadrants. Expand the four-probe layout structure. During the expansion process, four new probe positions are added to the original probe structure, which are set at the horizontal midpoint, longitudinal midpoint, center point and distal longitudinal midpoint, including the four coordinate points (d1 / 2,0), (0,d2 / 2), (d1 / 2,d2 / 2) and (d1,d2 / 2). The introduction of the extended probe enables the formation of symmetrical encrypted points in the original structure, enhances the covariance sample density in the spatial direction, and improves the ability to capture vibration change trends. Through the setting of these supplementary points, the original four-point structure is expanded to an eight-probe structure. The probe layout structure is adaptively adjusted in combination with the shape characteristics of the wafer test area. During the adjustment process, the relative coordinate relationship between the probes is kept unchanged. Through overall translation, angle rotation, center outward movement or local retraction, the actual layout position of each probe is within the accessible test area and avoids the protective layer or blank boundary on the wafer that is not allowed to be touched.For the probes near the wafer edge, appropriately contract them to the inner ring area to ensure that they neither deviate from the designed structure nor can stably collect signal data. In the case of the presence of a missing corner, an irregular edge, or a blind area in some regions on the wafer surface, adjust the actual placement strategy of some probes by means of local replacement or point displacement. For example, shift some of the probes near the boundary to the densely sampled area in the middle of the test area, or relocate them to the symmetric area on the premise of keeping the placement logic unchanged to ensure the balance of covariance data and the integrity of signal coverage. When necessary, locally contract the overall eight-point structure or redefine the direction of the symmetry axis to form an irregular probe structure such as a semi-circular, fan-shaped, or eccentric symmetric type. Through the above steps, a multi-probe measurement structure is finally obtained.

[0027] In a specific embodiment, the process of performing step 200 may specifically include the following steps: Apply a fixed contact pressure to each probe in the multi-probe measurement structure and set the sampling frequency to synchronously collect the vibration signals of the wafer surface of each probe; Sort out the vibration signals of the wafer surface of each probe to obtain the original vibration sampling data, and preprocess the original vibration sampling data to obtain the standardized vibration signals; According to the eight-probe layout structure in the multi-probe measurement structure, select four pairs of probe positions: (0,0) and (d1,0), (0,0) and (0,d2), (d1,0) and (d1,d2), and (0,d2) and (d1,d2), and extract the corresponding standardized vibration signals to obtain four signal pairs; Calculate the covariance for each of the four signal pairs to obtain four covariance values with different phase differences; Construct a time series based on the four covariance values with different phase differences and perform a sliding window process on this time series to obtain a covariance sample series; Perform detrending processing and smoothing filtering on the covariance sample series to obtain a covariance sample set; Perform Fourier transform and iterative least squares fitting on the covariance sample set to obtain the probe vibration characteristic data.

[0028] Specifically, a fixed contact pressure is applied to each probe in the multi-probe measurement structure, so that all probes are on the same force reference during the working state, avoiding the contact resistance fluctuation or probe suspension phenomenon caused by pressure differences. The realization of the fixed contact pressure is completed by a precise micro-mechanical loading mechanism or a constant-force elastic element, and it is ensured that no additional stress concentration or local deformation is introduced during the contact process. After the contact state is stable, all probes are synchronously triggered according to the set unified sampling frequency. Through the time synchronization mechanism, it is ensured that each probe collects the wafer surface response signal on exactly the same time axis, forming a multi-channel vibration data sequence with a spatial distribution structure and high time alignment accuracy. After the sampling is completed, the original vibration signal is sorted and processed, including data channel number verification, timing integrity check, abnormal data elimination, and dropped frame repair, to obtain the original vibration sampling data. The original vibration sampling data is preprocessed. The amplitude normalization operation is performed to make the signal amplitudes of all probe channels in a unified numerical range, avoiding signal energy deviation caused by inconsistent channel gains; the DC offset correction is performed on each channel to eliminate the baseline change caused by the initial voltage drift of the system. At the same time, in cooperation with the band-pass filtering operation, the low-frequency noise and high-frequency interference components are filtered out, so that the processed signal has a clear frequency band characteristic and a good periodic distribution. The standardized vibration signal is obtained through the above processing. According to the eight-probe layout scheme in the multi-probe measurement structure, typical probe pairs are selected to construct covariance signal pairs. Four probe pairs with representative spatial positions and distributed at the edges of the structure are selected, namely (0,0) and (d1,0), (0,0) and (0,d2), (d1,0) and (d1,d2), and (0,d2) and (d1,d2). These four probe pairs respectively correspond to the horizontal starting point and the horizontal boundary point, the vertical starting point and the vertical boundary point, the horizontal end point and the vertical extension point, and the vertical end point and the horizontal extension point. They form four boundary sampling paths within the rectangular area, respectively representing the spatial propagation behaviors of the vibration wave in the horizontal, vertical, and diagonal directions. The vibration signals of these corresponding probes are extracted from the standardized data and combined in pairs to form four signal pairs with clear spatial directions. The covariance is calculated for each signal pair respectively to measure its synchronism and joint change trend within the local time window. As a measure of the average deviation degree between two groups of signals, the covariance effectively captures the phase difference characteristics of spatially adjacent probes in the vibration state. Since there is a fixed relative position between each pair of probes in the layout, the corresponding covariance values show a stable phase relationship difference, thus forming four covariance reference samples in different spatial directions. These four covariance values are constructed into a continuous time series in time order to reveal the change pattern and frequency distribution of the covariance over the entire period. In order to improve the observation resolution of the covariance change trend, a sliding window process is performed on the covariance time series.Perform point-by-point sliding with a time window of fixed length over the entire time series. Extract local covariance segments within each window and construct a covariance sample sequence. This operation forms a large number of subsequence samples with overlapping structures, enhancing the statistical stability of the signal and the ability to capture fluctuation trends. Perform detrending and smoothing filtering on the covariance sample sequence. The detrending process fits the long-term trend line in the sample and removes it, making the sample sequence more prominent in medium- and short-term variation characteristics and enhancing the perception ability of high-frequency main oscillation components. The filtering operation eliminates short-term sharp fluctuations or occasional interferences by introducing a smoothing function, such as a moving average or a Gaussian kernel function, to form a covariance sample set. Perform Fourier transform on the covariance sample set to obtain the energy distribution map of covariance changes in the frequency domain. Identify the energy peak frequency range corresponding to the main vibration frequency through Fourier transform to provide an initial frequency reference for subsequent feature fitting. On this basis, combine the main peak information in the frequency domain and perform curve fitting on the covariance samples by the iterative least squares method. Set the initial values of the vibration frequency, response amplitude, and initial phase during the fitting process and gradually iterate and adjust these parameters to make its output waveform as close as possible to the covariance sample with the minimum error. After multiple rounds of convergent iteration, obtain a set of stable fitting parameters, which are the probe vibration characteristic data, including the vibration frequency, response amplitude, and phase angle.

[0029] In a specific embodiment, the process of performing Fourier transform and iterative least squares fitting on the covariance sample set to obtain the probe vibration characteristic data may specifically include the following steps: Perform Fourier transform on the covariance sample set to obtain a preliminary vibration frequency range; Determine a search interval based on the preliminary vibration frequency range, scan the frequency points within this search interval, and obtain a target frequency candidate set; Map the target frequency candidate set to the cosine space to obtain a vibration parameter solution model; Apply the iterative least squares method to the vibration parameter solution model, initialize the frequency, amplitude, and phase parameters, perform the first round of parameter fitting, and obtain preliminary vibration characteristic parameters; Calculate the residuals based on the preliminary vibration characteristic parameters, and iteratively optimize the vibration frequency, amplitude, and phase parameters by the gradient descent method until the change rate of the sum of squared residuals is less than a preset threshold to obtain the probe vibration characteristic data.

[0030] Specifically, perform a Fourier transform on the covariance sample set to convert the original time-domain signal to the frequency domain, obtaining a set of spectral structures, where each frequency point corresponds to the energy distribution intensity of a certain periodic component in the covariance signal. By analyzing the spectral results, several significant peaks with concentrated energy are observed, and these spectral peaks correspond to the main frequency or harmonic components contained in the actual vibration signal. When the vibration signal collected by the probe originates from high-frequency mechanical excitation or the internal acoustic response of the wafer, the Fourier spectrum will present an energy distribution pattern with strong concentration and clear primary and secondary components. Therefore, the role of this step is to clarify the main frequency range where the vibration signal exists and thus derive the preliminary vibration frequency range. Determine the search interval based on the preliminary vibration frequency range, that is, set a search interval that contains a high-probability frequency distribution, and perform a dense scanning operation on the frequency points within this interval. This scanning process evenly divides multiple points on the frequency axis with a fixed step size, and each point represents a candidate frequency to be fitted. The obtained set of frequency points constitutes the target frequency candidate set, which includes all possible frequencies covering the preliminary main frequency range and has sufficient density to ensure the fitting accuracy and the stability of the subsequent model convergence. Map the target frequency candidate set to the cosine space. Substitute each frequency to be measured into a cosine function model framework and construct a set of simulated signals. This signal is based on the cosine waveform, and the parameters include frequency, amplitude, and phase, which can approximately represent the change trend of the actual covariance sample. This cosine space model serves as a theoretical fitting template, and its output is used to match the real covariance data, thereby inversely deducing the parameter combination that best conforms to the actual observed data. During the modeling process, the model initially does not contain specific values but serves as an adjustable structural framework waiting for the input of the subsequent data-driven optimization process. The model itself reflects the fluctuation pattern of the vibration signal in the time series as the parameters change, so it has high fitting flexibility and iterability. Based on the parameter solution model established in the cosine space, introduce the iterative least squares method to perform the first-round parameter fitting, and initialize the key parameters in the model, namely frequency, amplitude, and phase. The initial frequency is taken from the frequency point corresponding to the energy peak in the previous Fourier transform, the initial amplitude is set to the maximum fluctuation amplitude in the covariance sample, and the initial phase is set to zero or any reference angle. Using these initial values as the input for the first-round fitting, calculate a set of preliminary vibration characteristic parameters by minimizing the sum of the squared errors between the model output and the covariance sample. Substitute the first-round fitting results into the original covariance sample, calculate the residual sequence between the actual signal and the fitted signal, and use the sum of the squared residuals as the objective function to evaluate the quality of the current fitting parameters. On this basis, gradually update the three key parameters through the gradient descent method, that is, calculate the partial derivative information of the current sum of the squared residuals with respect to frequency, amplitude, and phase, determine the update direction and step size of each parameter, and perform fine-tuning of the parameters in the direction of the minimum error in each round of iteration, thereby continuously optimizing the fitting accuracy.The gradient descent process uses a fixed or adaptive learning rate for step size control and combines a momentum factor to improve the convergence stability and the ability to prevent getting stuck in local optima. As the iteration progresses, the system recalculates the sum of squared residuals and determines its rate of change after each parameter update. When the rate of change of the sum of squared residuals for two consecutive rounds is less than a preset threshold, it indicates that the model has basically reached the optimal state and there is no significant room for improvement. At this time, the iteration process is terminated, and the current optimal fitting parameters are output as the final vibration characteristic identification result. This set of results includes three key indicators with clear physical meanings, namely the dominant frequency representing the actual vibration response, the amplitude parameter describing the signal amplitude intensity received by the probe at this frequency, and the initial phase angle determining the signal start position and propagation phase structure. These parameters together constitute the spatial vibration signal model observed by the probe in the working state.

[0031] In a specific embodiment, the process of performing step 300 may specifically include the following steps: Set multiple temperature test points on the wafer probe table; Use a multi-probe measurement structure to collect electrical signals of E measurement paths at multiple temperature test points. Each measurement path includes a signal source, a probe contact point, and a signal receiver, and obtain multi-measurement path electrical signals; Perform normalization and time-domain alignment processing on the multi-measurement path electrical signals to obtain preprocessed multi-measurement path signals; Through the path fusion module of the baseline autoencoder, perform weighted fusion on the preprocessed multi-measurement path signals to obtain a fused feature input vector; Pass the fused feature input vector through the input layer of the baseline autoencoder to the first hidden layer for non-linear transformation to obtain the first layer feature mapping, and pass the first layer feature mapping to the second hidden layer and use the ReLU activation function for processing to obtain the ambient temperature feature.

[0032] Specifically, multiple temperature test points are set within the structure of the wafer probe station to cover different thermal gradient regions. Different thermal distribution conditions are applied to the probe station through a heating device or an environmental control system to construct multiple stable temperature states. Data acquisition operations are performed at each temperature point to ensure the systematicness of the test and the continuity of the temperature field conditions. After the temperature control is completed and the thermal steady state is reached, a multi-probe measurement structure is used for signal acquisition, and complete electrical response data need to be collected under each temperature state. In this measurement structure, E electrical measurement paths are set, and each path includes three key links, namely a signal source, a probe contact point, and a signal receiver. The signal source is used to generate a standard electrical signal, which is injected into the wafer surface through the probe. The probe contact point couples the signal into the wafer material and finally transmits it back to the measurement terminal through the receiver. During this process, due to differences in the position, contact state, and thermal distribution of each path, the response signals will reflect the characteristic response patterns under the corresponding temperature states. By completely collecting the response data of these E paths at multiple temperature points, an original electrical signal data set containing multi-paths, multi-temperatures, and multi-time slices is formed. The electrical signals of multiple measurement paths are subjected to normalization and time-domain alignment processing. The normalization process adjusts the voltage or current amplitudes of all paths to a unified range, removes the amplitude offsets caused by the gain differences of each channel, makes the signal intensities between different paths comparable, and improves the numerical stability of model training, avoiding training offsets caused by inconsistent numerical scales. Subsequently, time-domain alignment processing is performed to ensure the synchronization of the collected signals of different paths in the time coordinate. Due to certain channel delays in hardware sampling, the sampling data is calibrated in time sequence through a reference signal or a synchronization mark to eliminate small time drifts and achieve the time consistency of the signals under the same temperature state, obtaining a set of multi-measurement path signals that have been standardized and time-aligned. The preprocessed multi-path signals are input into the path fusion module of the baseline autoencoder. This module performs weighted fusion on the multi-channel signals through a weight assignment mechanism. The path fusion module automatically assigns different fusion weights to each path according to the response stability and feature representativeness of each path under different temperature states. The weights are dynamically generated through methods such as attention mechanism, channel correlation scoring, or historical residual learning, so as to strengthen the expression ability of important paths during the fusion process while suppressing the noise interference of low-quality paths. The fusion result is a compressed feature input vector that retains the main change trends in the multi-channel signals and serves as the input to the encoder part in a unified structural format, realizing the aggregation of information and the compactness of the structure. The fused feature input vector is fed into the input layer of the baseline autoencoder, and the network structure transmits it to the first hidden layer for non-linear feature transformation. This hidden layer consists of multiple groups of neurons. Each neuron forms a preliminary activation value by performing a weighted sum with each element in the input vector and adding a bias, and then undergoes a transformation through a non-linear mapping function to achieve the extraction and compression of linearly inseparable features.In this hidden layer, the encoder captures the structural changes in the electrical signal that are highly correlated with temperature changes, and initially generates the first-layer feature mapping result. The first-layer feature mapping is then passed on to the second hidden layer, where the ReLU activation function is selected to perform a second non-linear transformation on the output of the first layer. As a sparsity activation function, ReLU only retains the positive signals and suppresses the negative fluctuations, and its characteristics enable the neural network to have stronger expressive power and gradient propagation stability. Under the action of ReLU, the second hidden layer extracts the key feature information in the signal to form a set of high-dimensional abstract features, which are regarded as the temperature characterization results in the current test environment, that is, the ambient temperature features.

[0033] In a specific embodiment, the process of executing step 400 may specifically include the following steps: Through the decoder part of the baseline autoencoder, the baseline signals of each measurement path are reconstructed according to the ambient temperature features to obtain the reconstructed baseline signals, and the difference between the actually measured electrical signal and the reconstructed baseline signal is calculated to obtain the signal deviation values of each measurement path; Perform time series analysis on the signal deviation values of each measurement path to obtain a time delay model; Based on the time delay model, calculate the signal transmission delay caused by temperature changes between different measurement paths, and perform time delay compensation on the electrical characteristic values of each measurement point to obtain a preliminary corrected signal; Based on the preliminary corrected signal, calculate the posterior probability distribution of the temperature under the condition of observing the signal deviation to obtain the probability weight coefficient, and perform weighted compensation on the electrical characteristic measurement values according to the probability weight coefficient to obtain the temperature-compensated electrical signal.

[0034] Specifically, through the decoder part of the baseline autoencoder, the temperature feature vector output by the encoder is used as the input signal and transmitted to the decoding network for restoration and reconstruction. The decoder structure is relatively symmetric to the encoder structure and contains multiple fully connected layers or convolutional expansion modules inside, which are used to gradually expand the compressed temperature features and reconstruct them into a signal format with the same structure as the original input data. The decoding process is essentially a fitting and restoration operation on the input features, and its goal is to use the neural network to map and fit the relationship between the environmental temperature and the electrical signal, so as to construct a theoretical undisturbed baseline electrical response that should be presented under the current temperature conditions. The reconstructed output is the standard baseline signal corresponding to multiple measurement paths in the current thermal environment. This signal represents the pure thermal influence caused by temperature changes and does not contain random errors or structural interferences. The reconstructed baseline signal is compared and analyzed with the electrical signal measured in the actual test channel by channel. The difference between the measured signal and the corresponding reconstructed baseline signal in each measurement path is calculated, and this difference is defined as the signal deviation value. The signal deviation reflects the abnormal disturbance part that exceeds the theoretical thermal response range in the actual measurement. This part comes from non-ideal temperature response, material transient behavior, impedance unevenness during signal propagation, contact nonlinearity or coupling changes caused by mechanical stress. The signal deviation values obtained for each path form a dynamic data sequence that changes with time. Time series analysis is performed on the signal deviation values of each measurement path to extract the potential time response structure law. During the time series analysis process, a representative observation time window is selected, and the change trend of the deviation values of each path within this time window is modeled. By comparing the response lag degrees between different paths, a delay relationship structure between the paths is established. This structure calculates the change curve of the similarity of the deviation signals between different paths with time delay, and extracts the time difference corresponding to the maximum similarity, thereby deriving the signal propagation time delay difference between different paths. The resulting time delay model reflects the response delay required when a temperature perturbation propagates from one path to another path, and has multiple characteristics such as spatial directionality, path correlation, and temperature dependence. According to the transmission delay law reflected by the time delay model, time domain correction processing is performed on the electrical characteristic values of each measurement point. The measured signal is shifted forward or backward on the time axis according to the time delay difference between the paths, so that all signals logically correspond to the same thermal response moment. Through the time series compensation based on path lag, the response misalignment caused by uneven temperature gradient or material thermal inertia difference is effectively corrected uniformly, and a set of preliminarily corrected signals with consistent structure time and aligned thermal responses is obtained. Based on the preliminarily corrected signal, the posterior probability distribution of the temperature under the condition of the observed signal deviation is calculated to evaluate the causal association strength between the current signal deviation and the temperature factor.The posterior probability distribution of temperature reflects the likelihood that a certain amplitude deviation is caused by temperature changes when observed. It is approximately solved by Bayesian inference or variational inference methods and estimated in combination with the deviation-temperature correlation pattern in historical samples. A set of posterior probability distributions is generated for the signal deviations on each path, and the probability weight coefficients representing the confidence of its thermal response are extracted from them. This coefficient is a quantitative expression of the degree to which the electrical characteristics value of the current path is dominated by temperature perturbations. The probability weight coefficients are used to weight and compensate the measured electrical characteristic values at each measurement point, and the deviation values are corrected according to the probability weights, so that the temperature-dominated component is completely regressed, and the perturbation part caused by non-thermal factors is weakened or eliminated, forming a set of temperature-compensated electrical signals that integrate the thermal response model, time alignment mechanism, and probability weight compensation.

[0035] In a specific embodiment, the process of performing steps to perform time series analysis on the signal deviation values of each measurement path and obtaining the time delay model may specifically include the following steps: Perform cross-correlation analysis on the signal deviation values of each measurement path to obtain the delay relationship matrix between the measurement paths; Extract the time delay value corresponding to the maximum cross-correlation coefficient from the delay relationship matrix, and summarize the time delay values of all measurement path pairs to obtain the original time delay data set; Perform hierarchical clustering on the original time delay data set to obtain a hierarchical time delay data structure; Based on the hierarchical time delay data structure, select control points for each layer of time delay data respectively, and determine the coefficients of the cubic spline function by the least squares method to obtain the time delay characteristic curve equation; Perform boundary constraint processing on the time delay characteristic curve equation to obtain a smooth time delay model function, and establish a mapping relationship between the time delay model function and the temperature change rate to obtain the time delay model.

[0036] Specifically, the time synchronization between each measurement path is systematically analyzed based on the signal deviation value. In multi-channel signal acquisition, all paths are theoretically sampled at the same starting time point. However, due to uneven thermal diffusion rates, different path structures, non-uniform material distributions, and differences in signal transmission path lengths, there are significant delay differences in the time responses of the signals of each path to temperature changes. Cross-correlation analysis is performed on the signal deviation values to quantify the time displacement relationship between each pair of measurement paths. Cross-correlation, as an analysis method to measure the similarity of two sets of signals at different time offsets, reveals the delay amount corresponding to the optimal alignment point between the two signals. By taking the signal deviation of each path as a reference benchmark and performing cross-correlation operations with the deviation signals of other paths one by one, that is, a set of time delay response curves is established between all pairs of measurement paths, and a symmetric delay relationship matrix is constructed therefrom. Each element of this matrix represents the delay time amount corresponding to the optimal cross-correlation between two paths. The time delay value corresponding to the maximum cross-correlation coefficient is extracted from the delay relationship matrix. This delay time is the optimal synchronization point between the two paths under the current thermal perturbation condition, representing the true transmission time difference of thermal diffusion or signal propagation between the path pair. The maximum cross-correlation time points of the path pairs are respectively extracted and sorted in a unified numbering manner according to the path pair combination method, and a set of time delay data covering all path pairs is summarized to form the original time delay data set. This data set is a two-dimensional structure composed of path pair indices and corresponding delay amounts in form, which is mathematically represented as an asymmetric time transfer matrix and physically represents the lagged propagation behavior actually existing in the multi-path response system. Structural analysis is performed on the original time delay data set to identify the hierarchical laws or regional response characteristics existing therein. On this basis, a hierarchical clustering algorithm is introduced to group and cluster the time delay data. During the clustering process, according to the relative proximity of the delay values between the path pairs, they are gradually merged into several classes, and a tree-like clustering structure including different aggregation levels is constructed. This hierarchical time delay data structure is presented as a top-down, coarse-to-fine path lag behavior organization system in form. In each layer, path pairs with similar lag degrees are classified into a time synchronization subgroup, and the differences between different layers reflect the fact that there are structural differences in local area thermal diffusion or signal response within the system. Curve function fitting is performed on each layer in the hierarchical time delay data structure. The cubic spline function is used as the function form. In each layer, several representative control points are selected from the set of path pairs belonging to this layer. The selection of the control points is based on criteria such as the extreme value, mean value, or change inflection point of the lag amount to ensure its structural representativeness within this layer. Using these control points as nodes, the coefficients of the cubic spline function are solved by the least squares method.The least squares method determines a set of optimal curve parameters by minimizing the sum of the squares of the errors between the fitted curve and the true time-delay data, so that the generated spline curve not only passes through the control points, but also has a high overall fitting degree for all path pair data. After fitting, a set of independent time-delay characteristic curve equations are established for each clustering level to reflect the response lag patterns of path pairs within that level under different temperature or time conditions. A boundary constraint processing mechanism is introduced on the fitted spline curve to limit the derivative values at both ends of the function, restrict the curvature change range, or impose boundary extension conditions, so that the curve maintains a stable transition at the boundary of its domain and has physical interpretability. The curve after boundary processing has overall continuity, piecewise differentiability, and local convergence, forming a time-delay model function. A mapping relationship is established between the time-delay model function and the temperature change rate, so that the model adaptively adjusts its output structure according to different temperature change rates. This process analyzes the response delay data of path pairs in multiple different temperature change rate fields and constructs a functional mapping relationship between the temperature change rate and the time-delay function parameters. Methods such as polynomial regression, spline interpolation, neural networks, or basis function expansion are used to establish this relationship model, so that in practical applications, only the current temperature change rate needs to be input to dynamically generate the corresponding path time-delay function, thereby completing the entire system-level thermal coupling delay compensation mechanism and finally obtaining the time-delay model.

[0037] In a specific embodiment, the process of executing step 500 may specifically include the following steps: Based on the vibration frequency and amplitude parameters in the probe vibration characteristic data, perform compensation calculation on the contact impedance to obtain the physical layer calibration result; Fuse and process the physical layer calibration result with the temperature-compensated electrical signal to obtain the signal layer calibration result; According to the signal layer calibration result and the probe position parameters, establish an electrical characteristic parameter mapping function to obtain the parameter layer calibration result; Perform adaptive weight calculation on the physical layer calibration result, the signal layer calibration result, and the parameter layer calibration result to obtain the weighted calibration coefficient; Based on the weighted calibration coefficient, perform weighted fusion on the three-layer calibration results to obtain the calibrated electrical characteristic data, and perform consistency verification and outlier detection on the calibrated electrical characteristic data to obtain the wafer electrical characteristic measurement result.

[0038] Specifically, based on the vibration frequency and amplitude parameters in the probe vibration characteristic data. The vibration amplitude reflects the elastic deformation or clearance swing behavior that occurs during the contact process of the probe, while the vibration frequency reflects the periodic speed at which such disturbances occur. Therefore, both together determine the fluctuation range of the impedance characteristics of the contact surface at the microscale. The greater the vibration intensity, the more unstable the contact area, and the charge transfer efficiency per unit area also decreases accordingly, resulting in the periodic drift characteristics of the contact impedance. Based on this physical mechanism, an impedance compensation model with frequency and amplitude as input conditions is constructed. This model predicts the impedance deviation through the structural response function and outputs a set of compensation values corresponding to the actual vibration state. These compensation values act on the voltage or current indicators reflecting the contact part in the original electrical measurement results, performing corrections in the amplitude or phase direction, so that the compensated signal is closer to the electrical response under the ideal contact state. The corrected data is the physical layer calibration result. The physical layer calibration result is fused with the temperature-compensated electrical signal. Temperature changes can cause thermal expansion of materials, resistance drift, and thermoelectric effects, while vibration mainly leads to unstable contact states. The two act on the electrical signal responses in different dimensions. The two types of compensation results are structurally fused to eliminate their respective local residuals under separate processing and form an integrated signal representation with overall robustness in terms of space, thermodynamics, and dynamics. During the fusion process, signal weighted superposition, moving average, or filter fusion mechanisms are used, with the temperature compensation signal as the reference state and the physical layer compensation result as the high-frequency correction term, ensuring that the fused signal achieves a balance between long-term stability and short-period response. After the fusion is completed, the signal layer calibration result is obtained. According to the signal layer calibration result and the probe position parameters, an electrical characteristic parameter mapping function is established. This function establishes a non-linear mapping relationship between the calibrated electrical response and the probe geometric coordinates, and corrects the spatial distribution error by means of interpolation, neural network, or locally weighted regression. The parameter layer calibration result is obtained through this process. Adaptive weight calculations are performed on the physical layer calibration result, the signal layer calibration result, and the parameter layer calibration result to obtain the weighted calibration coefficient. The weight allocation process is determined based on the characteristic intensity of the dominant interference source in the current test conditions. For example, when the vibration amplitude increases significantly, the physical layer weight is increased; when the temperature fluctuates violently, the signal layer weight is increased; and when the structural differences are obvious or the probe layout density is uneven, the influence ratio of the parameter layer calibration is strengthened. The three-layer calibration results are fused according to their weighted coefficients, and the calibrated electrical characteristic data is output. Consistency verification and outlier detection are performed on the calibrated electrical characteristic data. Consistency verification is carried out by statistically comparing the current result with historical similar test data to determine whether it is within a reasonable physical range and trend interval; outlier detection is carried out by identifying mutation values, outliers, or continuous deviation trends to eliminate measurement errors introduced by factors such as probe wear, data drift, or system misjudgment. Through the above steps, the wafer electrical characteristic measurement result is obtained.

[0039] The probe calibration method in the embodiments of the present application has been described above. Next, the probe calibration device 10 in the embodiments of the present application will be described. Please refer to Figure 2 In one embodiment, the probe calibration device 10 in the embodiments of the present application includes: A construction module 11, configured to calculate a vibration wavelength according to a preset vibration frequency and a sampling frequency and construct a multi-probe measurement structure; A calculation module 12, configured to synchronously collect vibration signals on the surface of a wafer by using the multi-probe measurement structure, calculate the covariance value between each pair of probes, and obtain probe vibration characteristic data; A processing module 13, configured to collect multi-measurement path electrical signals of the multi-probe measurement structure under different temperature conditions and process them through a baseline autoencoder to obtain environmental temperature characteristics; A correction module 14, configured to calculate a signal deviation value between a measured signal and a baseline signal according to the environmental temperature characteristics and correct the signal deviation value to obtain temperature-compensated electrical signals; A hierarchical calibration module 15, configured to perform hierarchical calibration on the probe vibration characteristic data and the temperature-compensated electrical signals to obtain a measurement result of the electrical characteristics of the wafer.

[0040] Through the collaborative cooperation of the above-mentioned various components, the probe calibration method proposed by the present invention constructs a multi-probe measurement structure with a general covariance layout, physically ensuring the measurability of continuous covariance samples, fundamentally solving the under-sampling problem caused by the mismatch between high-frequency probe vibration and low sampling frequency, and significantly improving the accuracy of vibration parameter identification. The baseline autoencoder calibration technology of the present invention does not require any biased samples for training, can adaptively extract environmental temperature characteristics from measurement data, realize real-time monitoring and compensation of temperature changes, and greatly improve the stability of electrical characteristic measurement in a variable-temperature environment. The present invention innovatively designs a delay-based probability compensation algorithm, which not only considers the influence of temperature changes on the signal amplitude, but also captures the influence of the temperature change rate on the signal phase through a time-delay model, achieving a more comprehensive and accurate temperature compensation effect. The present invention adopts a hierarchical calibration strategy, dividing the calibration process into three levels: the physical layer, the signal layer, and the parameter layer, respectively performing precise calibration on the contact instability caused by probe vibration, the signal offset caused by temperature changes, and the final electrical characteristic parameters, avoiding error accumulation and mutual interference. The present invention introduces an adaptive weight mechanism, which can dynamically adjust the weights of each layer of calibration according to the probe vibration state and the degree of temperature change, optimize the calibration effect, and make the calibration method have stronger environmental adaptability and robustness.

[0041] The probe control device provided by the embodiments of the present application is used to execute any one of the above-mentioned probe calibration methods.

[0042] It should be noted that those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working process of the probe control device described above can refer to the corresponding process of the foregoing probe calibration method, and will not be elaborated here.

[0043] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the systems, devices, and units described above can refer to the corresponding processes in the foregoing method embodiments, and will not be elaborated here.

[0044] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to enable a probe control device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of this application. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store program codes.

[0045] As described above, the above embodiments are only used to illustrate the technical solution of this application, rather than to limit it; although this application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of various embodiments of this application.

Claims

1. A probe calibration method, characterized in that, Including: Calculating the vibration wavelength according to a preset vibration frequency and a sampling frequency, and constructing a multi-probe measurement structure; Synchronously collecting vibration signals on the wafer surface by using the multi-probe measurement structure, calculating the covariance values between each pair of probes, and obtaining probe vibration characteristic data; Collecting multi-measurement path electrical signals of the multi-probe measurement structure under different temperature conditions, and processing them through a baseline autoencoder to obtain environmental temperature characteristics; Calculating the signal deviation value between the measured signal and the baseline signal according to the environmental temperature characteristics, and correcting the signal deviation value to obtain temperature-compensated electrical signals; Performing hierarchical calibration on the probe vibration characteristic data and the temperature-compensated electrical signals to obtain the measurement result of the wafer electrical characteristics.

2. The probe calibration method according to claim 1, wherein The calculating the vibration wavelength according to a preset vibration frequency and a sampling frequency, and constructing a multi-probe measurement structure includes: Calculating the vibration wavelength λ according to the preset vibration frequency and sampling frequency, and determining the probe spacing parameters d1 and d2 according to the vibration wavelength λ to obtain a probe spacing design scheme; Performing probe positioning on the wafer probe table according to the probe spacing design scheme, placing the first probe at the coordinate (0,0), the second probe at the coordinate (d1,0), the third probe at the coordinate (0,d2), and the fourth probe at the coordinate (d1,d2) to obtain a four-probe layout structure; Expanding the four-probe layout structure, and setting four probes at the coordinates (d1 / 2,d2 / 2), (d1 / 2,0), (0,d2 / 2), and (d1,d2 / 2) to obtain an eight-probe layout structure; Placing each probe in the eight-probe layout structure according to the shape characteristics of the wafer test area to obtain a multi-probe measurement structure.

3. The probe calibration method according to claim 1, wherein The synchronously collecting vibration signals on the wafer surface by using the multi-probe measurement structure, calculating the covariance values between each pair of probes, and obtaining probe vibration characteristic data includes: Applying a fixed contact pressure to each probe in the multi-probe measurement structure, and setting the sampling frequency to synchronously collect the vibration signals on the wafer surface of each probe; Sorting out the vibration signals on the wafer surface of each probe to obtain original vibration sampling data, and preprocessing the original vibration sampling data to obtain standardized vibration signals; According to the eight-probe layout structure in the multi-probe measurement structure, selecting four pairs of probe positions of (0,0) and (d1,0), (0,0) and (0,d2), (d1,0) and (d1,d2), and (0,d2) and (d1,d2), and extracting the corresponding standardized vibration signals to obtain four signal pairs; Calculating the covariance for each of the four signal pairs to obtain four covariance values with different phase differences; Constructing a time series based on the four covariance values with different phase differences, and performing a sliding window process on this time series to obtain a covariance sample series; Performing detrending processing and smoothing filtering on the covariance sample series to obtain a covariance sample set; Performing Fourier transform and iterative least squares fitting on the covariance sample set to obtain probe vibration characteristic data.

4. The probe calibration method according to claim 3, wherein Performing Fourier transform and iterative least squares fitting on the covariance sample set to obtain probe vibration characteristic data, including: Performing Fourier transform on the covariance sample set to obtain a preliminary vibration frequency range; Determining a search interval based on the preliminary vibration frequency range, scanning frequency points within the search interval to obtain a target frequency candidate set; Mapping the target frequency candidate set to the cosine space to obtain a vibration parameter solving model; Applying the iterative least squares method to the vibration parameter solving model, initializing frequency, amplitude, and phase parameters, performing the first round of parameter fitting to obtain preliminary vibration characteristic parameters; Calculating the residual based on the preliminary vibration characteristic parameters, and iteratively optimizing the vibration frequency, amplitude, and phase parameters through the gradient descent method until the change rate of the sum of squared residuals is less than a preset threshold to obtain the probe vibration characteristic data.

5. The probe calibration method according to claim 1, wherein Collecting multi-measurement path electrical signals of the multi-probe measurement structure under different temperature conditions and processing them through a baseline autoencoder to obtain ambient temperature characteristics, including: Setting multiple temperature test points on the wafer probe station; Collecting electrical signals of E measurement paths using the multi-probe measurement structure under the multiple temperature test points, each measurement path including a signal source, a probe contact point, and a signal receiver, to obtain multi-measurement path electrical signals; Performing normalization and time-domain alignment processing on the multi-measurement path electrical signals to obtain preprocessed multi-measurement path signals; Performing weighted fusion on the preprocessed multi-measurement path signals through the path fusion module of the baseline autoencoder to obtain a fused feature input vector; Passing the fused feature input vector through the input layer of the baseline autoencoder to the first hidden layer for non-linear transformation to obtain a first-layer feature map, and passing the first-layer feature map to the second hidden layer and processing it using the ReLU activation function to obtain the ambient temperature characteristics.

6. The probe calibration method according to claim 1, wherein Calculating the signal deviation value between the measured signal and the baseline signal based on the ambient temperature characteristics and correcting the signal deviation value to obtain temperature-compensated electrical signals, including: Reconstructing the baseline signal of each measurement path according to the ambient temperature characteristics through the decoder part of the baseline autoencoder to obtain a reconstructed baseline signal, and calculating the difference between the actually measured electrical signal and the reconstructed baseline signal to obtain the signal deviation value of each measurement path; Performing time series analysis on the signal deviation values of each measurement path to obtain a time delay model; Calculating the signal transmission delay amount caused by temperature changes between different measurement paths based on the time delay model and performing time delay compensation on the electrical characteristic values of each measurement point to obtain a preliminary corrected signal; Calculating the posterior probability distribution of temperature under the condition of observing the signal deviation based on the preliminary corrected signal to obtain a probability weight coefficient, and performing weighted compensation on the electrical characteristic measurement values according to the probability weight coefficient to obtain temperature-compensated electrical signals.

7. The probe calibration method according to claim 6, characterized in that Performing time series analysis on the signal deviation values of each measurement path to obtain a time delay model, including: Perform cross-correlation analysis on the signal deviation values of each measurement path to obtain a delay relationship matrix between the measurement paths; Extract the time delay value corresponding to the maximum cross-correlation coefficient from the delay relationship matrix, and summarize the time delay values of all measurement path pairs to obtain an original time delay data set; Perform hierarchical clustering on the original time delay data set to obtain a hierarchical time delay data structure; Based on the hierarchical time delay data structure, select control points for each layer of time delay data respectively, and determine the coefficients of the cubic spline function by the least squares method to obtain a time delay characteristic curve equation; Perform boundary constraint processing on the time delay characteristic curve equation to obtain a smooth time delay model function, and establish a mapping relationship between the time delay model function and the temperature change rate to obtain a time delay model.

8. The probe calibration method according to claim 1, characterized in that The hierarchical calibration of the probe vibration characteristic data and the temperature-compensated electrical signal to obtain the wafer electrical characteristic measurement result includes: Based on the vibration frequency and amplitude parameters in the probe vibration characteristic data, perform compensation calculation on the contact impedance to obtain a physical layer calibration result; Fuse the physical layer calibration result with the temperature-compensated electrical signal to obtain a signal layer calibration result; According to the signal layer calibration result and the probe position parameters, establish an electrical characteristic parameter mapping function to obtain a parameter layer calibration result; Perform adaptive weight calculation on the physical layer calibration result, the signal layer calibration result, and the parameter layer calibration result to obtain a weighted calibration coefficient; Based on the weighted calibration coefficient, perform weighted fusion on the three-layer calibration results to obtain calibrated electrical characteristic data, and perform consistency verification and outlier detection on the calibrated electrical characteristic data to obtain the wafer electrical characteristic measurement result.

9. A probe calibration device, characterized in that, For performing the probe calibration method according to any one of claims 1-8, the probe calibration device includes: A construction module for calculating the vibration wavelength according to a preset vibration frequency and sampling frequency and constructing a multi-probe measurement structure; A calculation module for synchronously collecting the wafer surface vibration signals by using the multi-probe measurement structure and calculating the covariance values between each pair of probes to obtain probe vibration characteristic data; A processing module for collecting multi-measurement path electrical signals of the multi-probe measurement structure under different temperature conditions and processing them through a baseline autoencoder to obtain environmental temperature characteristics; A correction module for calculating the signal deviation value between the measured signal and the baseline signal according to the environmental temperature characteristics and correcting the signal deviation value to obtain a temperature-compensated electrical signal; A hierarchical calibration module for performing hierarchical calibration on the probe vibration characteristic data and the temperature-compensated electrical signal to obtain the wafer electrical characteristic measurement result.

10. A probe control device, characterized in that, The probe control device is used to execute the probe calibration method according to any one of claims 1-8.

Images