A method and system for evaluating the full-chain application performance of a precision grinding wheel

Abstract

This invention proposes a method and system for evaluating the full-chain application performance of precision grinding wheels, including: collecting relevant data on grinding wheel performance from spline / block material properties, grinding wheel body characteristics, grinding process, and wafer post-grinding quality; establishing a multi-source heterogeneous performance detection database; performing preprocessing and key feature analysis to construct a dataset of key features of grinding wheel performance; constructing a hierarchical performance mapping model from material properties to wafer processing quality, and training it using the dataset of key features of grinding wheel performance; establishing quantitative evaluation criteria for four-dimensional feature indicators; constructing cross-scenario evaluation criterion transfer and correction rules; comprehensively evaluating the performance of the wafer precision grinding wheel in the evaluation scenario based on the key feature data of grinding wheel performance in each dimension; and using the trained hierarchical performance mapping model to trace the source of problems and optimize the process based on the comprehensive evaluation results. This invention achieves comprehensive evaluation of the application performance of grinding wheels across the entire chain and can locate grinding wheel performance defects at their root, avoiding the limitations of single-indicator evaluation.

Description

Technical Field

[0001] This invention relates to the technical fields of semiconductor manufacturing, ultra-precision grinding and grinding wheel performance, and in particular to a method and system for evaluating the full-chain application performance of wafer precision grinding wheels. Background Technology

[0002] Precision grinding for wafer thinning is a critical process in semiconductor chip manufacturing. Diamond grinding wheels are the core tools in this process, and their performance directly determines the wafer's surface roughness, subsurface damage, thickness uniformity (TTV / LTV), and batch stability, ultimately affecting chip yield and reliability. With the widespread application of third-generation semiconductor (silicon carbide, gallium nitride) wafers and the advancement of semiconductor manufacturing processes, higher demands are placed on the precision, consistency, and lifespan of precision grinding wheels for thinning.

[0003] A precise and comprehensive grinding wheel performance evaluation system is the core tool for controlling grinding wheel quality, optimizing manufacturing processes, and stabilizing wafer thinning processes. It is not only an important support for promoting the domestic production and upgrading of high-end thinning grinding wheels, but also a necessary prerequisite for ensuring the quality of advanced semiconductor wafer processing, improving chip production yield, and reducing process production losses. It has extremely strong practical application value and industrial promotion significance in the field of semiconductor precision manufacturing.

[0004] Current grinding wheel performance evaluation methods in the industry suffer from the following core pain points: ① Fragmented evaluation, lacking a complete chain of correlation: Existing evaluations often focus on a single grinding wheel index or the quality of the wafer after grinding. For example, the invention patent with publication number CN111222258A discloses a grinding wheel grinding performance classification method based on the crystal plane orientation of diamond abrasive grains. By statistically analyzing the three-dimensional data of diamond abrasive grains and performing crystal plane orientation simulation before grinding wheel production, it solves the destructive problem of grinding wheel grinding performance evaluation in existing technologies and achieves non-destructive and widely representative grinding wheel performance evaluation. However, this method does not establish a closed-loop mapping relationship of "grinding wheel material characteristics - grinding wheel body state - grinding process behavior - wafer processing quality", and cannot locate the root cause. ① Grinding wheel performance defects; ② Lack of evaluation of key interface performance: The interfacial bonding strength between diamond abrasive grains and the binder is the core influencing factor on grinding wheel life and anti-scratch ability, but existing methods lack quantitative and reproducible evaluation means, and can only be indirectly evaluated through offline wear tests, which cannot predict the risk of interface failure in advance; ③ Insufficient online monitoring of the grinding process: The wear, clogging, dynamic balance deterioration, slippage and other states of grinding wheel during the grinding process cannot be monitored in real time, and offline sampling inspection is relied upon, which can easily lead to batch wafer scrapping; ④ Inconsistent evaluation standards and poor transferability: The evaluation indicators, test methods and pass thresholds of different manufacturers and different models of grinding wheels vary greatly, and a universal and transferable evaluation standard system has not been formed, making it difficult to achieve performance comparison and optimization across scenarios. Summary of the Invention

[0005] To address the technical problems of existing technologies, such as fragmented evaluation dimensions of wafer precision thinning grinding wheels leading to inability to achieve linked traceability, lack of quantitative evaluation of interface performance, insufficient online monitoring of the grinding process, and lack of unified and adaptable evaluation standards, this invention proposes a full-chain application performance evaluation method and system for precision grinding wheels. This system achieves a closed-loop evaluation from all dimensions, from wheel material characteristics to wafer processing quality, and establishes a transferable evaluation standard system, providing a scientific basis for wheel selection, optimization, and production process control.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows: a method for evaluating the performance of a precision grinding wheel across the entire chain, comprising the following steps:

[0007] S1: Collect data related to grinding wheel performance from four dimensions: spline / block material properties, grinding wheel body characteristics, grinding process, and wafer post-grinding quality, and establish a multi-source heterogeneous performance testing database in four dimensions;

[0008] S2: Preprocess and perform key feature analysis on multi-source heterogeneous performance test data to construct a four-dimensional dataset of key features of grinding wheel performance;

[0009] S3: Construct a hierarchical performance mapping model from material properties to wafer processing quality, and train the hierarchical performance mapping model using a dataset of key features of grinding wheel performance to obtain a trained hierarchical performance mapping model.

[0010] S4: Establish quantitative evaluation criteria for characteristic indicators covering four dimensions: spline / block material properties, grinding wheel body characteristics, grinding process, and wafer post-grinding quality;

[0011] S5: Construct cross-scenario evaluation criterion transfer and correction rules: set different correction coefficients for different processing scenarios, and dynamically adjust the scoring threshold according to the correction coefficient of the corresponding scenario;

[0012] S6: Based on quantitative evaluation criteria and cross-scenario evaluation criterion transfer correction rules, and according to the key feature data of grinding wheel performance in each dimension of the evaluation scenario, the performance of precision grinding wheel in the evaluation scenario is comprehensively evaluated. Based on the comprehensive evaluation results, the trained hierarchical performance mapping model is used to trace the source of problems and optimize the process.

[0013] Preferably, the performance dimensions of the spline / block material include basic physicochemical indicators, mechanical property parameters, microstructure characteristics, interfacial bonding strength indicators, and tribological and wear performance parameters;

[0014] The characteristic dimensions of the grinding wheel body include the macroscopic and microscopic characteristic parameters of the surface before grinding;

[0015] The grinding process dimension includes the dynamic signals of the grinding process and the evolution of the grinding wheel body characteristics;

[0016] The quality dimensions of the wafer after grinding include precision and surface integrity.

[0017] Preferably, the basic physicochemical properties include: density, porosity, hardness, abrasive concentration, and particle size distribution; the mechanical property parameters include: flexural strength, tensile strength, impact strength, and elastic modulus; the microstructure characteristics include: exposed abrasive particles, uniform distribution, breakage, and interfacial porosity; the interfacial bonding strength indicators include: shear strength, pull-out strength, interfacial fracture strength, and interfacial wettability; and the tribological properties include: force ratio, coefficient of friction and stability, material wear rate, and temperature rise.

[0018] The macroscopic characteristic parameters include: dynamic rotational runout of the grinding wheel, rotational profile, and tooth height wear; the microscopic characteristic parameters include: terrain roughness RSa, terrain dispersion RSq, average abrasive grain exit height h, and microscopic profile skewness Rsk.

[0019] The dynamic signals of the grinding process include: grinding force, spindle load power, grinding vibration, acoustic emission, and temperature of the grinding zone under all working conditions; the evolution of the grinding wheel body characteristics includes: the increase value of dynamic rotational runout of the grinding wheel, the range of changes in the grinding wheel rotational profile, the difference in tooth height wear, and the decrease values ​​of topographic roughness RSa, topographic dispersion RSq, average abrasive grain exit height h, and microscopic profile skewness Rsk.

[0020] The accuracy includes: total wafer thickness deviation (TTV), local thickness deviation (LTV), surface shape peak-valley deviation (PV), curvature (Bow), and warp (Warp); the surface integrity includes: surface roughness (Ra), surface micro-defects, subsurface damage layer thickness, and surface residual stress; surface micro-defects include fisheyes, scratches, cracks, and chipping.

[0021] Preferably, the preprocessing includes: applying a wavelet threshold filtering algorithm to the dynamic signal of the grinding process in the grinding process dimension for filtering and noise reduction; and using a timestamp synchronization method to perform time-series alignment between the dynamic signal of the grinding process after filtering and noise reduction and the data on the evolution of the grinding wheel body features in the stage interval.

[0022] For all index data of spline / block material performance dimension, grinding wheel body characteristic dimension, and wafer grinding quality dimension, the mean and standard deviation of the test data of each index are calculated according to the 3σ error judgment criterion, an outlier judgment threshold is set, and outlier data is removed in batches; the dataset after removing outlier data is completed by linear interpolation.

[0023] The key feature analysis employs feature extraction algorithms to mine and extract key features from the preprocessed data across four dimensions, including:

[0024] Correlation feature screening was performed on all index data related to the basic physicochemical indicators, mechanical property parameters, interfacial bonding strength indicators, and tribological wear performance parameters of the preprocessed spline / block material properties. Strong collinear redundant indicators with correlation coefficients ≥0.85 were removed. The min-max normalization method was used to uniformly map all remaining indicators to [0,1] to eliminate dimensional interference. The normalized features were then extracted by PCA principal component analysis to obtain the comprehensive features of all material properties data with a cumulative variance contribution rate of not less than 95%, resulting in the principal component index set A.

[0025] The macroscopic and microscopic feature parameters of the preprocessed grinding wheel body feature dimensions are subjected to correlation feature screening to remove the indicator data with correlation coefficient ≥ 0.9; the remaining grinding wheel indicator data are uniformly mapped to the [0,1] interval by normalization to obtain the feature indicator set B.

[0026] The preprocessed dynamic signal of the grinding process was segmented according to time sequence alignment. For each stage of the signal, the time domain features and wavelet time-frequency energy features were extracted by joint time-frequency domain feature analysis. The peak and mean values ​​of grinding force, the mean value of spindle grinding power, the ratio of grinding vibration amplitude to wavelet energy, the ratio of acoustic emission RMS to wavelet energy, and the peak and mean values ​​of grinding zone temperature were obtained for each stage. The correlation coefficient between the signal features of each stage was calculated to screen for correlation features. Strong collinearity and redundant features with correlation coefficient ≥ 0.85 were removed. The remaining features were mapped to the [0,1] interval by minimum-maximum normalization. Then, the comprehensive features with a cumulative variance contribution rate of not less than 95% for each stage were extracted by PCA principal component analysis to obtain the principal component index set C, D, E, ...

[0027] The number of indicators for the evolution of the preprocessed grinding wheel body features is divided into segments according to time alignment. The correlation coefficient between the signal features of each stage is calculated to carry out correlation feature screening. Collinear redundant features with correlation coefficient ≥0.9 are removed. The remaining features are mapped to the [0,1] interval through minimum-maximum normalization to obtain the feature index set J, K, L, ...;

[0028] For all the indicators of precision and surface integrity of the pre-processed wafer grinding quality dimension, correlation feature screening was carried out to remove strongly collinear redundant indicators with correlation coefficients ≥0.85. The min-max normalization method was used to uniformly map all remaining indicator data to the [0,1] interval. The normalized features were extracted by PCA principal component analysis to obtain the principal component indicator set X with a cumulative variance contribution rate of not less than 95%.

[0029] Preferably, the hierarchical performance mapping model includes a material-bulk state mapping model, a bulk state-grinding condition mapping model, and a grinding condition-wafer quality mapping model; the material-bulk state mapping model is a first-level model, the bulk state-grinding condition mapping model is a second-level model, and the grinding condition-wafer quality mapping model is a third-level model.

[0030] Using the principal component index set A of spline / block material properties as the input variable of the material-bulk state mapping model, and the feature index set B of the grinding wheel body characteristics and the feature index sets J, K, L, ... of each stage of the grinding wheel body characteristic evolution in the grinding process dimension as the output variables of the material-bulk state mapping model, the model is trained by multivariate nonlinear regression analysis combined with gradient boosting machine learning algorithm. The model is fitted with quantitative constraint formulas for the degradation rate and degree of the grinding wheel body state of static material parameters, and a quantitative relationship between material properties and the surface state changes of the grinding wheel body is established.

[0031] The feature index set B of the grinding wheel body feature dimension and the feature index set J, K, L, ... of each stage of the grinding wheel body feature evolution in the grinding process dimension are used as input variables of the body state-grinding condition mapping model. The principal component index set C, D, E, ... of each stage of the grinding process dynamic signal in the grinding process dimension are used as model output variables. The body state-grinding condition mapping model is built and trained by using multivariate nonlinear regression analysis combined with gradient boosting machine learning algorithm and time series correlation fitting method. The multi-input multi-output mapping function is obtained, and the intrinsic correspondence between the surface state change of the grinding wheel body and the characteristics of the on-site grinding dynamic condition signal is established.

[0032] The principal component index sets C, D, E, ... of each stage of the grinding process dynamic signal in the grinding process dimension are used as input variables of the grinding condition-wafer quality mapping model, and the principal component index set X of the wafer post-grinding quality dimension is used as output variables of the grinding condition-wafer quality mapping model. The grinding condition-wafer quality mapping model is iteratively trained by using multivariate nonlinear regression analysis combined with gradient boosting machine learning algorithm, and a quantitative prediction function between the grinding process dynamic signal and the wafer grinding quality index is established.

[0033] Preferably, the evaluation criteria in step S4 are processed by extracting standardized key features of each dimension, including: grinding wheel body morphology features and process evolution indicators that have only undergone minimum-maximum normalization; and comprehensive principal component indicators of material properties, grinding dynamic signals, and wafer post-grinding quality obtained by normalization and PCA principal component analysis dimensionality reduction.

[0034] The attributes of all evaluation features are defined as positive performance indicators, negative performance indicators, and neutral performance indicators.

[0035] For the indicators of grinding wheel body morphology and process evolution that are normalized without PCA principal component analysis, the positive, negative, or neutral attributes corresponding to the original indicators are directly inherited. The eigenvector weight matrix corresponding to the comprehensive principal component indicators is extracted, and the original indicator with the highest absolute weight value is selected as the dominant term. The overall category of the principal component is defined according to the original indicator attributes of the dominant term: if the dominant term is a positive performance indicator, the principal component is a positive comprehensive performance indicator; if the dominant term is a negative performance indicator, the principal component is a negative comprehensive performance indicator; if the dominant term is a neutral performance indicator, the principal component is a neutral comprehensive performance indicator.

[0036] A 0-100 point quantification system is used to score individual indicator attributes. Two-level scoring thresholds are set for all features: a minimum acceptable threshold T1 and a minimum excellent threshold T2. Based on the individual attribute score, three performance level intervals are defined: a single attribute score ≥ the minimum excellent threshold T2 is judged as Grade A (excellent); a minimum acceptable threshold T1 ≤ a single attribute score < the minimum excellent threshold T2 is judged as Grade B (acceptable); and a single attribute score < the minimum acceptable threshold T1 is judged as Grade C (unacceptable).

[0037] The implementation method of the cross-scenario evaluation criterion migration correction rule is as follows: Select multiple grinding test machines to conduct rough grinding, semi-fine grinding, and fine grinding process tests, covering various wafer substrates such as silicon wafers, silicon carbide, gallium nitride, and sapphire, to obtain data index samples under multiple scenarios; set standard machines, semi-fine grinding, and silicon wafers as benchmark scenarios, corresponding to basic qualified lower limit thresholds. Basic high-quality lower limit threshold For different machine tools, semi-finished grinding, and silicon wafer processing scenarios, the ratio of the average score of key feature indicators for each dimension to the average score of key feature indicators for each dimension in the benchmark scenario is taken as the machine tool system correction coefficient K. M For standard equipment, different processes, and silicon wafer processing scenarios, the ratio of the average score of key feature indicators for each dimension to the average score of key feature indicators for each dimension in the benchmark scenario is taken as the process correction coefficient K. P For processing scenarios involving standard equipment, semi-finished grinding, and different wafer materials, the ratio of the average score of key feature indicators for each dimension to the average score of key feature indicators for each dimension in the benchmark scenario is taken as the material correction coefficient K. w For multi-factor scenario changes, the corresponding correction coefficients are multiplied to obtain the overall scenario correction coefficient; the scoring threshold is dynamically adjusted based on the overall scenario correction coefficient K: the lower limit of qualification threshold. High-quality lower limit threshold .

[0038] Preferably, the problem tracing and process optimization involves tracing the source of performance non-compliance, abnormal operating conditions, or wafer quality defects encountered during the evaluation process layer by layer: When the wafer grinding quality is substandard, a three-level model is used to reverse solve for the principal component index set C, D, E, ... of the grinding dynamic signal causing the substandard quality. The located grinding principal component indexes are then subjected to inverse projection transformation using a pre-stored PCA feature vector weight matrix to obtain normalized original index features. Inverse normalization is then used to reconstruct the actual grinding dynamic signal data with true dimensions, thus locating the abnormal grinding operating conditions. Finally, the located grinding principal component indexes are input into a two-level model for inverse... The model deduces and solves the normalized characteristic index of the grinding wheel body under the matching working conditions. The grinding wheel body characteristics are restored to the actual measured index of the grinding wheel surface morphology through inverse normalization. Using the actual measured index of the grinding wheel body morphology as the output constraint, the corresponding spline / block material performance principal component index is located through the first-level model. The located material principal component index is converted by PCA principal component analysis inverse projection and inverse normalization to restore the original material inherent measured index data. In addition, the qualitative identification images of abrasive grain exposure, uniform distribution, damage and breakage, and interface porosity detected by microstructure feature detection are used for auxiliary analysis and verification to accurately pinpoint the root cause of grinding quality problems.

[0039] Based on the problem tracing results and model correlation patterns, we propose improvement schemes for optimizing grinding wheel material formulation, grinding wheel structural parameters, and matching dressing process with grinding process.

[0040] After obtaining the principal component indices of the spline / block material properties, these indices are used as input constraints to solve the characteristic indices of the grinding wheel body feature dimension and the characteristic indices of each stage of the grinding wheel body feature evolution in the grinding process dimension through a first-level model. Then, the characteristic indices of the grinding wheel body feature dimension and the characteristic indices of each stage of the grinding wheel body feature evolution in the grinding process dimension are used as input constraints to solve the principal component indices of each stage of the grinding process dynamic signal in the grinding process dimension through a second-level model. Finally, the principal component indices of each stage of the grinding process dynamic signal are used as input constraints to solve the principal component indices of the wafer post-grinding quality dimension through a third-level model.

[0041] Preferably, a linearly increasing transverse shear force is applied to a single abrasive grain using the diamond indenter of a nano-scratch instrument, and the critical load at the time of abrasive grain detachment is recorded simultaneously as the shear detachment strength; a normal pull-out test is performed on a composite plate sample of diamond sheet and binder, with the pull-out force increasing linearly until the interface between the diamond sheet and the binder separates, and the detached section is observed and analyzed under a microscope. If the detached section shows the complete separation of the diamond sheet and the binder, the pull-out force is recorded as the pull-out bond strength; if there is binder residue on the detached section, the pull-out force is recorded as the interfacial fracture strength; the wetting angle of the binder on the diamond surface is observed after the binder is melted to test and evaluate the interfacial wettability;

[0042] The laser triangular displacement sensor is used to scan and collect the global macroscopic morphological fluctuation data of the surface under the state of grinding wheel rotation. The collected data is processed by periodic filtering, outer envelope feature point extraction and spline fitting. The fitted curve is converted into polar coordinates and displayed as the rotation profile of the grinding wheel. The height difference between the highest and lowest points in the rotation profile is calculated as the dynamic rotation runout of the grinding wheel. The difference between the mean values ​​of all data points in the rotation profile before and after grinding wheel wear is the tooth height wear amount.

[0043] The micro-topographic contour data of the grinding wheel surface was collected by in-situ scanning using a point spectral sensor. The collected data was processed by sliding window threshold filtering, contour median line extraction, peak and valley feature point extraction, and feature parameter calculation to obtain the topographic roughness RSa, topographic dispersion RSq, average abrasive grain tipping height h, and micro-profile skewness Rsk.

[0044] The terrain roughness RSa is: ;

[0045] The terrain dispersion RSq is: ;

[0046] The micro-profile skewness Rsk is: ;

[0047] The average abrasive grain exit height h is: ;

[0048] In the formula: n is the total number of data points collected for the micro-topographic contour of the entire surface of the grinding wheel. Let i be the measured value of the i-th collection point. Let be the value on the median line of the contour corresponding to the i-th sampling point, and t be the total number of data points of the contour peak data of the micro-topography of the entire surface of the grinding wheel. Let j be the measured value of the j-th peak. The value on the midline of the contour corresponding to the j-th peak point;

[0049] A three-dimensional force gauge, power meter, accelerometer, acoustic emission sensor, and infrared thermal imager were set up in the grinding area of ​​the grinding test platform. The grinding force, spindle load power, grinding vibration, acoustic emission, and grinding zone temperature were acquired using a data acquisition card during the entire grinding process. A wafer full inspection instrument equipped with infrared interferometry and spectral confocal probes was used to detect the surface peak-valley deviation PV, curvature Bow, and warp under standard three-point support conditions. The total thickness deviation TTV and local thickness deviation LTV were detected under full-width vacuum plane adsorption support conditions.

[0050] The surface roughness Ra and Rz of the ground workpiece were detected using a white light interferometer; the fisheye diameter and depth, scratch depth and width, crack length, and chipping size of the surface were observed using a confocal microscope and an industrial camera; the thickness of the subsurface damage layer was detected using laser scattering; and the residual stress of the surface layer was analyzed using wafer curvature analysis and Raman spectroscopy.

[0051] The wavelet threshold filtering algorithm sets the wavelet basis function and threshold, and uses the 4th-order Daubechies wavelet basis to carry out multi-level discrete wavelet decomposition with 5 wavelet decomposition levels; it uses SURE unbiased adaptive thresholding, and the threshold of wavelet detail coefficients in each level is determined by the noise standard deviation and length.

[0052] The positive performance indicators include: flexural strength, tensile strength, impact strength, shear spalling strength, pull-out bond strength, fracture strength, wettability, terrain roughness RSa, terrain dispersion RSq, average abrasive grain exit height h, and microscopic profile skewness Rsk; the negative performance indicators include: friction coefficient and stability, temperature rise, material wear rate, grinding wheel dynamic rotational runout, rotational profile, tooth height wear, grinding vibration, grinding zone temperature, increase in grinding wheel dynamic rotational runout, range of grinding wheel rotational profile changes, tooth height wear difference, and terrain roughness. The reduction values ​​of surface roughness RSa, topographic dispersion RSq, average abrasive grain exit height h, micro-profile skewness Rsk, total wafer thickness deviation TTV, local thickness deviation LTV, surface peak-valley deviation PV, curvature Bow, warp, surface roughness Ra and Rz, surface micro-defects, subsurface damage layer thickness, and surface residual stress; neutral performance indicators include: density, porosity, hardness, abrasive concentration, grain size distribution, elastic modulus, force ratio, grinding force, acoustic emission, and spindle load power;

[0053] Set the global maximum value of the key feature index across the entire sample. Global minimum value The feature index to be evaluated, X, and the score of each feature index. Positive performance indicators negative performance indicators Set the optimal range for neutral performance indicators. Full sample limit boundary A score of 100 is obtained within the optimal interval. The score is calculated towards both boundaries using the positive formula and the negative index, until it decreases linearly to 0.

[0054] The key characteristic index attribute scores of the grinding wheel performance in each dimension under the evaluation scenario and the score judgment threshold after evaluation criterion transfer correction are obtained, and the obtained grades are assigned scores: Grade A is excellent with 10 points, Grade B is qualified with 5 points, and Grade C is unqualified with 0 points. The graded scores of all index features in each dimension are calculated by weighted average. The features of the four dimensions of spline / block material performance, grinding wheel body features, grinding process, and wafer post-grinding quality are assigned weights of 20%, 20%, 20%, and 40%, respectively, to obtain the comprehensive performance score of the grinding wheel under the evaluation scenario.

[0055] A full-chain application performance evaluation system for precision grinding wheels includes:

[0056] The data acquisition module provides complete and authentic multi-source heterogeneous data for the entire chain performance evaluation;

[0057] The data processing module is used to standardize and organize multi-source heterogeneous data, remove invalid and interfering data, deeply mine and extract key feature parameters, and construct a dataset of key features of grinding wheel performance.

[0058] The model building and analysis module is used to achieve model parameter fitting and dynamic updating, and to conduct performance correlation inference and trace the cause of grinding wheel performance defects based on the model.

[0059] The evaluation output module is used to calculate the comprehensive performance level according to the established standards, integrate the analysis results, automatically generate an evaluation report, and output optimization and improvement suggestions.

[0060] The human-computer interaction module is used to input the working conditions and basic parameters of the sample, and to visually present the evaluation process, analysis results and performance judgment conclusions.

[0061] Preferably, the data acquisition module includes a spline / nodal testing unit, a grinding wheel body inspection unit, a grinding process monitoring unit, and a wafer quality inspection unit. The spline / nodal testing unit integrates a universal material testing machine, an impact testing machine, a nano-scratch tester, a friction and wear testing machine, a sandblasting hardness tester, a laser particle size analyzer, and a scanning electron microscope to collect intrinsic material parameters such as mechanical properties, interface bonding characteristics, and friction and wear. The grinding wheel body inspection unit integrates a laser triangular displacement sensor and a point spectral sensor to perform in-situ detection of macroscopic feature parameters and microscopic topographic features of the grinding wheel. The grinding process monitoring unit is mounted on a wafer precision grinding test platform and integrates a three-dimensional force gauge, a power meter, an accelerometer, an acoustic emission sensor, and an infrared thermal imager to achieve synchronous, high-frequency, and accurate acquisition of multi-source physical signals under all grinding conditions. The wafer quality inspection unit includes a wafer full inspection instrument, a white light interferometer, a confocal microscope, an industrial camera, a laser scattering detector, and a Raman spectrometer to comprehensively detect quality indicators related to wafer processing accuracy and surface integrity.

[0062] The data processing module includes a data preprocessing unit and a feature extraction unit. The data preprocessing unit completes noise reduction of multiple types of sensor signals, time synchronization of data at different sampling frequencies, and normalization of data across the entire domain. It also automatically removes abnormal interference data based on the 3σ criterion. The feature extraction unit performs targeted feature operations on data of various dimensions, automatically mines and quantifies various core feature parameters, and constructs a standardized and structured dataset of key features of grinding wheel performance.

[0063] The model building and analysis module has a built-in three types of hierarchical performance mapping models with progressive association, which are embedded with machine learning training iterations. Based on the hierarchical mapping relationship, it completes multi-dimensional performance correlation inference and realizes the linkage analysis of grinding wheel material performance, body state, grinding conditions and wafer processing quality.

[0064] The evaluation output module has a built-in data storage unit that pre-stores quantitative evaluation criteria and cross-scenario evaluation criterion migration and correction rules. It is configured to perform comprehensive evaluation calculations and automatically generate reports. It completes weighted calculations based on the score weights of each dimension indicator and integrates information from each dimension to compile evaluation reports and optimization and improvement plans.

[0065] The human-computer interaction module includes an information input terminal, a visualization display component, and an interactive control program. It supports operators in inputting basic information and visually displays the data acquisition process, data processing results, model correlation analysis curves, grinding wheel performance levels, and defect tracing results.

[0066] The beneficial effects of this invention are as follows:

[0067] (1) This invention has for the first time constructed a closed-loop evaluation system for the entire chain from grinding wheel material to wafer quality, established the performance mapping relationship of each link, and can locate the grinding wheel performance defects from the root, avoiding the limitations of single index evaluation and solving the pain point of fragmented evaluation in the industry.

[0068] (2) This invention achieves quantitative characterization of the bonding strength between diamond abrasive grains and binder through single-particle shear test, composite plate pull-out test and cross-sectional morphology analysis, and solves the problem that the interface performance cannot be quantitatively evaluated in the industry.

[0069] (3) The present invention sets up instruments such as a three-dimensional force meter, power meter, acceleration sensor, acoustic emission sensor, and infrared thermal imager on the wafer grinding test platform. Through the fusion of multi-sensor signals and signal feature recognition during the grinding process, the grinding conditions and workpiece quality are monitored in real time, which can provide timely warnings and avoid batch grinding quality defects.

[0070] (4) By setting cross-scenario evaluation criteria and transfer rules, this invention realizes universal evaluation under different grinding wheel types and different wafer materials, establishes transferable evaluation standards, and solves the problem of inconsistent industry evaluation standards.

[0071] (5) The analysis results of the mapping model can provide a scientific basis for the optimization of grinding wheel formula, structural design and grinding process adjustment, output optimization guidance, support process improvement and help improve grinding wheel life and wafer processing yield. Attached Figure Description

[0072] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0073] Figure 1 This is a schematic diagram of the structure of the present invention. Detailed Implementation

[0074] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0075] Example 1

[0076] A method for evaluating the performance of a precision grinding wheel's entire chain application, such as... Figure 1 As shown, the steps include:

[0077] S1: Collect grinding wheel performance-related data from four dimensions: spline / block material properties, grinding wheel body characteristics, grinding process, and wafer post-grinding quality, and establish a multi-source heterogeneous performance testing database in four dimensions.

[0078] In this embodiment, the performance dimensions of the spline / block material include basic physicochemical properties, mechanical properties, microstructure characteristics, interfacial bonding strength, and tribological properties. Specifically:

[0079] The basic physicochemical properties include: density, porosity, hardness, abrasive concentration, and particle size distribution. Density and porosity were tested using the Archimedes water displacement method, hardness was tested using a sandblasting hardness tester, and abrasive concentration and particle size were tested using a laser particle size analyzer.

[0080] Mechanical performance parameters include: flexural strength, tensile strength, impact strength, and modulus of elasticity. Flexural strength, tensile strength, and impact strength are tested using a universal testing machine and an impact testing machine, while modulus of elasticity is tested using the natural frequency method of impact vibration test.

[0081] The microstructure features include: exposed abrasive grains, uniform distribution, breakage and fragmentation, and interfacial porosity; all were detected by micromorphological observation using a scanning electron microscope with an energy dispersive spectroscopy (EDS) instrument, and auxiliary verification images were obtained for qualitative identification of the microstructure features.

[0082] The interfacial bonding strength indicators include: shear spalling strength, pull-out bonding strength, interfacial fracture strength, and interfacial wettability. A linearly increasing transverse shear force is applied to a single abrasive grain using a diamond indenter of a nano-scratch instrument, and the critical load at the time of grain detachment is recorded as the shear spalling strength (single-particle shear test). A normal pull-out test is performed on a composite plate sample of diamond sheet and binder sintered together. The pull-out force is linearly increased until the interface between the diamond sheet and binder separates. The detached section is observed and analyzed under a microscope. If the detached section shows complete separation of the diamond sheet and binder, the pull-out force is recorded as the pull-out bonding strength. If binder residue remains at the detached section, indicating internal fracture of the binder, the pull-out force is recorded as the interfacial fracture strength (composite plate pull-out test). The interfacial wettability is evaluated by observing the wetting angle of the binder on the diamond surface after the binder has melted.

[0083] Friction and wear performance parameters include: force ratio (the ratio of frictional force to pressure), coefficient of friction and stability, material wear rate, and temperature rise. Friction and wear testing machines are used to detect these performance indicators, including frictional force, pressure, coefficient of friction, material wear rate, stability, and temperature rise.

[0084] In this embodiment, the feature dimensions of the grinding wheel body include macroscopic and microscopic feature parameters of the surface before grinding. Specifically:

[0085] Macroscopic characteristic parameters include: dynamic rotational runout of the grinding wheel, rotational profile, and tooth height wear. A laser triangular displacement sensor is used to collect global macroscopic morphological fluctuation data of the grinding wheel's surface during rotation. The collected data is processed using algorithms such as periodic filtering, extraction of outer envelope feature points, and spline fitting. The fitted curve is converted to polar coordinates to represent the grinding wheel's rotational profile. The height difference between the highest and lowest points in the rotational profile is calculated as the dynamic rotational runout of the grinding wheel. The difference between the mean values ​​of all data points in the rotational profile before and after grinding wheel wear is the tooth height wear. This achieves in-situ detection of the macroscopic characteristic parameters of the grinding wheel body.

[0086] The microscopic feature parameters include: terrain roughness RSa, terrain dispersion RSq, average abrasive grain tip height h, and microscopic profile skewness Rsk. A point spectral sensor is used to collect high-frequency in-situ scanning data of the microscopic terrain profile of the grinding wheel surface. Algorithms are applied to the collected data, including sliding window threshold filtering, profile median line extraction, peak and valley feature point extraction, and feature parameter calculation, to obtain the microscopic feature parameter values ​​and achieve in-situ detection of the microscopic feature parameters of the grinding wheel body.

[0087] The formula for calculating terrain roughness RSa is: ;

[0088] The formula for calculating the topographic dispersion RSq is: ;

[0089] The formula for calculating the micro-profile skewness Rsk is: ;

[0090] The formula for calculating the average abrasive grain exit height h is: ;

[0091] In the above formula: n represents the total number of data points collected for the micro-topographic contour of the entire surface of the grinding wheel. Let i be the measured value of the i-th collection point. Let be the value on the median line of the contour corresponding to the i-th sampling point, and t be the total number of data points of the contour peak data of the micro-topography of the entire surface of the grinding wheel. Let j be the measured value of the j-th peak. is the value on the midline of the contour corresponding to the j-th peak point.

[0092] In this embodiment, the grinding process dimension includes dynamic signals of the grinding process and the evolution of grinding wheel body features. Specifically:

[0093] The dynamic signals of the grinding process include: grinding force, spindle load power, grinding vibration, acoustic emission, and temperature of the grinding zone under all working conditions. Instruments such as a three-dimensional force gauge, power meter, accelerometer, acoustic emission sensor, and infrared thermal imager are set up in the grinding area of ​​the grinding test platform. The force, power, vibration, acoustic emission, and temperature signals of the entire grinding process are acquired using a data acquisition card to realize real-time online monitoring of the grinding conditions.

[0094] The evolution of grinding wheel characteristics includes: the degree of change of macroscopic and microscopic characteristic parameters of the grinding wheel detected in situ during the grinding process; the increase in dynamic rotational runout of the grinding wheel to reflect the grinding stability of the grinding wheel; the range of change in the rotational profile of the grinding wheel to reflect the degree of uneven wear of the grinding wheel; the difference in tooth height wear to reflect the wear ratio of the grinding wheel; and the decrease in terrain roughness RSa, terrain dispersion RSq, average abrasive grain exit height h, and microscopic profile deviation Rsk to reflect the degree of grinding wheel passivation and clogging.

[0095] In this embodiment, the quality dimensions of the wafer after grinding include precision and surface integrity. Specifically:

[0096] The accuracy includes grinding accuracy indicators such as total wafer thickness deviation (TTV), local thickness deviation (LTV), surface peak-valley deviation (PV), bend (Bow), and warp (Warp). A full wafer inspection instrument equipped with infrared interferometry and spectral confocal probes is used to detect surface peak-valley deviation (PV), bend (Bow), and warp under standard three-point support conditions, and to detect total thickness deviation (TTV) and local thickness deviation (LTV) under full-area vacuum plane adsorption support conditions.

[0097] Surface integrity includes surface roughness (Ra, Rz), surface micro-defects (fisheye, scratches, cracks, chipping), subsurface damage layer thickness, and surface residual stress as surface integrity indices. The surface roughness values ​​Ra and Rz of the ground workpiece are detected using a white light interferometer. The diameter and depth of fisheyes, the depth and width of scratches, the length of cracks, and the size of chipping are observed using a confocal microscope and an industrial camera. The thickness of the subsurface damage layer is detected using laser scattering. Surface residual stress is analyzed using wafer curvature analysis and Raman spectroscopy.

[0098] S2: Preprocess and perform key feature analysis on multi-source heterogeneous performance test data to construct a four-dimensional dataset of key features of grinding wheel performance.

[0099] In this embodiment of the application, the preprocessing and key feature analysis of multi-source heterogeneous performance testing data includes:

[0100] S21: Wavelet threshold filtering algorithm is used to filter and reduce noise for the dynamic signal of the grinding process. By setting the wavelet basis function and threshold parameters, the noise signal caused by environmental electromagnetic interference, equipment operating noise and sensor error is effectively filtered out, which significantly improves the signal-to-noise ratio of the collected data and ensures the authenticity and effectiveness of the dynamic signal.

[0101] In this embodiment, a 4th-order Daubechies wavelet basis is used for multi-level discrete wavelet decomposition, with a fixed number of 5 layers. This accurately matches the time-frequency characteristics of non-stationary grinding signals, fully preserving the transient impact characteristics corresponding to abrasive grain shedding, grinding wheel passivation, and wafer scratches, while effectively separating broadband random noise. Specifically, this embodiment uses a SURE unbiased adaptive threshold, with the threshold calculation formula for each wavelet detail coefficient being: In the formula: Let be the noise standard deviation of the wavelet detail coefficients of the i-th layer. Let be the length of the wavelet detail coefficients of the i-th layer.

[0102] Meanwhile, the timestamp synchronization method is used to time-align the dynamic signal of the grinding process after filtering and noise reduction with the data on the evolution of the grinding wheel body characteristics in the stage interval. These different types of data with different sampling frequencies are time-series aligned. By calibrating the time base of each detection instrument, the time period of all multi-source data is accurately synchronized and matched, ensuring the temporal correlation between data and providing data support with a unified time dimension for subsequent correlation analysis.

[0103] S22: For all index data in the dimensions of spline / block material performance, grinding wheel body characteristics, and wafer post-grinding quality, following the 3σ error judgment criterion, by calculating the mean and standard deviation of the test data of each dimension index, anomaly judgment thresholds are set, and invalid abnormal data caused by factors such as instrument system error, experimental operation deviation, environmental temperature and humidity interference, and equipment vibration interference are batched out; at the same time, the dataset after removing abnormal data is completed using linear interpolation to ensure the integrity of the dataset and comprehensively guarantee the accuracy, effectiveness, and overall reliability of the multi-dimensional evaluation data.

[0104] S23: Conduct systematic feature engineering and use feature extraction algorithms to mine and extract key features from the four dimensions of data.

[0105] Specifically, for all index data related to the basic physicochemical indices, mechanical property parameters, interfacial bonding strength indices, and friction and wear performance parameters of the spline / block material performance dimension after step S22, correlation feature screening is carried out, the Pearson correlation coefficient between features is calculated, and strongly collinear redundant indices with a threshold |r|≥0.85 are removed; then, the min-max normalization method is used to uniformly map all remaining features to the [0,1] interval to eliminate dimensional interference; finally, the normalized features are extracted by PCA principal component analysis to obtain the principal component index set that integrates the comprehensive features of all material performance data with a cumulative variance contribution rate of not less than 95%, thus completing the dimensionality reduction processing of high-dimensional feature data. As input or output of subsequent models, This represents the principal component index variables of the calculated material properties dimension, avoiding problems such as model overfitting and slow training convergence.

[0106] Specifically, for the macroscopic and microscopic feature parameters of the grinding wheel body feature dimension processed in step S22, correlation feature screening is also performed, and the threshold for eliminating collinearity of indicators |r|≥0.9 is appropriately relaxed to eliminate highly overlapping and redundant features; then, the remaining grinding wheel body features are uniformly mapped to the [0,1] interval through normalization. Since the number of indicators of the grinding wheel body feature dimension is small and the correlation between indicators is weak, the normalized feature indicator set is directly extracted. As input or output of subsequent models. This represents the index variables after normalization of each indicator in the feature dimension of the grinding wheel body.

[0107] Specifically, the dynamic signal of the grinding process after step S21 is segmented according to the time-aligned stages. For each stage signal, time-frequency joint feature analysis is used to extract the time-domain features and wavelet time-frequency energy features of the signal. The continuous waveform signal is transformed into fixed-dimensional quantized features to obtain the peak and mean grinding force, the mean spindle grinding power, the grinding vibration amplitude and wavelet energy ratio, the acoustic emission RMS and wavelet energy ratio, and the peak and mean temperature of the grinding zone for each stage. The Pearson correlation coefficient between the signal features of each stage is calculated to carry out correlation screening. Strong collinear redundant features with a threshold |r|≥0.85 are removed. The remaining features are mapped to the [0,1] interval through min-max normalization to eliminate dimensional differences. Then, PCA principal component analysis is added to extract the comprehensive feature principal component index set C, D, E, ... with a cumulative variance contribution rate of not less than 95% for each stage, which is used as the input or output of the subsequent model.

[0108] Then, for the grinding wheel body feature evolution index data processed in step S21, the Pearson correlation coefficient between the signal features of each stage is calculated according to the same time-series alignment stage to carry out correlation screening. Collinear redundant features with a threshold |r|≥0.9 are removed. The remaining features are mapped to the [0,1] interval through min-max normalization to eliminate dimensional differences. The normalized feature index set J, K, L, ... of each stage is extracted as the input or output of the subsequent model.

[0109] Specifically, for all the index data of precision and surface integrity of the wafer after grinding in step S22, correlation feature screening is also carried out, the Pearson correlation coefficient between features is calculated, and strongly collinear redundant indicators with a threshold |r|≥0.85 are removed; the min-max normalization method is used to uniformly map all remaining features to the [0,1] interval; the normalized features are then subjected to PCA principal component analysis to extract the principal component index set with a cumulative variance contribution rate of not less than 95%. , which serves as the input or output of subsequent models. These represent the principal component index variables of the calculated wafer post-grinding quality dimension.

[0110] S3: Construct a hierarchical performance mapping model from material properties to wafer processing quality, and train the hierarchical performance mapping model using a dataset of key features of grinding wheel performance in four dimensions to obtain a trained hierarchical performance mapping model.

[0111] In this embodiment of the application, the hierarchical performance mapping model from material properties to wafer processing quality includes a material-bulk state mapping model, a bulk state-grinding condition mapping model, and a grinding condition-wafer quality mapping model.

[0112] Specifically, the material-bulk state mapping model is set as a first-level model, using the principal component index set A of the spline / block material performance dimension in step S23 as the model input variable, and the feature index set of the grinding wheel bulk feature dimension in step S23 as the input variable. The model uses the characteristic index set J, K, L, ... of each stage of the grinding wheel body feature evolution in the grinding process dimension of S23 as the model output variables. It employs multivariate nonlinear regression analysis combined with gradient boosting machine learning algorithm to train the model, fitting quantitative constraint formulas for the degradation rate and degree of the grinding wheel body state of static material parameters. This accurately establishes a quantitative relationship between material properties and changes in the surface state of the grinding wheel body, allowing for positive prediction of the actual impact of different material formulations on the grinding wheel's performance, wear resistance, and anti-clogging properties, or for tracing back to the root cause of grinding wheel body state degradation. This model is a single-input, multi-output fusion analysis model. The multivariate nonlinear regression + gradient boosting machine learning (GBDT) algorithm adapts to the strong coupling and non-monotonic changes of multiple stages and variables in grinding, and can output complete grinding wheel body service state parameters at once, resulting in high inference efficiency. Simultaneously, it can provide input feature importance, supporting the full-link defect tracing from grinding anomalies to the root cause of grinding wheel body degradation and relocating the underlying material performance shortcomings.

[0113] Specifically, the body state-grinding condition mapping model is set as a two-level model, using the feature index set of the grinding wheel body feature dimension in step S23. The model input variables are the feature index set J, K, L, ... of each stage of the grinding wheel body feature evolution in step S23 (grinding process dimension). The model output variables are the principal component index set C, D, E, ... of each stage of the grinding process dynamic signal in step S23 (grinding process dimension). The model is built and trained using multivariate nonlinear regression analysis combined with gradient boosting machine learning algorithm and time-series correlation fitting method. After training, the multi-input multi-output mapping function is solidified, establishing an intrinsic correspondence between the surface state changes of the grinding wheel body and the characteristics of the dynamic working condition signal in the field grinding. This allows for rapid inference of the current grinding wheel body passivation and wear degree based on the dynamic signal during grinding, enabling real-time online accurate identification and status determination of abnormal working conditions such as grinding wheel passivation, wear, and performance degradation without downtime.

[0114] Specifically, the grinding condition-wafer quality mapping model is set as a three-level model. The principal component index set C, D, E, ... of each stage of the grinding process dynamic signal in step S23 (grinding process dimension) is used as the model input variable, and the principal component index set of the wafer post-grinding quality dimension in S23 is used as the input variable. As the output variable of the model, multivariate nonlinear regression analysis combined with gradient boosting machine learning algorithm is used to complete the model iterative training, establish a quantitative prediction function between the dynamic signal of the grinding process and the wafer grinding quality index, and form a stable and reliable intelligent prediction mechanism for wafer processing quality. It can predict the trend of finished product processing quality in advance based on real-time dynamic signal data of the grinding process, analyze the matching degree between process parameters and current grinding conditions, and avoid batch quality defects.

[0115] S4: Establish quantitative evaluation criteria for four-dimensional characteristic indicators.

[0116] In this embodiment, a quantitative evaluation criterion is established covering four dimensions: spline / block material properties, grinding wheel body characteristics, grinding process, and wafer post-grinding quality. The evaluation processing object is the standardized key features of each dimension extracted in step S23, including two categories: grinding wheel body morphology features and process evolution indicators that have only undergone min-max normalization processing; and comprehensive principal component indices of material properties, grinding dynamic signals, and wafer post-grinding quality obtained by normalization and PCA principal component dimensionality reduction.

[0117] First, the attributes of all evaluation features are defined as positive performance indicators (higher values ​​indicate better performance), negative performance indicators (higher values ​​indicate worse performance), and neutral performance indicators (optimal performance only within a preset reasonable range). The judgment rules are as follows:

[0118] The first step is to pre-distinguish the attributes of the underlying original indicators based on the process physical meaning of the original measured indicators. The attributes of the original indicators are divided as follows:

[0119] Positive performance indicators include: flexural strength, tensile strength, impact strength, shear spalling strength, pull-out bond strength, fracture strength, wettability, terrain roughness RSa, terrain dispersion RSq, average abrasive tip height h, and micro profile skewness Rsk.

[0120] Negative performance indicators include: friction coefficient and stability, temperature rise, material wear rate, dynamic rotational runout of grinding wheel, rotational profile, tooth height wear, grinding vibration, grinding zone temperature, increase in dynamic rotational runout of grinding wheel, range of change in rotational profile of grinding wheel, difference in tooth height wear, decrease in topographic roughness RSa, decrease in topographic dispersion RSq, decrease in average abrasive grain tip height h, decrease in microscopic profile skewness Rsk, total wafer thickness deviation TTV, local thickness deviation LTV, surface peak-valley deviation PV, curvature Bow, warp Warp, roughness (Ra, Rz), surface micro-defects (fisheye, scratches, cracks, chipping), subsurface damage layer thickness, and surface residual stress.

[0121] Neutral performance indicators include: density, porosity, hardness, abrasive concentration, particle size distribution, elastic modulus, force ratio, grinding force, acoustic emission, and spindle load power.

[0122] The second step is to address the indicators of the physical characteristics and process evolution of the grinding wheel that are normalized without PCA dimensionality reduction. Since normalization does not change the trend of the indicator's quality, the positive, negative, or neutral attributes corresponding to the original indicator are directly inherited.

[0123] The third step involves extracting the corresponding eigenvector weight matrix for the PCA comprehensive principal component index, selecting the normalized original index with the highest absolute weight value as the dominant term, and defining the overall category of the principal component based on the original index attributes of the dominant term: if the dominant term is a positive index, then the principal component is a positive comprehensive performance index; if the dominant term is a negative index, then the principal component is a negative comprehensive performance index; if the dominant term is a neutral index, then the principal component is a neutral comprehensive performance index.

[0124] Then, a unified scoring system of 0-100 points is used to quantify the individual indicator attributes, with higher scores representing better performance. Let the global maximum value of a key feature indicator be set across the entire sample. Global minimum value The feature index to be evaluated, X, and the score of each feature index. Regarding positive indicators: Regarding negative indicators: For neutral indicators, set the optimal range. Full sample limit boundary A score of 100 is awarded within the interval. The score is then calculated towards both boundaries using the positive and negative index formulas, until it decreases linearly to 0.

[0125] Finally, two levels of scoring thresholds are independently calibrated for all features: a qualified lower limit threshold T1 and a high-quality lower limit threshold T2. Based on the individual attribute score, three performance level intervals are divided: a single attribute score ≥ T2 is judged as Grade A excellent; T1 ≤ single attribute score < T2 is judged as Grade B qualified; and a single attribute score < T1 is judged as Grade C unqualified.

[0126] By standardizing and converting all key features across the four dimensions and calibrating grading thresholds, a comprehensive, standardized, and quantitatively controllable performance benchmark evaluation system for the entire grinding wheel chain is formed. Furthermore, by combining the performance results from the multi-level mapping model in step S3, and after obtaining the principal component indices of spline / block material performance, these indices are used as input constraints to forward solve the feature indices of the grinding wheel body feature dimension and the feature indices of each stage of the grinding wheel body feature evolution in the grinding process dimension through a first-level model. Then, using the feature indices of the grinding wheel body feature dimension and the feature indices of each stage of the grinding wheel body feature evolution in the grinding process dimension as input constraints, the principal component indices of each stage of the grinding process dynamic signal in the grinding process dimension are forward solved through a second-level model. Finally, using the principal component indices of each stage of the grinding process dynamic signal as input constraints, the principal component indices of the wafer post-grinding quality dimension are forward solved through a third-level model. Through the joint layer-by-layer calculation of the hierarchical performance mapping model, the lower-level related performance indices are automatically predicted, thereby pre-quantifying and determining the quality level of the grinding wheel material performance, body morphology characteristics, grinding process conditions, and post-grinding wafer quality, and pre-identifying performance shortcomings.

[0127] S5: Different correction coefficients are set for different processing scenarios, and cross-scenario evaluation criterion transfer correction rules are constructed. The scoring judgment threshold is dynamically adjusted according to the correction coefficient of the corresponding scenario.

[0128] In this embodiment, a cross-scenario evaluation criterion migration and correction rule is constructed to enable the evaluation standard to adaptively adapt to various processing conditions. Multiple grinding test machines are selected to conduct rough grinding, semi-fine grinding, and fine grinding process tests, covering various wafer substrates such as silicon wafers, silicon carbide, gallium nitride, and sapphire, to obtain data index samples under multiple scenarios.

[0129] First, we set standard equipment, semi-finished grinding, and silicon wafers as the benchmark scenarios, and corresponding basic acceptable lower threshold values. Basic high-quality lower limit threshold For different machine tools, semi-finished grinding, and silicon wafer processing scenarios, the ratio of the average score of key feature index attributes for each dimension to the average score of key feature index attributes for the corresponding dimensions in the benchmark scenario is taken as the machine tool system correction coefficient K. M For standard equipment, different processes, and silicon wafer processing scenarios, the ratio of the average score of key feature indicators for each dimension to the average score of key feature indicators for each dimension in the benchmark scenario is taken as the process correction coefficient K. P For processing scenarios involving standard equipment, semi-finished grinding, and different wafer materials, the ratio of the average score of key feature indicators for each dimension to the average score of key feature indicators for each dimension in the benchmark scenario is taken as the material correction coefficient K. w For multi-factor scene changes, the corresponding correction coefficients are multiplied to obtain the overall scene correction coefficient.

[0130] Then, the scoring threshold is dynamically adjusted based on the adjustment factor K for the corresponding scenario. The calculation formula is as follows: , For precision grinding applications involving high-hardness, high-brittleness, and wide-bandgap semiconductor wafers such as silicon carbide and gallium nitride, the material correction factor K... w >1, process correction factor K P >1, the comprehensive correction system is K w •K P If the value is greater than 1, the scoring threshold is increased after amplification by a comprehensive correction coefficient, raising the standard requirements for "qualified" and "high-quality". Based on a unified basic evaluation threshold system, the scoring threshold is adaptively scaled using a triple correction coefficient to adapt to diverse processing scenarios involving multiple machines, processes, and wafer materials, forming a transferable evaluation criterion correction mechanism that balances the universality of the evaluation system with the specificity of different working conditions.

[0131] S6: Based on quantitative evaluation criteria and cross-scenario evaluation criterion transfer correction rules, and according to the key characteristic data of grinding wheel performance in each dimension of the evaluation scenario, the performance of the wafer thinning grinding wheel in the evaluation scenario is comprehensively evaluated. Based on the comprehensive evaluation results, problem tracing and process optimization are performed. The specific implementation method is as follows:

[0132] S61. Assign scores to the key characteristic indicators of grinding wheel performance in each dimension obtained in step S4, and to the resulting grades after threshold shifting and correction in step S5: Grade A (excellent) receives 10 points, Grade B (qualified) receives 5 points, and Grade C (unqualified) receives 0 points. A weighted average is calculated for the graded scores of all features in each dimension. The features of the four dimensions—spline / block material properties, grinding wheel body characteristics, grinding process, and wafer post-grinding quality—are assigned weights of 20%, 20%, 20%, and 40%, respectively. Finally, the comprehensive performance score of the grinding wheel in the evaluated scenario is obtained, achieving a standardized and quantitative evaluation of the grinding wheel's comprehensive grinding capability, service reliability, and process adaptability.

[0133] S62. Trace the source of performance non-compliance items, abnormal operating conditions items, or wafer quality defects that occur during the evaluation process. When the wafer quality fails to meet the standards after grinding, firstly, use the three-level model in step S3 to solve inversely to locate the grinding dynamic signal principal component index set C, D, E, ... that causes the quality non-compliance. Use the pre-stored PCA feature vector weight matrix to perform inverse projection transformation on the located grinding principal component index set to obtain the normalized original index features. Then, use Min-Max inverse normalization to convert and restore the actual grinding dynamic signal data with true dimensions to complete the location of grinding operating condition abnormalities.

[0134] Then, input the abnormal grinding principal component indexes located above into the secondary model in step S3 for reverse deduction, and solve the normalized characteristic index of the grinding wheel body matching the working condition. The grinding wheel body characteristics can be restored by Min-Max inverse normalization to obtain the actual measured index of the grinding wheel surface morphology, and lock the deterioration state of the grinding wheel body such as passivation, wear, and blockage.

[0135] Finally, using the actual morphological indicators of the grinding wheel body as the output constraint, the principal component indicators of the corresponding spline / nodal material properties are located by solving the first-level model in step S3. PCA inverse projection and inverse normalization are performed on the located principal component indicators to restore the original material's inherent measured index data. In this way, the entire link of "wafer quality defects - abnormal grinding conditions - grinding wheel body deterioration - material performance shortcomings" is located layer by layer. Furthermore, the qualitative identification images of abrasive grain exposure, uniform distribution, breakage, and interface porosity detected by microstructure feature detection are used for auxiliary analysis and verification to accurately pinpoint the root cause of grinding quality problems. For example, scratches and burns on the wafer surface were located and analyzed using a multi-level performance mapping model. This revealed abnormal stress and vibration during grinding, which led to the identification of deterioration in the morphology of the grinding wheel itself. Ultimately, the source was determined to be insufficient material interface bonding strength, which made the grinding process prone to abrasive grain shedding. Qualitative analysis of the microstructure of the grinding wheel surface revealed numerous pits of abrasive grain shedding on the surface after grinding, as well as obvious multiple areas of chip blockage, which further verified the conclusion that the material performance was insufficient.

[0136] The feature vector weights and global maximum and minimum values ​​of each dimension required for PCA inverse projection and normalization inverse conversion are all stored uniformly in step S3 during the model training phase, providing a fixed conversion basis for the reverse reconstruction of principal components and normalization indicators into actual engineering indicators.

[0137] S63: Based on the problem tracing results and model correlation patterns, targeted improvement schemes are proposed for optimizing grinding wheel material formulations, grinding wheel structural parameters, dressing processes, and grinding process matching. For insufficient inherent material properties, formulation optimization suggestions are proposed, including abrasive concentration, binder hardness, and interface modification processes. For structural problems such as grinding wheel surface clogging and uneven wear, optimization schemes are proposed for grinding wheel dressing and forming process control. For unreasonable matching of grinding process parameters, matching and adjustment schemes for process parameters such as grinding feed rate, grinding speed, and cooling conditions are proposed. Ultimately, a closed-loop output of grinding wheel performance evaluation, problem tracing, and process optimization is achieved, providing accurate and implementable technical guidance for grinding wheel R&D iteration, factory performance judgment, and on-site grinding process control.

[0138] Example 2

[0139] To adapt to the aforementioned full-chain grinding application performance evaluation method for wafer thinning grinding wheels, a full-chain application performance evaluation system for precision grinding wheels is proposed through a modular and integrated hardware and software collaborative architecture. This system enables full-dimensional data acquisition, intelligent processing, model analysis, grade evaluation, and result visualization output, and includes the following core modules:

[0140] The data acquisition module comprises four subdivided testing units. The spline / block testing unit integrates various instruments and equipment, including a universal testing machine, impact testing machine, nano-scratch tester, tribometer, sandblasting hardness tester, laser particle size analyzer, and scanning electron microscope, for precise acquisition of intrinsic material parameters such as mechanical properties, interfacial bonding characteristics, and tribological wear. The grinding wheel body testing unit integrates laser triangulation sensors and point spectral sensors for in-situ detection of macroscopic characteristic parameters and microscopic topographic features of the grinding wheel. The grinding process monitoring unit, mounted on a dedicated wafer precision grinding test platform, integrates a triaxial force gauge, power meter, accelerometer, acoustic emission sensor, and infrared thermal imager, enabling synchronous, high-frequency, and precise acquisition of multi-source physical signals across all grinding conditions. The wafer quality inspection unit integrates a wafer full inspection instrument, white light interferometer, confocal microscope, industrial camera, laser scattering detector, and Raman spectrometer for comprehensive testing of wafer processing accuracy and surface integrity-related quality indicators. The data acquisition module provides complete and authentic raw data support for the performance evaluation of the entire chain.

[0141] The data processing module, comprised of software, includes data denoising, time-series synchronization, anomaly detection, and feature extraction algorithms. It comprises a data preprocessing unit and a feature extraction unit. The data preprocessing unit performs denoising on various types of sensor signals, time synchronization of data at different sampling frequencies, and global data normalization. It also automatically removes abnormal and interfering data based on the 3σ criterion, ensuring the validity and consistency of the original data. The feature extraction unit performs targeted feature operations on data across various dimensions, automatically mining and quantifying various core feature parameters to construct a standardized and structured model training feature dataset, providing high-quality data support for subsequent model fitting and correlation analysis. The data processing module is used to standardize and organize multi-source heterogeneous data, remove invalid and interfering data, deeply mine and extract key feature parameters, and construct a standardized feature dataset.

[0142] The model building and analysis module, comprised of software, is the core intelligent computing unit of the system. It incorporates three types of hierarchically related performance mapping models and embeds a machine learning training iteration program. It can import pre-processed standardized feature datasets and, based on the hierarchical mapping relationships, perform multi-dimensional performance correlation inferences, enabling linked analysis of grinding wheel material properties, body condition, grinding conditions, and wafer processing quality. The model building and analysis module is used to achieve model parameter fitting and dynamic updates, conduct performance correlation inferences based on the model, and trace the root causes of grinding wheel performance defects.

[0143] Evaluation Output Module: Composed of hardware and software, this module includes a built-in data storage unit, pre-stored quantitative evaluation criteria and scenario adaptation rules, and is configured with a comprehensive evaluation calculation program and an automatic report generation program. It can perform weighted comprehensive calculations based on the scores and weights of each dimension's indicators, integrate information from all dimensions to compile an evaluation report and optimization / improvement plans. The evaluation output module is used to calculate the comprehensive performance level according to predetermined standards, integrate the analysis results, automatically generate an evaluation report, and output optimization / improvement suggestions.

[0144] The human-computer interaction module, composed of hardware and software, includes an information input terminal, a visualization display component, and an interactive control program. It supports operators in inputting basic information such as grinding wheel model, wafer material, grinding process parameters, and test conditions. Simultaneously, it can visually display the data acquisition process, data processing results, model correlation analysis curves, grinding wheel performance grade, and defect tracing results. The human-computer interaction module is used to input basic parameters of the working conditions and samples, and visually presents the evaluation process, analysis results, and performance judgment conclusions, facilitating operators' real-time monitoring of the grinding wheel's grinding performance status and process compatibility.

[0145] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for evaluating the full-chain application performance of precision grinding wheels, characterized in that, Includes the following steps: S1: Collect data related to grinding wheel performance from four dimensions: spline / block material properties, grinding wheel body characteristics, grinding process, and wafer post-grinding quality, and establish a multi-source heterogeneous performance testing database in four dimensions; S2: Preprocess and perform key feature analysis on multi-source heterogeneous performance test data to construct a four-dimensional dataset of key features of grinding wheel performance; S3: Construct a hierarchical performance mapping model from material properties to wafer processing quality, and train the hierarchical performance mapping model using a dataset of key features of grinding wheel performance to obtain a trained hierarchical performance mapping model. S4: Establish quantitative evaluation criteria for characteristic indicators covering four dimensions: spline / block material properties, grinding wheel body characteristics, grinding process, and wafer post-grinding quality; S5: Construct cross-scenario evaluation criterion transfer and correction rules: set different correction coefficients for different processing scenarios, and dynamically adjust the scoring threshold according to the correction coefficient of the corresponding scenario; S6: Based on quantitative evaluation criteria and cross-scenario evaluation criterion transfer correction rules, and according to the key feature data of grinding wheel performance in each dimension of the evaluation scenario, the performance of precision grinding wheel in the evaluation scenario is comprehensively evaluated. Based on the comprehensive evaluation results, the trained hierarchical performance mapping model is used to trace the source of problems and optimize the process.

2. The method for evaluating the full-chain application performance of precision grinding wheels according to claim 1, characterized in that, The performance dimensions of the spline / block material include basic physicochemical properties, mechanical property parameters, microstructure characteristics, interfacial bonding strength, and tribological properties. The characteristic dimensions of the grinding wheel body include the macroscopic and microscopic characteristic parameters of the surface before grinding; The grinding process dimension includes the dynamic signals of the grinding process and the evolution of the grinding wheel body characteristics; The quality dimensions of the wafer after grinding include precision and surface integrity.

3. The method for evaluating the full-chain application performance of precision grinding wheels according to claim 2, characterized in that, The basic physicochemical properties include: density, porosity, hardness, abrasive concentration, and particle size distribution; the mechanical properties include: flexural strength, tensile strength, impact strength, and elastic modulus; the microstructure characteristics include: exposed abrasive grains, uniform distribution, breakage, and interfacial porosity; the interfacial bonding strength indicators include: shear strength, pull-out strength, interfacial fracture strength, and interfacial wettability; the tribological properties include: force ratio, coefficient of friction and stability, material wear rate, and temperature rise. The macroscopic characteristic parameters include: dynamic rotational runout of the grinding wheel, rotational profile, and tooth height wear; the microscopic characteristic parameters include: terrain roughness RSa, terrain dispersion RSq, average abrasive grain exit height h, and microscopic profile skewness Rsk. The dynamic signals of the grinding process include: grinding force, spindle load power, grinding vibration, acoustic emission, and temperature of the grinding zone under all working conditions; the evolution of the grinding wheel body characteristics includes: the increase value of dynamic rotational runout of the grinding wheel, the range of changes in the grinding wheel rotational profile, the difference in tooth height wear, and the decrease values ​​of topographic roughness RSa, topographic dispersion RSq, average abrasive grain exit height h, and microscopic profile skewness Rsk. The accuracy includes: total wafer thickness deviation (TTV), local thickness deviation (LTV), surface shape peak-valley deviation (PV), curvature (Bow), and warp (Warp); the surface integrity includes: surface roughness (Ra), surface micro-defects, subsurface damage layer thickness, and surface residual stress; surface micro-defects include fisheyes, scratches, cracks, and chipping.

4. The method for evaluating the full-chain application performance of precision grinding wheels according to claim 2 or 3, characterized in that, The preprocessing includes: applying a wavelet threshold filtering algorithm to the dynamic signal of the grinding process in the grinding process dimension for filtering and noise reduction; and using a timestamp synchronization method to time-align the dynamic signal of the grinding process after filtering and noise reduction with the data of the evolution of the grinding wheel body features in the stage interval. For all index data of spline / block material performance dimension, grinding wheel body characteristic dimension, and wafer grinding quality dimension, the mean and standard deviation of the test data of each index are calculated according to the 3σ error judgment criterion, an outlier judgment threshold is set, and outlier data is removed in batches; the dataset after removing outlier data is completed by linear interpolation. The key feature analysis employs feature extraction algorithms to mine and extract key features from the preprocessed data across four dimensions, including: Correlation feature screening was performed on all index data related to the basic physicochemical indicators, mechanical property parameters, interfacial bonding strength indicators, and tribological wear performance parameters of the preprocessed spline / block material properties. Strong collinear redundant indicators with correlation coefficients ≥0.85 were removed. The min-max normalization method was used to uniformly map all remaining indicators to [0,1] to eliminate dimensional interference. The normalized features were then extracted by PCA principal component analysis to obtain the comprehensive features of all material properties data with a cumulative variance contribution rate of not less than 95%, resulting in the principal component index set A. The macroscopic and microscopic feature parameters of the preprocessed grinding wheel body feature dimensions are subjected to correlation feature screening to remove the indicator data with correlation coefficient ≥ 0.9; the remaining grinding wheel indicator data are uniformly mapped to the [0,1] interval by normalization to obtain the feature indicator set B. The preprocessed dynamic signal of the grinding process was segmented according to time sequence alignment. For each stage of the signal, the time domain features and wavelet time-frequency energy features were extracted by joint time-frequency domain feature analysis. The peak and mean values ​​of grinding force, the mean value of spindle grinding power, the ratio of grinding vibration amplitude to wavelet energy, the ratio of acoustic emission RMS to wavelet energy, and the peak and mean values ​​of grinding zone temperature were obtained for each stage. The correlation coefficient between the signal features of each stage was calculated to screen for correlation features. Strong collinearity and redundant features with correlation coefficient ≥ 0.85 were removed. The remaining features were mapped to the [0,1] interval by minimum-maximum normalization. Then, the comprehensive features with a cumulative variance contribution rate of not less than 95% for each stage were extracted by PCA principal component analysis to obtain the principal component index set C, D, E, ... The number of indicators for the evolution of the preprocessed grinding wheel body features is divided into segments according to time alignment. The correlation coefficient between the signal features of each stage is calculated to carry out correlation feature screening. Collinear redundant features with correlation coefficient ≥0.9 are removed. The remaining features are mapped to the [0,1] interval through minimum-maximum normalization to obtain the feature index set J, K, L, ...; For all the indicators of precision and surface integrity of the pre-processed wafer grinding quality dimension, correlation feature screening was carried out to remove strongly collinear redundant indicators with correlation coefficients ≥0.

85. The min-max normalization method was used to uniformly map all remaining indicator data to the [0,1] interval. The normalized features were extracted by PCA principal component analysis to obtain the principal component indicator set X with a cumulative variance contribution rate of not less than 95%.

5. The method for evaluating the full-chain application performance of precision grinding wheels according to claim 4, characterized in that, The hierarchical performance mapping model includes a material-bulk state mapping model, a bulk state-grinding condition mapping model, and a grinding condition-wafer quality mapping model; the material-bulk state mapping model is a first-level model, the bulk state-grinding condition mapping model is a second-level model, and the grinding condition-wafer quality mapping model is a third-level model. Using the principal component index set A of spline / block material properties as the input variable of the material-bulk state mapping model, and the feature index set B of the grinding wheel body characteristics and the feature index sets J, K, L, ... of each stage of the grinding wheel body characteristic evolution in the grinding process dimension as the output variables of the material-bulk state mapping model, the model is trained by multivariate nonlinear regression analysis combined with gradient boosting machine learning algorithm. The model is fitted with quantitative constraint formulas for the degradation rate and degree of the grinding wheel body state of static material parameters, and a quantitative relationship between material properties and the surface state changes of the grinding wheel body is established. The feature index set B of the grinding wheel body feature dimension and the feature index set J, K, L, ... of each stage of the grinding wheel body feature evolution in the grinding process dimension are used as input variables of the body state-grinding condition mapping model. The principal component index set C, D, E, ... of each stage of the grinding process dynamic signal in the grinding process dimension are used as model output variables. The body state-grinding condition mapping model is built and trained by using multivariate nonlinear regression analysis combined with gradient boosting machine learning algorithm and time series correlation fitting method. The multi-input multi-output mapping function is obtained, and the intrinsic correspondence between the surface state change of the grinding wheel body and the characteristics of the on-site grinding dynamic condition signal is established. The principal component index sets C, D, E, ... of each stage of the grinding process dynamic signal in the grinding process dimension are used as input variables of the grinding condition-wafer quality mapping model, and the principal component index set X of the wafer post-grinding quality dimension is used as output variables of the grinding condition-wafer quality mapping model. The grinding condition-wafer quality mapping model is iteratively trained by using multivariate nonlinear regression analysis combined with gradient boosting machine learning algorithm, and a quantitative prediction function between the grinding process dynamic signal and the wafer grinding quality index is established.

6. The method for evaluating the full-chain application performance of precision grinding wheels according to any one of claims 2, 3, and 5, characterized in that, The evaluation criteria in step S4 are processed by extracting standardized key features of each dimension, including: grinding wheel body morphology features and process evolution indicators that have only undergone minimum-maximum normalization; and comprehensive principal component indicators of material properties, grinding dynamic signals, and wafer post-grinding quality obtained by normalization and PCA principal component analysis dimensionality reduction. The attributes of all evaluation features are defined as positive performance indicators, negative performance indicators, and neutral performance indicators. For the indicators of grinding wheel body morphology and process evolution that are normalized without PCA principal component analysis, the positive, negative, or neutral attributes corresponding to the original indicators are directly inherited. The eigenvector weight matrix corresponding to the comprehensive principal component indicators is extracted, and the original indicator with the highest absolute weight value is selected as the dominant term. The overall category of the principal component is defined according to the original indicator attributes of the dominant term: if the dominant term is a positive performance indicator, the principal component is a positive comprehensive performance indicator; if the dominant term is a negative performance indicator, the principal component is a negative comprehensive performance indicator; if the dominant term is a neutral performance indicator, the principal component is a neutral comprehensive performance indicator. A 0-100 point quantification system is used to score individual indicator attributes. Two-level scoring thresholds are set for all features: a minimum acceptable threshold T1 and a minimum excellent threshold T2. Based on the individual attribute score, three performance level intervals are defined: a single attribute score ≥ the minimum excellent threshold T2 is judged as Grade A (excellent); a minimum acceptable threshold T1 ≤ a single attribute score < the minimum excellent threshold T2 is judged as Grade B (acceptable); and a single attribute score < the minimum acceptable threshold T1 is judged as Grade C (unacceptable). The implementation method of the cross-scenario evaluation criterion migration correction rule is as follows: Select multiple grinding test machines to conduct rough grinding, semi-fine grinding, and fine grinding process tests, covering various wafer substrates such as silicon wafers, silicon carbide, gallium nitride, and sapphire, to obtain data index samples under multiple scenarios; set standard machines, semi-fine grinding, and silicon wafers as benchmark scenarios, corresponding to basic qualified lower limit thresholds. Basic high-quality lower limit threshold For different machine tools, semi-finished grinding, and silicon wafer processing scenarios, the ratio of the average score of key feature indicators for each dimension to the average score of key feature indicators for each dimension in the benchmark scenario is taken as the machine tool system correction coefficient K. M For standard equipment, different processes, and silicon wafer processing scenarios, the ratio of the average score of key feature indicators for each dimension to the average score of key feature indicators for each dimension in the benchmark scenario is taken as the process correction coefficient K. P For processing scenarios involving standard equipment, semi-finished grinding, and different wafer materials, the ratio of the average score of key feature indicators for each dimension to the average score of key feature indicators for each dimension in the benchmark scenario is taken as the material correction coefficient K. w For multi-factor scenario changes, the corresponding correction coefficients are multiplied to obtain the overall scenario correction coefficient; the scoring threshold is dynamically adjusted based on the overall scenario correction coefficient K: the lower limit of qualification threshold. High-quality lower limit threshold .

7. The method for evaluating the full-chain application performance of precision grinding wheels according to claim 6, characterized in that, The aforementioned problem tracing and process optimization involves a layer-by-layer tracing of performance non-compliance items, abnormal operating conditions, or wafer quality defects encountered during the evaluation process: When the wafer grinding quality fails to meet standards, a three-level model is used to reverse-engineer the principal component index set C, D, E, ... of the grinding dynamic signal causing the quality non-compliance. The located grinding principal component indexes are then subjected to inverse projection transformation using a pre-stored PCA eigenvector weight matrix to obtain normalized original index features. Inverse normalization is then used to reconstruct the actual grinding dynamic signal data with true dimensions, thus locating the abnormal grinding operating conditions. Finally, the located grinding principal component indexes are input into a two-level model for inverse derivation. The normalized characteristic index of the grinding wheel body under the matching working conditions is solved. The grinding wheel body characteristics are restored to the actual measured index of the grinding wheel surface morphology through inverse normalization. Using the actual measured index of the grinding wheel body morphology as the output constraint, the principal component index of the corresponding spline / block material properties is solved through a first-level model. The principal component index of the located material is converted by PCA principal component analysis inverse projection and inverse normalization to restore the original material inherent measured index data. In addition, the qualitative identification images of abrasive grain exposure, uniform distribution, damage and breakage, and interface porosity detected by microstructure feature detection are used for auxiliary analysis and verification to accurately pinpoint the root cause of grinding quality problems. Based on the problem tracing results and model correlation patterns, we propose improvement schemes for optimizing grinding wheel material formulation, grinding wheel structural parameters, and matching dressing process with grinding process. After obtaining the principal component indices of the spline / block material properties, these indices are used as input constraints to solve the characteristic indices of the grinding wheel body feature dimension and the characteristic indices of each stage of the grinding wheel body feature evolution in the grinding process dimension through a first-level model. Then, the characteristic indices of the grinding wheel body feature dimension and the characteristic indices of each stage of the grinding wheel body feature evolution in the grinding process dimension are used as input constraints to solve the principal component indices of each stage of the grinding process dynamic signal in the grinding process dimension through a second-level model. Finally, the principal component indices of each stage of the grinding process dynamic signal are used as input constraints to solve the principal component indices of the wafer post-grinding quality dimension through a third-level model.

8. The method for evaluating the full-chain application performance of precision grinding wheels according to claim 7, characterized in that, A linearly increasing transverse shear force was applied to a single abrasive grain using a diamond indenter of a nano-scratch instrument, and the critical load at the time of abrasive grain detachment was recorded as the shear detachment strength. Normal pull-out tests were conducted on a composite plate sample of diamond sheet and binder, with the pull-out force increasing linearly until the interface between the diamond sheet and binder detached. The detachment fracture surface was observed and analyzed under a microscope. If the detachment fracture surface showed complete separation of the diamond sheet and binder, the pull-out force was recorded as the pull-out bond strength; if binder residue remained at the detachment fracture surface, the pull-out force was recorded as the interfacial fracture strength. The interfacial wettability was evaluated by observing the wetting angle of the binder on the diamond surface after the binder was melted. The laser triangular displacement sensor is used to scan and collect the global macroscopic morphological fluctuation data of the surface under the state of grinding wheel rotation. The collected data is processed by periodic filtering, outer envelope feature point extraction and spline fitting. The fitted curve is converted into polar coordinates and displayed as the rotation profile of the grinding wheel. The height difference between the highest and lowest points in the rotation profile is calculated as the dynamic rotation runout of the grinding wheel. The difference between the mean values ​​of all data points in the rotation profile before and after grinding wheel wear is the tooth height wear amount. The micro-topographic contour data of the grinding wheel surface was collected by in-situ scanning using a point spectral sensor. The collected data was processed by sliding window threshold filtering, contour median line extraction, peak and valley feature point extraction, and feature parameter calculation to obtain the topographic roughness RSa, topographic dispersion RSq, average abrasive grain tipping height h, and micro-profile skewness Rsk. The terrain roughness RSa is: ; The terrain dispersion RSq is: ; The micro-profile skewness Rsk is: ; The average abrasive grain exit height h is: ; In the formula: n is the total number of data points collected for the micro-topographic contour of the entire surface of the grinding wheel. Let i be the measured value of the i-th collection point. Let be the value on the median line of the contour corresponding to the i-th sampling point, and t be the total number of data points of the contour peak data of the micro-topography of the entire surface of the grinding wheel. Let j be the measured value of the j-th peak. The value on the midline of the contour corresponding to the j-th peak point; A three-dimensional force gauge, power meter, accelerometer, acoustic emission sensor, and infrared thermal imager were set up in the grinding area of ​​the grinding test platform. The grinding force, spindle load power, grinding vibration, acoustic emission, and grinding zone temperature were acquired using a data acquisition card during the entire grinding process. A wafer full inspection instrument equipped with infrared interferometry and spectral confocal probes was used to detect the surface peak-valley deviation PV, curvature Bow, and warp under standard three-point support conditions. The total thickness deviation TTV and local thickness deviation LTV were detected under full-width vacuum plane adsorption support conditions. The surface roughness Ra and Rz of the ground workpiece were detected using a white light interferometer; the fisheye diameter and depth, scratch depth and width, crack length, and chipping size of the surface were observed using a confocal microscope and an industrial camera; the thickness of the subsurface damage layer was detected using laser scattering; and the residual stress of the surface layer was analyzed using wafer curvature analysis and Raman spectroscopy. The wavelet threshold filtering algorithm sets the wavelet basis function and threshold, and uses the 4th-order Daubechies wavelet basis to carry out multi-level discrete wavelet decomposition with 5 wavelet decomposition levels; it uses SURE unbiased adaptive thresholding, and the threshold of wavelet detail coefficients in each level is determined by the noise standard deviation and length. The positive performance indicators include: flexural strength, tensile strength, impact strength, shear spalling strength, pull-out bond strength, fracture strength, wettability, terrain roughness RSa, terrain dispersion RSq, average abrasive grain exit height h, and microscopic profile skewness Rsk; the negative performance indicators include: friction coefficient and stability, temperature rise, material wear rate, grinding wheel dynamic rotational runout, rotational profile, tooth height wear, grinding vibration, grinding zone temperature, increase in grinding wheel dynamic rotational runout, range of grinding wheel rotational profile changes, tooth height wear difference, and terrain roughness. The reduction values ​​of surface roughness RSa, topographic dispersion RSq, average abrasive grain exit height h, micro-profile skewness Rsk, total wafer thickness deviation TTV, local thickness deviation LTV, surface peak-valley deviation PV, curvature Bow, warp, surface roughness Ra and Rz, surface micro-defects, subsurface damage layer thickness, and surface residual stress; neutral performance indicators include: density, porosity, hardness, abrasive concentration, grain size distribution, elastic modulus, force ratio, grinding force, acoustic emission, and spindle load power; Set the global maximum value of the key feature index across the entire sample. Global minimum value The feature index to be evaluated, X, and the score of each feature index. Positive performance indicators negative performance indicators Set the optimal range for neutral performance indicators. Full sample limit boundary A score of 100 is obtained within the optimal interval. The score is calculated towards both boundaries using the positive formula and the negative index, until it decreases linearly to 0. The key characteristic index attribute scores of the grinding wheel performance in each dimension under the evaluation scenario and the score judgment threshold after evaluation criterion transfer correction are obtained, and the obtained grades are assigned scores: Grade A is excellent with 10 points, Grade B is qualified with 5 points, and Grade C is unqualified with 0 points. The graded scores of all index features in each dimension are calculated by weighted average. The features of the four dimensions of spline / block material performance, grinding wheel body features, grinding process, and wafer post-grinding quality are assigned weights of 20%, 20%, 20%, and 40%, respectively, to obtain the comprehensive performance score of the grinding wheel under the evaluation scenario.

9. A full-chain application performance evaluation system for precision grinding wheels, utilizing the full-chain application performance evaluation method for precision grinding wheels according to any one of claims 1-8, characterized in that, include The data acquisition module provides complete and authentic multi-source heterogeneous data for the entire chain performance evaluation; The data processing module is used to standardize and organize multi-source heterogeneous data, remove invalid and interfering data, deeply mine and extract key feature parameters, and construct a dataset of key features of grinding wheel performance. The model building and analysis module is used to achieve model parameter fitting and dynamic updating, and to conduct performance correlation inference and trace the cause of grinding wheel performance defects based on the model. The evaluation output module is used to calculate the comprehensive performance level according to the established standards, integrate the analysis results, automatically generate an evaluation report, and output optimization and improvement suggestions. The human-computer interaction module is used to input the working conditions and basic parameters of the sample, and to visually present the evaluation process, analysis results and performance judgment conclusions.

10. The full-chain application performance evaluation system for precision grinding wheels according to claim 9, characterized in that, The data acquisition module includes a spline / nodal testing unit, a grinding wheel body inspection unit, a grinding process monitoring unit, and a wafer quality inspection unit. The spline / nodal testing unit integrates a universal material testing machine, an impact testing machine, a nano-scratch tester, a friction and wear testing machine, a sandblasting hardness tester, a laser particle size analyzer, and a scanning electron microscope to collect intrinsic material parameters such as mechanical properties, interfacial bonding characteristics, and friction and wear. The grinding wheel body inspection unit integrates a laser triangular displacement sensor and a point spectral sensor to perform in-situ detection of macroscopic feature parameters and microscopic topographic features of the grinding wheel. The grinding process monitoring unit is mounted on a wafer precision grinding test platform and integrates a triaxial force gauge, a power meter, an accelerometer, an acoustic emission sensor, and an infrared thermal imager to achieve synchronous, high-frequency, and accurate acquisition of multi-source physical signals under all grinding conditions. The wafer quality inspection unit includes a wafer full inspection instrument, a white light interferometer, a confocal microscope, an industrial camera, a laser scattering detector, and a Raman spectrometer to comprehensively detect quality indicators related to wafer processing accuracy and surface integrity. The data processing module includes a data preprocessing unit and a feature extraction unit. The data preprocessing unit completes noise reduction of multiple types of sensor signals, time synchronization of data at different sampling frequencies, and normalization of data across the entire domain. It also automatically removes abnormal interference data based on the 3σ criterion. The feature extraction unit performs targeted feature operations on data of various dimensions, automatically mines and quantifies various core feature parameters, and constructs a standardized and structured dataset of key features of grinding wheel performance. The model building and analysis module has a built-in three types of hierarchical performance mapping models with progressive association, which are embedded with machine learning training iterations. Based on the hierarchical mapping relationship, it completes multi-dimensional performance correlation inference and realizes the linkage analysis of grinding wheel material performance, body state, grinding conditions and wafer processing quality. The evaluation output module has a built-in data storage unit that pre-stores quantitative evaluation criteria and cross-scenario evaluation criterion migration and correction rules. It is configured to perform comprehensive evaluation calculations and automatically generate reports. It completes weighted calculations based on the score weights of each dimension indicator and integrates information from each dimension to compile evaluation reports and optimization and improvement plans. The human-computer interaction module includes an information input terminal, a visualization display component, and an interactive control program. It supports operators in inputting basic information and visually displays the data acquisition process, data processing results, model correlation analysis curves, grinding wheel performance levels, and defect tracing results.

Images