TW201135474A - Method for sampling workpiece for inspection and computer program product performing the same - Google Patents

Abstract

A method for sampling a workpiece for inspection and a computer program product performing the method are disclosed. The method calculates a reliance index (RI) and a RI threshold (RIT) associated with a virtual measurement value of a workpiece; and a global similarity index (GSI) and a GSI threshold (GSIT) associated with the process data of the workpiece by analyzing process data applied on a unit of production equipment. When the RI of the workpiece is smaller than the RIT or the GSI thereof is greater than the GSIT, the workpiece can be chosen for measurement.

Description

201135474 VI. Description of the Invention: [Technical Field] The present invention relates to a method for sampling inspection of a workpiece, and more particularly to a method for effectively sampling a defective guard and a computer program product thereof. [Prior Art] At present, most semiconductor and TFT-LCD factories adopt the sampling method for the quality monitoring methods of the products or workpieces of the production machine. The workpiece can be the wafer of the semiconductor industry or the glass of the TFT-LCD industry. Substrate ^When the production machine completes the processing of several workpieces, the workpieces are placed in the Front Opening Unified Pod; FOUP for transmission to the measuring machine. Detecting the quality of the workpiece. Generally speaking, the measuring machine will randomly select one workpiece from the entire number of workpieces (for example: 25 pieces) for measurement, for example: the first one in the cassette The conventional sampling method is based on the assumption that the process quality of the production machine will not suddenly be abnormal, so the measurement results of the product or workpiece being sampled can be used to infer the same cassette or wafer transfer box. The quality of all products. However, the method of sampling inspection can only know the quality of the workpiece that is actually being tested, and the workpiece that is actually sampled is not necessarily a potential risk. The workpiece often causes Miss Detection (MD). In addition, if the production machine is abnormal between the two tests, the conventional method cannot be found in time, resulting in many defects. And cause considerable cost loss. In theory, if all the workpieces 201135474 in the same cassette or wafer transfer box can be measured, the above-mentioned leak detection can be avoided, and the production machine can be found in time. Abnormal. However, actual measurement of each workpiece in the same cassette or wafer transfer cassette is quite time consuming and requires a lot of manpower and material resources. Moreover, for wafers or TFT_LCD plants with hundreds of processes It is almost impossible to perform actual measurement for each workpiece of each process. Therefore, in order to avoid the above problems, it is necessary to provide a method of sample sampling inspection and computer program products, thereby effectively The appropriate workpiece is selected for measurement, and it can be found in time when an abnormality occurs in the production machine. [Invention] Therefore, one of the present inventions This is to provide a method of sample sampling inspection and its computer program product 'by determining whether the confidence index of the workpiece (Reliance Index; RI) is less than the confidence threshold value; or GSI (Global Similarity Index; overall similarity) Whether the value of the indicator is greater than the GSI threshold value (GSIT) 'to effectively select the appropriate workpiece for measurement' and it is free from the occurrence of the leak test, and can find the abnormality of the production machine in time. Objective, a method is proposed. In the embodiment of the present invention, first: sampling the historical process parameter data of the test, and from the measuring machine "the complex array value of the machine, wherein the historical measurements are respectively The measured value of the workpiece based on historical measurements. Then, using the historical process=parameters to produce the measured values to establish-estimate the model and the reference-wide data and historical quantity basis--estimation algorithm, reference mode; The reference 201135474 algorithm 'infers the algorithm and the reference parameters to the estimated mode R. Then enter 'the historical measurement value> > test mode, and calculate a plurality of historical virtual plural historical reference _ value 1, the distribution is historically virtual

The assignment of the gate 2 (10) (4) This 011) and the area of the historical reference prediction value are generated to generate a plurality of confidence index values, wherein when the overlap, the larger the product, the higher the heart index value, the representative corresponds to the historical virtual measurement. The higher the credibility of the value. Then, based on the historical virtual (four) value, the historical reference predicted value, and the historical measured value, the -indicator threshold value (8)T is calculated. Secondly, collect the process parameter data of the plurality of workpieces in the cassette sent by the production machine, and input the process parameter data of each workpiece to the estimation mode and the reference mode, and calculate the virtual measurement value and reference of each workpiece. Predictive value. Then, calculating the overlap area between the allocation of the virtual measurement value of each workpiece and the allocation of the reference prediction value to generate a confidence index value of each workpiece, wherein the larger the overlap area, the higher the confidence index value, the representative office The higher the credibility corresponding to its virtual measurement. Then, at least one first workpiece whose confidence index value is less than the confidence threshold value is selected from the workpieces and the first workpiece is sent to the measuring machine for detection. According to still another embodiment of the present invention, in the method of sample sampling inspection, the data of the complex array history of the production machine is first obtained. Then, using these group history process parameters, and based on a statistical distance algorithm, a statistical distance mode is established. Then, using these historical process data and these historical measurements, and applying the Leave-One-Out (LOO) principle in Cross Validation to reconstruct the statistical distance pattern and calculate the phase The corresponding GSI value is used to calculate a GSI threshold (GSIT). Then, the process parameter data of each workpiece is input to the statistical distance mode, and 201135474 · «+ calculates the GSI value of the process parameter data corresponding to the virtual measurement value of the parent workpiece. Then, at least one second workpiece whose GSI value is greater than the Gsi threshold value is selected from the workpieces, and the second workpiece is transferred to a measuring machine for detection. According to the above object of the present invention, a computer program product for storing sample sampling inspection is provided, and when the computer is loaded into the computer program product and executed, the method of sampling inspection of the workpiece as described above can be completed. Therefore, by applying the invention, it is possible to evaluate whether the quality of the workpiece may be abnormal by using the process parameter data of a workpiece, so as to effectively select a suitable workpiece for measurement, and avoid the occurrence of leak detection, and timely Found that the production machine is abnormal. [Embodiment] Referring to Fig. 1, there is shown a system architecture diagram showing a method for performing sample inspection of a workpiece of the present invention. The present invention provides a fully automated Virtual Virtual Metrology (AVM) system 90 between the production machine φ 20 and the measuring machine 30 for assisting the process parameter data 22 of all workpieces 82 in the cassette 80. The measuring machine 30 draws the appropriate workpiece for measurement. In an embodiment, the AVM system 90 first notifies the Manufacturing Execution System (MES) (not shown) the code of the selected workpiece, and the manufacturing execution system then gives instructions to the selected workpiece according to the code. The measuring machine 30 measures the workpiece to be selected. Moreover, in an embodiment, the AVM system 90 is embedded in the measurement machine 30. In yet another embodiment, the AVM system 90 is embedded in the production machine 20. Of course, the AVM system 90 can also independently perform the workpiece sampling method of the 201135474, so the present invention is not limited thereto. Please refer to FIG. 2, which is a schematic diagram showing the architecture of an AVM system according to an embodiment of the present invention. The AVM system 90 of the present embodiment includes at least: a process parameter data pre-processing module 10, a measurement data pre-processing module 12, an estimation mode 60, a confidence indicator module 40, and a similarity index module 50. The process parameter data pre-processing module 10 collates and standardizes the original process parameter data from the production machine 20, deletes the abnormal data and filters out the important parameters, and excludes the unimportant parameters to avoid the interference effect and affect the prediction accuracy. . The measurement data pre-processing module 12 filters the measurement data from the measurement machine 30 to remove the abnormal value therein. The estimation mode 60 can use the estimation algorithm to estimate the first stage virtual measurement value (VMi) of the plurality of workpieces 82 in the cassette 80, and can also selectively utilize the two-stage operation mechanism 62 and the estimation algorithm. The second stage virtual measurement value (VM„) of the plurality of workpieces 82 in the cassette 80 is estimated. The possible estimation algorithms are: regression algorithm, neural network algorithm and other prediction algorithms. The confidence indicator module 40 is used to evaluate the credibility of the virtual measurement value, and generates a confidence indicator (RI). The similarity indicator module 50 is used to evaluate the currently input process parameter data and the estimation mode 60. The similarity degree of all parameter data of the training model is trained, and the similarity index (GSI) of the process parameter is generated, which is used to assist the confidence index to judge the confidence of the virtual measurement system. Before the estimation mode 60 operates The process parameter data (historical process parameter data) obtained from the production machine 20 and the product quality measurement data (historical measurement value) obtained from the measuring machine 30 are respectively transmitted to the process parameter data pre-processing module. 10 and measurement Before the material processing module 12, for 201,135,474 treatment before information on these ί after pre-treatment and standardization *;. Γ (νΜ,) " good ^ mT, i index values (RI) and the overall similarity and "value (). The so-called "second stage" virtual confidence indicator and similarity indicator county are getting jobs from the machine;; ^ == adding the process parameters and actual measurement values of the two pieces of 82 to the historical system & parameter reading and historical volume dragon , to re-evaluate mode 60, confidence indicator module 4〇 reference mode and similarity index = group 50 statistical distance mode, and then recalculate each of the card = 8 虚拟 virtual value (VMlI) and its The accompanying confidence indicator and the following theoretical basis for the description of the estimation model, the confidence index value (reference model) and the overall similarity index value (statistical distance model). Estimation mode and confidence index (as shown in Table 1, assume that the current E set to n sets of measurement data, including process data (^=/, 2,..., „) and their corresponding The actual measured value data (is ^^ 2,...,π) 'where each set of process data contains p parameters (from parameter 1 to parameter p), ie \ = [\,,···,~b, It also collects ((4)) the process data of the actual production of the pen, but there is no actual measurement data except for „+1, that is, in the actual workpiece produced by the pen, only the first workpiece is actually measured for actual measurement. Then, based on its actual measurement of eight +1, the other (w_w") pen workpiece is inferred. 201135474 Table 1 original data sample sample data point parameter 1 parameter 2 parameter P actual measurement value 1 ^1,2 • · · Xhp less 1 2 χ2, \ x2,2 • * « Χ2,Ρ yi ... ... ... ... ... ... η 〜,1心,2 • · · XfhP yn η+1 ^/7+1,1 ^m+1, 2 • · · ^n+\yp yn+\ η+2 χη+2,\ χη+2,2 # # · ^n+2,p Zip ... ... ... · · ... ... m ^m,\ •V2 ... Xm , p Zip

In Table 1, y, h, ..., Λ are historical measurements, and Λ +1 is the actual measured value of the first workpiece in the workpiece batch being produced. Usually, a set of actual measured values (ie, ζ·=1, 2, ..., η) is a normal distribution with an average of / /, the standard deviation σ, that is, ~Λ^, σ2). For the sample group (the average of the brothers _, / = 1, 2, ..., 4 and the standard deviation to normalize all the actual measured data, you can get \, Ζ, 2, ..., (also known as Ζ score (z Scores)), where the average of each score is 0, and the standard deviation is 1, ie \~#(〇,1). For actual measurement data, if 4 is closer to 〇, it means measurement The closer the data is to the central value of the specification, the standardized formula such as 201135474 7 , ΐ-1, 2, ···, η σ is less (1) y=-(yi+y2+---+yn) η (2) iyi~ y)2 +(y2~y)2 +··^^-j)2] (3) where the brother is the actual measured value of the group/group; ^ is the actual measured value after the standardization of the Z-group data Data; J is the average of all actual measured data; 'S is the standard deviation of all actual measured data; the description here is the application of neural network (NN) algorithm estimation: 寅 algorithm to establish The estimation mode of the virtual measurement is performed and the reference mode for verifying the estimation mode is established by, for example, a reference prediction algorithm of the regression algorithm. However, the present invention may also use other algorithms as the estimation algorithm or the reference prediction method. As long as the reference prediction algorithm is different from the estimation algorithm, the present invention is not limited thereto. The estimation algorithm and the reference prediction algorithm of the present invention may be, for example, a Back Propagation Neural Network (BPNN), a General Regression Neural Network (GRNN), a radial basis. Radial Basis Function Neural Network (RBFNN), Simple Recurrent Network (SRN), Support Vector Data Description (SVDD), Support 201135474 Vector Machine (Support Vector Machine; SVM) ), Multiple Regression (MR); Partial Least Squares (PLS), Nonlinear Iterative Partial Least Squares (NIPALS) or Generalized Linear Models (Generalized Linear Models; GLMs).

When applying the neural network algorithm and the complex regression algorithm, if the convergence condition is the smallest Sum of Square Error (SSE), and the normalization of the two modes after the +〇〇 If the measured values are ambiguous and ^, they should be the same as the actual measured values & In other words, at the time, today = 2 outside = ^ represents the actual measured value after standardization, but the name is changed for the purpose of different modes. Therefore, ~, eat)' and a. ~ bucket &, σ!), indicating that 'and ~ are the same assignment, because the different estimates of the § 10 mode, the average of the two prediction algorithms is different from the estimate of the quasi-difference. The normalized average number estimation formula (\=~,) and the fresh error estimation formula (V&) are different from the standard deviation estimation formula (which is different from the standard deviation estimation formula (qi^^). The value is designed to calculate the trustworthiness of the virtual measured value. The value of this U index should be (4) the statistical distribution of the virtual (four) value ~ the statistical distribution of the actual measured value z _ phase _ degrees. However, ❹ 1 In the virtual measurement, there is no actual measurement value that can be used to evaluate the reliability of the virtual measurement (obviously, if the actual measurement is obtained, it does not need to be measured 12 201135474). Therefore, the present invention adopts The statistical allocations estimated by reference prediction algorithms (such as complex regression algorithms) are substituted for the statistical allocation of & the reference prediction algorithm of the present invention may also be other related prediction algorithms, so the present invention is not limited thereto. Eyes refer to Figure 3, which illustrates the hair Schematic diagram of the confidence indicator value of the embodiment. The confidence indicator value of the present invention is defined as the allocation and reference mode of the prediction (virtual measurement value) of the calculation estimation mode (for example, using a neural network (NN) algorithm) ( For example, the distribution of the prediction (reference measurement value) using the complex regression algorithm allows the intersection area coverage value (overlap area A) between the two. Therefore, the formula of the confidence index value is as follows: πσ dx (4) where Lu When ζ 则 2 & then 4 = ^. σ is set to 1 私山 t. The 彳 彳 值 随着 随着 随着 随着 随着 。 彳 彳 彳 彳 彳 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 It is closer to using the reference mode f, Λ U and the corresponding virtual measurement value is more reliable. Otherwise, the reliability of the phase measurement value decreases with the decrease of the overlap area A. ;, and the distribution; the estimated distribution & completely according to the statistical distribution theory, the confidence index value is equal to i; and when the two distributions are divided into six, separated by Ding, the confidence index value approaches 0. His description of the estimation mode to calculate the virtual measurement value $ 13 20113547 4 The method of allocation (SSE) is: 2 ί: ': The convergence condition is to minimize the squared error and the variance is the allocation of the bite, that is, to give the underarm, the mean of the needle is equal to %, and the estimate is "2" The ΝΝ estimate of the outer 呤 is: the modeling system of the NN estimation model ~ the standardization step. j and the advanced 仃 process data NN estimation model process data standardization formula is as follows: zv 5«—, factory no / xj ·ΐ = 1,2,··,n,n^-1,---,m; j = 1,2,-..,

P (5) 3c,. + Ar,,. + ...+ ΛΓ (6)

XJ

Xl (7) where ~ is the y. process parameter in the /group process data; is the number; z is the average of the process parameters of the process / the standard process after the standard process; The standard deviation of the first y· process parameters; use this “standardized process data h _ 2.12 to input the standardized process data should be marked (3⁄4, 卜,”·, cut, household/previous to NN estimation In the mode, the virtual measured value after the jitter is ~, : is obtained in relative (8) 201135474

0ZyN = zyNi,i = J,2r",n,n+J,...,m =i(zyNI + zh2 + - + zhn) where is the average of the normalized virtual measurements (9) (1〇) Method The following is a basic hypothesis that the complex regression algorithm for calculating reference prediction values (&, and ~) from the complex regression mode is "under a given heart, the distribution of ~ is equal to %, and the variance is the distribution of food." That is, given 2, and the complex regression estimate of % is a complex regression estimate of ~4 eat = 3⁄4. • In order to obtain the relationship between the process data after standardization and the actual measured value of the η group /=仏,”), the weights corresponding to these parameters in the complex regression analysis must be defined. The construction and relationship are as follows: fir0+βΓ , ΖΧιι + fir2ZXi2 + ·. - + prp2^ ^ = β^βΓΙζΧ2+βΓ2ζΧ22+-+βκρΖ^ρ =z' ... lp y2 (11)

Pr0 + βτ, Ζχη, + β^Ζ\, 2 + ' + KZX = Zv n, py» r L ύ 15 (12) 201135474 Suppose k,, Ζ· 2 yi Ζ y" ζχ 2,1 ζχ , ι, ρ ζν Ιρ 2 V η,Ι 夂,··· ζν η.ρ (13) Number analysis: the least square method 'can be obtained as reference (14) β^{ζτχΖχγ zrZy Then, the complex regression mode can be obtained [1-7] , 2,...,η,η+人...,m .+ β^ζ} ip (15) 艮. Therefore, in the estimation stage, after the process data comes in, the complex regression estimate Z&+ corresponding to it can be obtained according to formulas (15) and P. The standard regression number ~ complex regression estimator is: π ' (16) ~ 4 (W"+z:pJ (17) Tian Qiu NN estimation model estimation formula ^, and 彡 ~ and complex regression model 201135474 The estimation formula, and the post-6' can be plotted as the normal distribution map shown in Figure 3, and the calculation uses the estimation (such as the neural network (NN) algorithm) to predict (virtual guess). The intersection area coverage value (overlap area A) between the distribution of the prediction (reference measurement value) of the reference mode (for example, using the complex regression algorithm) can be used to obtain the confidence index value of each virtual measurement value. After obtaining the confidence index value (RI), a confidence index threshold (RIt) must be established. If RI < RIt, the virtual quantity ^ of the RI workpiece has a low degree of reliability, that is, has this RI The quality of the workpiece is more likely to be abnormal, so the actual measurement is required. The following describes the method of determining the confidence threshold (4): Before setting the confidence indicator (RIt), the maximum allowable error upper limit must be set first (6) The error of the virtual measurement value (called the actual measurement value minus NN) The percentage of the absolute difference between the average value, divided by two actual measurement value, i.e.; estimation & amp pattern of the obtained

Εηοη y> y χίοο% (18) Then 'the maximum allowable error upper limit (6) can be reduced according to the error defined by equation (18) and the virtual quantity measurement specification. Therefore, the confidence index threshold (RIT) is defined as the reference index value (RI) corresponding to the maximum allowable error upper limit ,, as shown in Figure 4. which is,

(19)

17 201135474 //The σ is defined in equation (4); and 2center = Zym + [?X ^)]/ay (20) where t is defined in equation (3). Twice Similarity Indicator (GSI) As mentioned above, when virtual measurement is applied, no actual measurement is available to verify the accuracy of the virtual measurement. Therefore, the confidence index value (RI) is calculated by substituting the normalized complex regression estimate\ instead of the normalized actual measured value 4. However, such substitution may result in an error in the confidence index value (RI). To compensate for this situation, the present invention proposes a global similarity index (GSI) of the process to help determine the reliability of the virtual measurement. The concept of the GSI proposed by the present invention compares the device process data currently used as the input of the virtual measurement system with all the historical parameter data at the time of modeling, and obtains the similarity between the input process data and all historical parameter data. index. The present invention can quantize similarities using various statistical distance algorithms, such as Mahalanobis Distance, Euclidean Distance, and Centroid Method. The Markov distance is a statistical distance algorithm introduced by P.C. Mahalanobis in 1936. This technique is based on the correlation between variables to identify and analyze the patterns of different sample groups. The Mahalanobis distance is a method for determining the similarity between an unknown sample group and a known sample group. This method considers the correlation between data sets and has a scale invariant, that is, it is not related to the magnitude of the measured value. . If the data has a high similarity, the calculated Mahalanobis distance will be smaller. 201135474 The present invention utilizes the calculated magnitude of GSI (Machine Distance) to distinguish whether the new process data is similar to all process data of the model. If the calculated GSI is small, it means that the new process data is similar to the model process data, so the virtual measurement value of the new process data (high similarity) will be more accurate. Conversely, if the calculated GSI is too large, it means that the new process data is somewhat different from the modeled process data. Therefore, the quality of the workpiece with the new process data (low similarity) is more likely to be abnormal, so actual measurement is required. The calculation formula of the standardized process parameter z of the estimation model is as follows

l, J (5), (6) and (7). First, define the pattern parameter data, which is equal to force. Thus, the parameters of the standardized modeling process data are all 〇 (that is, the standardized modeling parameter ‘is 〇). In other words, all parameters in 4 = are 0. Next, the correlation coefficient between each standardized modeling parameter is calculated. Suppose the correlation coefficient between the sth parameter and the tth parameter is rst, and there is a key information, then

Rst k— 1 1-} ^srztl=~,~:{zsrztl+zs2'zt2+- + zsk'ztk) k-1 (21) After completing the calculation of the correlation coefficient between the parameters, the correlation coefficient matrix can be obtained as follows: 1 r12 "· rlp r21 1 "· r2p ^pl ^p2 ... 7 19 (22) (23) 201135474 The hypothesis and the inverse matrix (and the system is defined as j, then

An a21 dpi °12 ··. A p a22 ... ^p2 ·· aPP A = RJ = Thus, the second standardized process parameter (zj and the standardized pattern data (the Mahalanobis distance between zj) The formula is as follows: Ό2λ=(Ζλ-ΖΜγκ-](ζλ-ΖΜ) (24) (25) Available = Σ Σ ^ij^azβ j=li=l and the GSI value of the second process data is z)f/ p. After obtaining the GSI value, apply the Leave-One-Out (LOO) principle in the cross validation to define the GSI threshold (GSIT). The formula for the GSI threshold (GSIT) is as follows: = a * GSI L〇〇(26) The so-called "Leave-One-Out (LOO) principle" extracts a test sample from the entire modeled sample as a simulated upper line, and then uses the remaining samples to establish a GSI. The model, then apply this newly created GSI model to calculate its GSI value for the test sample on the simulated line, which is represented by GSILOO. Then repeat the above steps until all the 201135474 samples in the modeled sample are calculated. GSIL00. Therefore, Equation (26) represents an example of all GSILOO calculated from all modeled samples by the LOO principle. Such as the 90% truncated average (Trimme (i Mean). The value of the formula (26) a is between 2 and 3, which can be fine-tuned according to the actual situation, the default value of & is 3. Inventive sample sampling inspection method. [Month] Referring to Figure 5, there is shown a flow chart of a sample sampling inspection method according to an embodiment of the present invention, in which a estimation mode, a reference mode, and a "ten distance mode" are established; After obtaining the confidence index threshold (RIT) and the GSI threshold (GS τ), the process parameter data of each workpiece in the cassette is transferred to the above-mentioned push mode, reference mode and statistical distance mode to calculate each workpiece. The centroid index value (RI) and the GSI value (step 1〇〇). Next, step 110 is performed for each ^8 to determine whether the confidence index value (RI) is less than the threshold value (RIt); Whether the GSI value is greater than the asi threshold value - if the determination result in step 110 is yes, proceed to step 120 to determine the condition of the step 11G to perform the measurement; otherwise, the end # is the workpiece sampling inspection method. The judgment result is 疋 仃 仃 step 120, to measure the machine This workpiece is measured by the amount of workpieces that are in the condition of U〇 in the card. In the other embodiment, the characteristics of each workpiece in the same cassette are the same, so only lit: The workpiece that meets the conditions of step 110 may be selected from at least - the mechanics / (four). If the result of the determination in step 12G is negative, the workpiece sampling inspection method of this embodiment is ended. It can be understood that the method for sampling inspection of the workpiece of the present invention is as described above.

21 201135474 Description of the Operation Steps The computer program for storing sample inspections of the present invention is used to perform the method of sampling inspection of the workpiece as described above. Please refer to FIG. 6A to FIG. 6C, and FIG. 6A is a diagram illustrating the present invention.

A schematic diagram of the results of the virtual measured value and the actual measured value of the application example; FIG. 6B is a schematic diagram showing the result of the confidence index value of the application example of the present invention; FIG. 6C is a diagram showing the overall similarity of the application example of the present invention Schematic diagram of the results of the indicator values. As shown in Figure 6A, the first to the hundredth data are the historical measurements of the first to the 100th workpieces, which correspond to the complex array history process parameter data, respectively, to establish the estimation mode, the reference mode, and the statistical distance. Mode; and obtain confidence index threshold (rIt) and GSI threshold (GSIT). The 10th to the 125th data is a plurality of workpieces in the cassette, wherein the 101st workpiece is sampled by a conventional method, so the actual measurement value is used to adjust the establishment of the estimation mode, the reference mode and the statistical distance. mode. The purpose of the sample sampling inspection method of the present invention is to effectively select a suitable workpiece from the 102nd to 125th workpieces for measurement, thereby avoiding the occurrence of leak detection, and detecting the abnormality of the production machine in time. As shown in Figure 6B, the confidence index value (RI) is used to determine that the first workpiece (indicated by the first data set) needs to be sent to the measuring machine for testing. Next, as shown in Fig. 6C, the overall similarity index value (Gsi) is used to determine the degree of similarity between the workpiece data and the modeling data. Among them, the 114th data group's although its RI is greater than RIT (0.567), but its GSI is greater than GSIT (5.093), which means that the 114th workpiece needs to be sent to the measuring machine for detection to prevent leakage detection. occur. The 107th and 12th data, because its RI is less than RIt and because its GSI is greater than GSIt, it is necessary to send the 22nd 201135474 ι〇7, 12〇 workpieces to the measuring machine for detection to prevent the RI. The situation happened. In addition to the 107th, 114th and 120th data, since the rest of the workpiece has an RI greater than RIt and its GSI is less than GSIt, it means that it is not necessary to detect the workpieces other than the 107th, 1Hth and 120th workpieces, thus saving manpower and material resources. In another embodiment, the present invention may also select at least one of the first, seventh, and 120 workpieces for measurement. It can be seen from the above preferred embodiment of the present invention that the method of the sample sampling inspection of the present invention can effectively extract a suitable workpiece for measurement, avoiding the occurrence of a leaking test, and can timely detect the abnormality of the production machine. . The present invention has been disclosed in the above embodiments, and it is not intended to limit the invention to any of the ordinary skill in the art, and various changes and modifications may be made without departing from the spirit and scope of the invention. Therefore, the scope of the invention is defined by the scope of the appended claims. BRIEF DESCRIPTION OF THE DRAWINGS The above description of the present invention and other objects, features, advantages and embodiments of the present invention are more clearly understood. The description of the drawings is as follows: Figure 1 is a green sampling of the workpiece to which the present invention is implemented. The system architecture of the method of inspection is not intended. FIG. 2 is a schematic diagram showing the architecture of the AVM system according to an embodiment of the present invention. Fig. 3 is a schematic view showing the value of the confidence index of the embodiment of the present invention. Figure 4 is a schematic diagram showing the confidence index threshold of the embodiment of the present invention only ^ 1 Ci 23 201135474. Fig. 5 is a flow chart showing the method of sampling inspection of a workpiece according to an embodiment of the present invention. Fig. 6A is a view showing the results of the virtual measured value and the actual measured value of the application example of the present invention. Fig. 6B is a diagram showing the result of the confidence index value of the application example of the present invention; and Fig. 6C is a diagram showing the result of the overall similarity index value of the application example of the present invention.

[Main component symbol description] 10: Process parameter data pre-processing module 12 20 Production machine 22 30 Measuring machine 40 50 Similarity index module 60 62 Two-stage operation mechanism 80 82 Workpiece 90 100: Calculate the RI of the workpiece GSI 110: Whether the RI is smaller than the RIT; or GSI; 120: Decide whether to perform the measurement 130: Measurement by the measuring machine A: Overlap area

Measurement Data Pre-Processing Module Process Parameter Data Confidence Index Module Estimation Mode Card 匣 AVM 24

Claims (1)

201135474 VII. Patent application scope: • ^ The method of sampling inspection, the deaf person obtains the complex array history of a production machine. The multi-measurement value obtained from the measurement machine is respectively based on the historical values of the groups. The measured value of the workpiece; the workpiece produced by the parameter Using the set of historical process parameters to establish a estimation mode, wherein the estimation model method; ~ data and the historical measurement values are established according to a estimation algorithm using the set of historical process parameters to establish - reference a mode in which the reference oscillates the historical measurement data to the algorithm 'the estimation algorithm and the reference performance. Built, according to a reference number = value history process parameters: and multiple complex == value history process parameters to the reference mode, and calculate the complex 】 Calculate the overlapping area between the distribution of the historical virtual measurement values and the allocation of the reference prediction values to generate a plurality of historical L-heart index values (Rdiance Index; RI), wherein when the overlapping area The larger the confidence indicator value is, the higher the reliability corresponding to the historical virtual measurement values is represented; according to the historical virtual measurement values, the historical reference prediction values, and the historical measurement values Calculating a confidence index threshold; collecting process parameter data of a plurality of workpieces in one of the cassettes sent by the production machine; 25 201135474 inputting process parameter data of the workpieces to the plurality of workpieces a virtual measurement value; a formula of a plurality of reference values of the workpieces to which the reference molds are output; At least one first workpiece having a confidence value is selected from the workpieces; and the workpiece is sent to the measuring machine to detect the at least one first row of the workpieces. # 汝 汝 汝 一 一 一 一 一 一 一 一 一 一 一 一 一 一 一 Ne Ne Ne Ne Ne Ne Ne Ne Ne 3. The method of sampling inspection of a workpiece according to claim 1, wherein the estimation algorithm and the reference algorithm are respectively selected from a Back Propagation Neural Network (ΒΡ> ΒΡ> ), a General Regression Neural Network (GRNN), a Radial Basis Function Neural Network (RBFNN), a simple regression network 26 201135474 (Simple Recurrent Network; SRN ), a Support Vector Data Description (SVDD), a Support Vector Machine (SVM), a Multiple Regression (MR); Partial Least Squares; PLS), a non-linear alternative partial least-square method (NIPALS) and a generalized linear model (GLMs). 4. The method for sampling inspection of a workpiece according to claim 1 of the patent application, further comprising: using the set of historical process parameters, and establishing a statistical distance mode according to a statistical distance algorithm; Data and these historical measurements, and apply the Leave-〇ne_0ut; l〇〇 principle in Cross Validation to reconstruct the statistical distance pattern and calculate the corresponding GSI (Global Similarity) Index; overall similarity index) value to calculate a GSI threshold value; input process parameter data of the workpieces to the statistical distance mode, and e ten calculate the process corresponding to the virtual measurement values of the workpieces a GSI value of the parameter data; selecting at least one second workpiece whose GSI value is greater than the GSI threshold value from the workpieces; and sending the at least one second workpiece of the workpieces to the measuring machine for inspection . 27 201135474 5. The method for sampling inspection of a workpiece according to claim 4, wherein the statistical distance algorithm is selected from a Mahalanobis Distance algorithm, an Euclidean Distance algorithm, and A group of the Centroid Method. 6. A method for sample sampling inspection, comprising: obtaining a complex array historical process parameter data of a production machine; using the group historical process parameters, and establishing a statistical distance mode according to a statistical distance algorithm; The historical process data of the group and the historical measurement values, and applying the retention-one principle in the interaction verification to reconstruct the statistical distance mode, and calculating a corresponding GSI value to calculate a GSI threshold; Process parameter data of the workpieces to the statistical distance mode, and calculating GSI values of the process parameter data corresponding to the virtual measurement values of the workpieces; • selecting GSI values from the workpieces that are greater than the GSI threshold value At least one first workpiece; and sending the at least one of the workpieces to a measuring machine for detection. 7. The method of sampling inspection of a workpiece according to claim 6, wherein the unified distance algorithm is selected from the group consisting of a Markov distance algorithm, a Euclidean distance algorithm and a central method. . 28 δ. If you apply for the Philip method, it also includes: The sampling test of the workpieces mentioned in the 6th item of the workpiece is obtained from the lure boring machine, and the measured values are based on the number of calendars, respectively. Under-measured value; based on the amount of the workpiece produced by the domain history system, the use of the group history to establish a estimation model, in which several Becco and these historical measurements are used to build the law; Based on a certain estimation and deduction, some historical quantity measurement is established. (4) * Method input method and the reference second system basis - refer to 'number heart history virtual = process parameter to the estimation mode' to calculate "several = process parameters To the reference mode's (4) predicting the distribution of the quasi-measured values and the overlap area between the historical parameters and the chores, resulting in a plurality of historical confidences # 'where the greater the overlap area, the higher the confidence index value , the higher the credibility of the representative virtual measurement values; 1 based on the historical virtual measurement values, the historical reference prediction values and the historical measurement values to calculate a confidence index threshold Value; ~ collect one of the production machines sent a plurality of workpiece process parameter data in the cassette; and $ input the process parameter data of the workpieces to the estimation mode, and calculate a plurality of virtual measurement values of the workpieces; 201135474 Some complex 1:= According to the reference mode, the calculation of the distribution of the predicted value of the work is calculated separately: the distribution between the reference and the value of the confidence indicator, and the area of the plurality of I pieces is the greater the distortion of the charm. The higher the higher the credibility of a virtual quantity depreciation, the higher the value of the threshold value: the value of the confidence indicator is less than the confidence finger-workpiece; and the workpiece is sent to the measuring machine to The at least one of the workpieces is inspected. The method, wherein the method of the sampling test of the workpiece described in Item 8 and the free-regressive method, wherein the push: the sample of the workpiece sampling test: neural network, 1 used back, - Simple regression network, one: - Obtained to the base-like neuro-regressive algorithm; one = data interpolation, - support generation partial least squares method and - generalized linear mode, flat method, - nonlinear π composition a group of people. 6, the method of sampling the inspection of the workpieces that are loaded into this brain type product. Where the requirements are 1 or