JP2009021348A - Abnormal factor identification method and system, program for causing a computer to execute the abnormal factor identification method, and computer-readable recording medium recording the program - Google Patents

Abstract

To provide an abnormality factor identification method capable of accurately identifying an abnormality factor even when a large-scale process data group obtained by a very long and complicated process such as a semiconductor product production process is targeted. To do.
At least one feature amount representing a feature of process data is extracted from a process data group (S5). The process data group is expressed by being converted into a feature space defined with the feature as an axis (S6, S7). Among the process data groups for defective products in the feature amount space, those having a predetermined relationship are classified into the same group (S8). For each classified group, the feature amount contributing to the distance from the reference point to the point represented by the process data for each defective product in the feature amount space is extracted (S9). Based on the extracted feature amount, an abnormal factor is identified.
[Selection] Figure 1

Description

  The present invention relates to an abnormality factor identification method and system, and more specifically, represents a manufacturing condition acquired from a production process when a defective product occurs in a production process having a plurality of manufacturing processes such as a production process of a semiconductor product. The present invention relates to a method and system for identifying an abnormal factor by analyzing process data using multivariate analysis.

  The present invention also relates to a program for causing a computer to execute such an abnormality factor identification method.

  The present invention also relates to a computer-readable recording medium recording such a program.

  In order to perform efficient production in a manufacturing process that produces a large amount of products, the process that causes the manufacturing process to be diagnosed as abnormal based on the inspection result of the manufactured product or the characteristic value of the manufactured product An important issue is to quickly return the manufacturing process to a normal state by specifying the data and resetting the specified process data to a normal value. However, when there are many manufacturing processes such as a semiconductor manufacturing process, the number of manufacturing variables that can cause anomalies is extremely large, and it is impossible for the quality manager to check the manufacturing variables one by one. Therefore, conventionally, for example, Patent Document 1 (Japanese Patent Publication No. 2002-525757) proposes a multidimensional method and system (hereinafter referred to as “prior art”) for statistical process management.

  In this prior art, as shown in FIG. 8, first, in step S101, the process data group in which each processing condition of each data observed in a certain process is recorded becomes a variable suitable for analysis. Conversion is performed to obtain a new variable (data conversion step). In this conversion, the scale of the variable is converted, for example, the average value of the variable is set to zero.

  Next, in step S102, the converted process data group is a process data group that is “in the management range” when a normal result is obtained in the process, and “out of management range” when an abnormality occurs in the process. To the process data group. These are hereinafter referred to as a normal process data group and an abnormal process data group, respectively.

  Next, in step S103, an abnormal process data group is mapped onto a space with each processing condition as an axis, and a vector connecting the center of the space and the mapped point is considered. In the above vector of the abnormal process data group, the absolute value of the cosine between the two vectors is used as the similarity, and the hierarchical ascending classification type automatic classification method is applied to indicate the direction in which the abnormal process data is concentrated and distributed. Extract (specify the cause direction). The extracted directions are designated as cause direction 1, cause direction 2,..., And abnormal process data is classified into any cause direction.

  Next, in step S104, the following two indexes (first index and second index) are calculated in order to identify each abnormality occurrence factor for the abnormal process data. The first index is the distance from the center of the normal process data group to the observation point (abnormal process data) in the cause direction, and is calculated as a scalar product of the vector of the abnormal process data group and the vector of the cause direction. This first index indicates to which cause direction the abnormal process data is strongly related. The second index is the angular proximity between the vector connecting the center of the normal process data group and the observation point (abnormal process data) and the vector in each cause direction, and is obtained as the cosine of the two vectors.

Then, in step S105, the cause direction of each data in the abnormal process data group and, in turn, which variable causes the abnormality are identified from the first index and the second index (identification of the abnormality factor).
JP-T-2002-525757

  As is known, in a very long and complicated process such as a production process of a semiconductor product, the number of items in the process data group is enormous.

  Here, in the above-described prior art, when the number of data items in the process data group, that is, the number of variables is enormous, the data is classified using only the directionality of the data in a high-dimensional space. . For this reason, it becomes a very difficult and inaccurate classification. The reason is that, even in a high-dimensional space, even a slight difference in data appears as a large distance difference by adding a small difference in each dimension. This kind of problem that makes it difficult to classify data in a high-dimensional space is known as the spherical concentration phenomenon (Introductory explanations (Kenichiro Ishii, Noriyoshi Ueda, Eisaku Maeda, Hiroshi Murase, “Easy to understand pattern recognition”) , Ohm, August 1998)).

  If the data classification accuracy deteriorates as described above, there is a problem in that the data when an abnormality occurs cannot be classified well for each tendency, and the abnormality factor cannot be accurately identified when a plurality of abnormality factors are mixed.

  As one countermeasure against the spherical concentration phenomenon, there is a technique of performing classification in a space in which the dimension of data is reduced by performing principal component analysis which is one type of multivariate analysis technique. However, in a very long and complicated production process such as a semiconductor manufacturing process, even if a dimension compression technique such as principal component analysis is applied to the process data group, the number of dimensions is, for example, 10 or more. For this reason, it is difficult to sufficiently improve the classification accuracy of abnormal factors.

  Furthermore, since the conventional technique employs a hierarchical classification method, there are too many calculation steps when analyzing a process data group obtained in a very long and complicated production process such as a semiconductor manufacturing process. There is also a problem.

  Therefore, an object of the present invention is to efficiently classify a tendency indicated by an abnormality even when a large-scale process data group obtained by a very long and complicated process such as a semiconductor product production process is targeted. It is an object of the present invention to provide an abnormality factor identification method and system capable of accurately identifying an abnormality factor.

  Moreover, the subject of this invention is providing the program for making a computer perform such an abnormal factor identification method.

  Moreover, this invention is providing the computer-readable recording medium which recorded such a program.

In order to solve the above problem, the abnormality factor identification method of the present invention is:
An abnormality factor identification method for identifying a factor of defective product generation in a process of performing one or more processes on an object and performing an inspection process for each object,
In the above process, one or more kinds of process data including process execution conditions for each target object are acquired in association with the inspection result of the inspection step, and the object is determined to be non-defective or defective by the inspection result. It is determined whether it is a non-defective product,
Extracting at least one feature amount representing the characteristics of the process data from a process data group formed by a plurality of the process data;
The process data group is expressed by converting it into a feature amount space defined with the feature amount as an axis, and the feature amount space is an average value after the conversion of the process data group for the non-defective product is a single reference point. Is set to be
In the feature amount space, those having a predetermined relationship among the process data groups for the defective product are classified into the same group,
For each of the classified groups, the feature amount contributing to the distance from the reference point to the point represented by the process data for each defective product in the feature amount space is extracted,
An abnormality factor is specified based on the extracted feature amount.

  Here, the “process data group” is, for example, a production process of a semiconductor product, such as conditions for forming a film, temperature for annealing, time, manufacturing conditions such as laser output, and electrical characteristics. All data representing process performance conditions such as intermediate inspection data can be included.

  According to the abnormality factor specifying method of the present invention, the process data group having the predetermined relevance among the process data groups for the defective product in the feature amount space is classified into the same group, whereby the process data group is abnormally trended. It can be classified by (that is, cause of abnormality). Subsequently, for each of the classified groups, feature amounts contributing to the distance from the reference point to the point represented by the process data for each defective product are extracted in the feature amount space. An abnormal factor is specified based on the extracted feature amount. In this way, according to this abnormality factor identification method, the abnormality factor is identified by classifying each group, that is, each abnormality tendency (abnormality cause), so even if a plurality of abnormality factors are mixed, the accuracy is improved. It is possible to identify abnormal factors well. Therefore, even when targeting a large group of process data obtained by a very long and complicated process such as a semiconductor product production process, the tendency of anomalies can be efficiently classified and the cause of anomalies can be accurately identified. Can be identified well.

  In the abnormality factor identification method according to an embodiment, the predetermined relationship for the process data for the defective product to be classified into the same group is that a distance in the feature amount space is a predetermined value or less. And

  In the abnormality factor identification method according to this embodiment, among the process data groups for the defective products, those having a short distance in the feature amount space are classified into the same group. Therefore, the process data group for the defective product can be efficiently classified for each abnormal tendency (cause of abnormality). Therefore, even if a plurality of abnormality factors are mixed, the abnormality factor can be specified with high accuracy.

  In the abnormality factor identification method according to an embodiment, the conversion into the feature amount space is performed by normalization so that an average of process data groups for the non-defective product is 0 and a standard deviation is 1. .

  In the abnormality factor identification method according to this embodiment, it is possible to treat and analyze all variables equally without being influenced by the difference in scale of each variable forming the feature amount space.

  The abnormality factor identification method according to an embodiment is characterized in that the production conditions obtained in the production process included in the process are used as the process data, and the cause of the abnormality that has occurred in the production process is identified.

  According to the abnormality factor identification method of this embodiment, when any abnormality occurs in the manufacturing process, it is possible to specify which manufacturing condition in the manufacturing process contributes to the occurrence of the abnormality.

  In the abnormality factor identification method according to an embodiment, the inspection result obtained in the inspection step is the quality of the manufactured product in the final inspection step of the process and the characteristic value of the manufactured product in the intermediate inspection step of the process. It is characterized by including the quality of.

  According to the abnormality factor identification method of this embodiment, not only the cause of the quality of the entire manufactured product by the final inspection process, but also the cause of the abnormality of the characteristic value of the manufactured product by the intermediate inspection process is specified. can do.

  In one embodiment of the abnormality factor identification method, the process is a semiconductor product production process.

  In general, the production process of semiconductor products is a very long and complicated process, and thus the obtained process data group is large-scale. According to the abnormality factor identification method of this embodiment, in the semiconductor product production process, the tendency of abnormality can be efficiently classified, and the abnormality factor can be identified with high accuracy.

  As a result, an operator (including a person in charge of maintenance) can immediately improve the manufacturing process in a production process in which one or more manufacturing processes are performed on a manufactured product. Therefore, improvement in production efficiency is realized.

  In the abnormality factor identification method according to an embodiment, an abnormality is detected for each process data based on a feature value contributing to the distance from the reference point to a point represented by the process data for each defective product in the feature value space. An evaluation index representing a suspicious degree as a factor is calculated, and abnormal factor candidates are presented based on the calculated evaluation index.

  In general, the production process of semiconductor products is a very long and complicated process, so there is not necessarily only one abnormal factor for one abnormal event, and several abnormal factors are acting. There is. Here, in the abnormality factor identification method of this embodiment, an evaluation index representing a suspicious degree as an abnormality factor is calculated for each process data, and an abnormality factor candidate is presented based on the calculated evaluation index. Therefore, it is possible to present in an easy-to-understand manner information how many of each abnormality factor are suspected as abnormality factors.

  In an abnormality factor identification method according to an embodiment, a principal component analysis is performed on the process data group, and the feature amount space is set with the obtained principal component as an axis.

  Generally, when a process data group includes many process variables and there is a strong correlation between the process variables, the subsequent analysis is adversely affected due to multicollinearity. Here, in the abnormality factor identification method according to this embodiment, principal component analysis is performed on the process data group, and the feature amount space is set with the obtained principal component as an axis. Therefore, it is possible to extract the feature amount in consideration of the correlation between the process variables, and it is possible to avoid the adverse effects described above.

  In the abnormality factor identification method according to an embodiment, in the feature amount space, a process for the defective product in a partial space defined with the first principal component and the second principal component obtained by the principal component analysis as axes. The data group is classified into the above groups.

  In the abnormality factor identification method according to an embodiment, a partial space defined with the first principal component and the second principal component obtained by the principal component analysis as axes, that is, a partial space defined with a salient feature amount as an axis. The process data group for the defective product is grouped. Therefore, it is possible to avoid the problem that it is difficult to classify data in a high dimension such as the spherical concentration phenomenon described above.

  In the abnormality factor identification method according to an embodiment, when the process data group for the defective product is classified into the group, a non-hierarchical clustering method using a distance between two points in the feature amount space as an evaluation index is used. It is characterized by.

  According to the abnormality factor identification method of one embodiment, the process data group for the defective product can be efficiently classified into data having similar characteristics of the process execution conditions, that is, for each abnormal tendency (abnormal cause).

In one embodiment of the abnormal factor specifying method, the contribution to the distance from the reference point in the feature space to a point represented by the process data for each defective, to T ² statistic Hotelling of each feature quantity It is calculated as a contribution.

  According to the abnormality factor identification method of this embodiment, it is possible to accurately determine the process variable contributing to the distance.

In the abnormality factor identification method according to an embodiment, the contribution of each feature value to the T ² statistic of the feature amount on the feature amount space for the non-defective product process data and the feature amount space for the process data of the defective product based on the significant differences between the contribution to T ² statistic Hotelling of each feature amount of the above and identifies the abnormal factor.

According to abnormal factor specifying method of this one embodiment, for the process variables do not follow a normal distribution, the significant difference between the time of normal process and process abnormal, T ² statistic Hotelling (Hotelling) of the process variable It is possible to appropriately evaluate the contribution to, that is, the suspicion as an abnormal factor candidate.

In the abnormality factor identification method according to an embodiment, for a process variable that gives certain process data, the contribution of the hoteling of each feature value on the feature value space to the T ² statistic for the process data of the defective product is predetermined. If it is smaller than the threshold, the process variable is excluded from the abnormal factor candidates.

According to the abnormality factor identification method of this embodiment, it is possible to avoid the problem that the evaluation value of the process variable becomes relatively high such that the contribution of the normal hoteling to the T ² statistic is extremely small. It becomes possible.

The abnormality factor identification system of this invention is
An abnormality factor identification system for identifying a factor of defective product generation in a process of performing one or more processes on an object and performing an inspection process on each object,
In the above process, one or more kinds of process data including process execution conditions for each target object are acquired in association with the inspection result of the inspection step, and the object is determined to be non-defective or defective by the inspection result. It is determined whether it is a non-defective product,
A storage unit for storing the process data and the inspection result in the inspection process in association with each object,
An abnormality factor identification unit that executes the abnormality factor identification method according to any one of claims 1 to 13 using the storage contents of the storage unit.

  In the abnormality factor identification system of this invention, a memory | storage part associates and memorize | stores the said process data and the test result of the said test process for every said target object. The abnormality factor identification unit executes the abnormality factor identification method using the stored contents of the storage unit. Therefore, process data that should originally be detected as an abnormal factor can be detected from a huge number of process data, and the process data that is the cause of defective product generation can be accurately identified.

  The abnormality factor identification program of the present invention is a program for causing a computer to execute the abnormality factor identification method of the present invention.

  According to the abnormality factor identification program of the present invention, it is possible to cause a computer to execute the abnormality factor identification method of the present invention.

  The recording medium of the present invention is a computer-readable recording medium in which the abnormality factor identification program of the present invention is recorded.

  According to the recording medium of the present invention, it is possible to cause the computer to execute the abnormality factor specifying method of the present invention by causing the computer to read the recorded content of the recording medium.

  Hereinafter, the present invention will be described in detail with reference to the illustrated embodiments.

  FIG. 7 shows a block configuration of an abnormal factor identification system (generally indicated by reference numeral 10) of one embodiment of the present invention for identifying the factor of defective product generation in the semiconductor production process 100.

  The semiconductor production process 100 executes one or more manufacturing steps A, B, C,... On a manufactured product (wafer in this example) as an object, and is performed between the manufacturing steps A, B, C. The intermediate inspection process E and the final inspection process F performed after the manufacturing processes A, B, C,... Are performed. In this example, the intermediate inspection process E includes quality determination of the characteristic value of the manufactured product, and the final inspection process F includes final quality determination for the entire manufactured product.

  In this example, each of the manufacturing processes A, B, and C is performed by a plurality of manufacturing apparatuses A1, A2, and A3; B1, B2, and B3; C1, C2, and C3 capable of performing the processes. Implemented in parallel. From the same viewpoint, the intermediate inspection process E is performed by a plurality of inspection apparatuses E1, E2, and E3, and the final inspection process F is performed by a plurality of inspection apparatuses F1, F2, and F3. In this example, for the sake of simplicity, each process is performed by three apparatuses. However, naturally, it may be performed by one apparatus, or may be performed by two apparatuses or four or more apparatuses.

  The abnormality factor identification system 10 includes a database 20 as a storage unit, an abnormality factor identification unit 30, an input device 50 for an operator to input to the abnormality factor identification unit 30, and the abnormality factor identification unit 30. And a display device 60 for displaying the processing result.

  The database 20 acquires one or more types of process data 12 including manufacturing conditions for each manufactured product from each manufacturing process A, B, C,... Of the production process, and from each inspection process E, F. Inspection data 13 representing the inspection result (including the pass / fail judgment result) of the manufactured product is acquired. Then, for each manufactured product, the process data 12 and the inspection data 13 for the manufactured product are associated (linked) and stored. The database 20 is configured by a known hard disk drive device, for example.

  Note that the “process data group” is a production process of a semiconductor product as in this example, conditions for film formation, manufacturing conditions such as annealing temperature, time, laser output, and electrical It includes all data representing process performance conditions such as intermediate inspection data such as characteristics.

  The abnormality factor identification unit 30 executes the abnormality factor identification method described later using the process data 12 and the inspection data 13 accumulated in the database 20, and obtains process data that is a factor (abnormal factor) of defective product generation. Identify. In this example, the abnormality factor specifying unit 30 includes a computer that executes arithmetic processing according to software (program). The processing contents and processing results by the abnormality factor identifying unit 30 are output to the display device 60.

  The display device 60 is composed of, for example, an LCD (Liquid Crystal Display) or a CRT (Cathode Ray Tube), and displays information output by the abnormality factor specifying unit 30.

  The input device 50 includes, for example, a known keyboard or mouse.

  An operator (including a person in charge of maintenance) instructs the abnormality factor identification unit 30 to execute processing by the input device 50 while setting the processing conditions while viewing the screen displayed on the display device 60. be able to.

  FIG. 1 shows a schematic flow of an abnormality factor identification process executed by the abnormality factor identification unit 30.

  First, in step S1, process data (process data group) is acquired from the database 20.

  Next, in step S2, the non-defective product data and the defective product data are separated based on whether each process data forming the process data group is a value within the management range. Note that the criterion for determining whether or not each process data is a value within the management range may coincide with the pass / fail determination result represented by the inspection data 13 or may be determined separately.

  In this example, a process data group as shown in FIG. 2 will be described. The horizontal axis in FIG. 2 represents an identification number for identifying a manufactured product, and the vertical axis represents a certain characteristic value Y obtained by the inspection process. The process data group when the characteristic value Y is within the management range from −5 to +5 is defined as “non-defective product data”, and the process data group when the characteristic value Y is outside the management range is defined as “defective product data”. "

  Next, in steps S3 and S4 in FIG. 1, matrices are determined based on the obtained good product data and defective product data, respectively.

とする（ここで、Ｎはデータ数、Ｐは変数の数を表す。）。そのうち良品データの行列Ｘｇを

とする（ここで、Ｎｇは正常な特性のデータ数を表す。）。また、不良品データの行列Ｘｂを

とする（ここで、Ｎｂは特性不良となったデータ数を表す）。 Here, the data matrix X created by the entire target process data group is

Where N represents the number of data and P represents the number of variables. Of these, the matrix Xg of good data

(Here, Ng represents the number of data of normal characteristics). Also, the matrix Xb of defective product data

(Nb represents the number of data with defective characteristics).

とする。不良品データの行列Ｘｂについては、行列Ｘｇの標準化に使用したＸｇの各項目についての平均値と標準偏差値とを用いて標準化し、得られた規格化行列を

とする。 The matrix Xg of the non-defective product data is standardized using the average value and the standard deviation value for each item of the matrix Xg, and the obtained normalized matrix is

And The matrix Xb of defective product data is standardized using the average value and the standard deviation value for each item of Xg used for standardization of the matrix Xg, and the resulting normalized matrix is

And

  Next, in step S5, in order to determine a principal component space as a feature amount space, a principal component analysis is performed to extract a principal component axis as a feature amount axis.

Specifically, principal component analysis is performed as follows. First, the singular value decomposition is performed on the normalization matrix of good product data to obtain the following equation (Equation 6).

In the above equation (Equation 6), U and V are orthogonal matrices. S is a diagonal matrix, and singular values s _r are arranged in descending order in the diagonal elements. R is the number of main components employed. U _R is the submatrix whose elements through R-th column of U, is S _R is a partial matrix, V _R is the submatrix whose elements through R-th column of V whose elements to R rows R column of S .

Next, in step S6, a principal component score Tg _R corresponding to the coordinates of the non-defective product data in the feature amount space is calculated by the following equation (Expression 7).

On the other hand, in step S7, the principal component score Tb _R corresponding to the coordinates of the defective product data in the feature amount space is calculated by the following equation (Equation 8).

  In the matrices (principal component scores) calculated by the above equations (Equation 7) and (Equation 8), the first column is the first principal component, the second column is the second principal component,. R is the main component.

Further, in steps S6 and S7 described above, the principal component scores Tg _R and Tb _R calculated by the above equations (Equation 7) and (Equation 8), respectively, as illustrated in FIG. The two principal components are plotted in the principal component space with the horizontal axis and the vertical axis, respectively.

  Generally, when a process data group includes many process variables and there is a strong correlation between the process variables, the subsequent analysis is adversely affected due to multicollinearity. Here, in the above example, the principal component analysis is performed on the process data group, and the feature amount space is set around the obtained principal component. Therefore, it is possible to extract the feature amount in consideration of the correlation between the process variables, and it is possible to avoid the adverse effects described above.

  In the principal component space as the feature amount space, the average value of the non-defective product data is a single reference point, that is, the origin of the principal component space. In addition, since the standard deviation of non-defective data is standardized to be 1, all variables can be handled and analyzed equally without being affected by the difference in the scale of each variable in this principal component space. It becomes.

  In FIG. 3, the data plotted in the vicinity have similar characteristics, and it is estimated that the characteristics are poor due to the same factors.

  Therefore, in step S8 of FIG. 1, clustering (grouping) is performed on the points plotted on the feature amount space.

  Specifically, in this example, clustering (grouping) is performed by a k-means method which is one of known non-hierarchical clustering methods on the subspace spanned by the first principal component and the second principal component. carry out. That is, among the defective product data, those having a short distance in the principal component space are classified into the same group. In the example of FIG. 3, as a result of clustering, the defective product data is classified into abnormality groups 1 to 4, respectively. Each data belonging to the abnormal group 1 is represented by a symbol ●, each data belonging to the abnormal group 2 is represented by a symbol x, each data belonging to the abnormal group 3 is represented by a symbol Δ, and each data belonging to the abnormal group 4 is represented by a symbol +.

  By this clustering, the process data group can be classified by abnormality tendency (that is, abnormality cause). In this example, defective product data is defined in a partial space defined with the first principal component and the second principal component obtained by principal component analysis as axes, that is, in a partial space defined with a significant feature amount as an axis. Group. Therefore, it is possible to avoid the problem that it is difficult to classify data in a high dimension such as the spherical concentration phenomenon described above. In this example, since the non-hierarchical clustering method is used, the defective product data can be efficiently classified into data having similar characteristics of the process execution conditions, that is, for each abnormal tendency (abnormal cause).

  Next, in step S9 of FIG. 1, a significant difference between the non-defective product data and the defective product data is calculated in order to identify the abnormality factor.

  Specifically, the following processing is performed for each classified abnormality group.

Here, the analysis is performed using the abnormality group 1 as an example. The process data matrix Xb and principal component score matrix Tb _R of defective products classified into the abnormal group 1 are replaced with the following equations (Equation 9) and (Equation 10), respectively.

Covariance matrix sigma _T of the principal component scores T _R is calculated by the following formula (11).

を用いて、ホテリングのＴ^２統計量は次式（数１３）によって算出される。

Here is the element of the matrix sigma _T

Using, T ² statistic Hotelling is calculated from the following equation (Equation 13).

と定義される。ここで、ｔは主成分得点からなるベクトル、つまりＴ_Ｒの行ベクトルである。 Furthermore, the contribution of the pth variable to the T ² statistic is

Is defined. Here, t is the row vector of the vector of principal component scores, i.e. T _R.

は主成分得点の共分散行列Σ_Ｔの逆行列である。ｘ_ｐは

の変数ｐに関するデータであり、スカラー量である。νｐは第ｐ変数に関する係数ベクトルの転置ベクトルであり、Ｖ_Ｒのｐ行と等しい。

Is the inverse matrix of the covariance matrix Σ _T of the principal component scores. x _p is

Is the data regarding the variable p, and is a scalar quantity. νp is the transpose vector of the coefficient vector related to the p-th variable, equal to p rows of V _R.

The calculated contribution of the p-th variable to the T ² statistic is a ratio of the p-th variable contributing to the distance from the origin in the feature amount space.

The operation for calculating the contribution of the p-th variable to the T ² statistic is performed for the non-defective product data and the defective product data using the main component score of the non-defective product data and the main component score of the defective product data.

と表す。この第ｐ変数についてのＴ^２統計量への寄与をＸｇの全データについて算出し累計すると、

となる。さらに、この（数１８）を、Ｘｇの全データ、全変数についてのＴ^２統計量への寄与の合計で除算して規格化すると、良品データのＴ^２統計量への規格化された寄与は、

として算出される。 Next, the contribution to the T ² statistic for the p-th variable of the i-th data of the matrix Xg, that is, the contribution to the T ² statistic calculated for the non-defective data is

It expresses. When the contribution to the T ² statistic for this p-th variable is calculated for all the data of Xg and accumulated,

It becomes. Furthermore, the (number 18), all the data of Xg, when normalized by dividing by the sum of the contribution to T ² statistic for all variables, the contribution which is normalized to T ² statistic good data ,

Is calculated as

と表す。この第ｐ変数についてのＴ^２統計量への寄与をＸｂの全データについて算出し累計すると、

となる。さらに、この（数２１）を、Ｘｂの全データ、全変数についてのＴ^２統計量への寄与の合計で除算して規格化すると、不良品データのＴ^２統計量への規格化された寄与は、

として算出される。 Similarly, the contribution to the T ² statistic for the p-th variable of the i-th data in the matrix Xb, that is, the contribution to the T ² statistic calculated for the defective product data is

It expresses. When the contribution to the T ² statistic for this p-th variable is calculated and accumulated for all the data of Xb,

It becomes. Furthermore, the contribution of this (Equation 21), all the data Xb, when normalized by dividing by the sum of the contribution to T ² statistic for all variables, which are normalized to T ² statistic defective data Is

Is calculated as

In this way, the normalized contribution of the good product data to the T ² statistic (Equation 19) and the normalized contribution of the defective product data to the T ² statistic (Equation 22) are obtained.

At this time, a significant difference between the normalized contribution of the good product data to the T ² statistic (Equation 19) and the normalized contribution of the defective product data to the T ² statistic (Equation 22) is evaluated. against the variable value of the contribution to T ² statistic abnormality is small, when that captures minute difference between the time and the abnormal normal (differences are not deemed essential by influence of outliers) This is excluded because it is possible. A threshold function for exclusion is given by the following equation (Equation 23).

The divergence degree index IoK (Index of Kairido) is defined by the following equation (Equation 24) using the above (Equation 19), (Equation 22), and (Equation 23).

This divergence degree index IoK is an evaluation index that represents a substantial significant difference between the contribution of the non-defective product data to the T ² statistic and the contribution of the defective product data to the T ² statistic.

  FIG. 4 shows an example in which the divergence index IoK is calculated for each item (parameter) of the process data and arranged in descending order using the calculated divergence index IoK. In FIG. 4, items whose divergence degree index IoK is 1 or more are candidates for abnormal factors. A candidate with a larger value of the divergence degree index IoK is more suspicious as an abnormality factor or represents a greater contribution to the abnormality.

  In this way, by calculating the divergence degree index IoK as an evaluation index, it is possible to appropriately evaluate information indicating how suspicious some abnormal factors are as abnormal factors, and to provide easy-to-understand information. .

  Next, in step S10 of FIG. 1, a process variable (parameter) having a large value of the divergence index IoK is specified as an abnormal factor.

  FIG. 5 shows the occurrence frequency (histogram) of non-defective / defective products for the process variable P5 listed at the top as the abnormal factor candidate in FIG. The horizontal axis in FIG. 5 represents values that the process variable P5 can take. From FIG. 5, when the process variable P5 has a relatively low value, it is confirmed that the occurrence frequency of defective products is high. Thereby, a decrease in the value of the process variable P5 is specified as a factor that the abnormal group 1 is away from the coordinate center (origin) in FIG.

  For reference, FIG. 6 shows a histogram corresponding to FIG. 5 when the above grouping (S8) is not performed. From FIG. 6, when no grouping is performed, it is difficult to determine a significant difference between the non-defective product manufacture and the defective product manufacture, and it is difficult to determine that the decrease in the process variable P5 value is an abnormal factor. I can say that.

  As described above, this abnormality factor identification method contributes to the distance from the origin to the point represented by the process data for each defective product in the principal component space as the feature amount space for each group clustered in step S8 described above. The extracted feature amount is extracted, and the abnormal factor is specified based on the extracted feature amount. That is, since the abnormality factor is specified by classifying each abnormality tendency (abnormal cause), even if a plurality of abnormality factors are mixed, the abnormality factor can be specified with high accuracy. Therefore, even when targeting a large group of process data obtained by a very long and complicated process such as a semiconductor product production process, the tendency of anomalies can be efficiently classified and the cause of anomalies can be accurately identified. Can be identified well.

  In addition, you may construct | assemble as a program for making a computer perform the above-mentioned abnormal equipment estimation method.

  Further, such a program may be recorded and distributed on a computer-readable recording medium such as a CD-ROM. By installing the program in a general-purpose computer, the abnormal equipment estimation method can be executed by the general-purpose computer.

  The method of the present invention can be used not only in a production process of a semiconductor product or the like but also in all processes that can obtain a process data group that is a process execution condition. Since the present invention can be carried out if a process data group that is the process execution condition is obtained, the present invention is not limited to the manufacture of chemicals, pharmaceuticals, agricultural products, iron making, plastic materials, etc. Applicable in a wide range of fields such as provision (banking, insurance, consulting). The present invention is particularly advantageous in the analysis of process data having many data items and the analysis of process data including a plurality of abnormal events.

It is a figure which shows the schematic flow of the abnormality factor identification method of one Embodiment of this invention. It is a graph which shows the tendency of the characteristic value Y which is a result of a process. It is the figure obtained by plotting the process data group when abnormality occurred in the process in the feature amount space. It is a figure which shows the extraction result of the abnormal factor candidate by the evaluation index employ | adopted by the said abnormal factor identification method. It is a figure which shows the histogram about normal time data and the abnormal group 1. FIG. It is a figure which shows the histogram about normal time data and abnormal time data. It is a figure which shows the structural example of the abnormality factor identification system of one Embodiment of this invention. It is a figure which shows the analysis processing flow of a prior art.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Abnormal factor identification system 12 Process data 13 Inspection data 20 Database 30 Abnormal factor identification part

Claims (16)

An abnormality factor identification method for identifying a factor of defective product generation in a process of performing one or more processes on an object and performing an inspection process for each object,
In the above process, one or more kinds of process data including process execution conditions for each target object are acquired in association with the inspection result of the inspection step, and the object is determined to be non-defective or defective by the inspection result. It is determined whether it is a non-defective product,
Extracting at least one feature amount representing the characteristics of the process data from a process data group formed by a plurality of the process data;
The process data group is expressed by converting it into a feature amount space defined with the feature amount as an axis, and the feature amount space is an average value after the conversion of the process data group for the non-defective product is a single reference point. Is set to be
In the feature amount space, those having a predetermined relationship among the process data groups for the defective product are classified into the same group,
For each of the classified groups, the feature amount contributing to the distance from the reference point to the point represented by the process data for each defective product in the feature amount space is extracted,
An abnormality factor identification method, wherein an abnormality factor is identified based on the extracted feature amount. In the abnormality factor identification method according to claim 1,
The abnormality factor identification method characterized in that the predetermined relevance for the process data of the defective products to be classified into the same group is that a distance in the feature amount space is a predetermined value or less. In the abnormality factor identification method according to claim 1,
The abnormality factor identification method characterized in that the conversion into the feature amount space is performed by normalization so that an average of process data groups for the non-defective product is 0 and a standard deviation is 1. In the abnormality factor identification method according to claim 1,
An abnormality factor identification method characterized by identifying the cause of an abnormality that has occurred in the manufacturing process using the manufacturing conditions obtained in the manufacturing process included in the process as the process data. In the abnormality factor identification method according to claim 4,
The inspection result obtained in the inspection step includes the quality of the entire manufactured product by the final inspection step of the process and the quality of the characteristic value of the manufactured product by the intermediate inspection step of the process. Abnormal factor identification method. In the abnormality factor identification method according to claim 5,
An abnormality factor identifying method, wherein the process is a production process of a semiconductor product. In the abnormality factor identification method according to claim 6,
Based on the feature amount contributing to the distance from the reference point to the point represented by the process data for each defective product in the feature amount space, an evaluation index representing the suspicious degree as an abnormal factor is calculated for each process data. And presenting an abnormality factor candidate based on the calculated evaluation index. In the abnormality factor identification method according to claim 1,
A method of identifying an abnormal factor, characterized in that a principal component analysis is performed on the process data group, and the feature amount space is set with the obtained principal component as an axis. In the abnormality factor identification method according to claim 8,
Classifying process data groups for the defective products into the groups in the partial space defined by using the first principal component and the second principal component obtained by the principal component analysis in the feature amount space. Characteristic abnormality factor identification method. In the abnormality factor identification method according to claim 1,
A non-hierarchical clustering method using a distance between two points in the feature amount space as an evaluation index when classifying process data groups for the defective products into the groups. In the abnormality factor identification method according to claim 1,
Abnormalities and calculating a contribution from the reference point in the feature space the contribution to the distance to the point representing the process data for each defective, the T ² statistic Hotelling of each feature quantity Factor identification method. In the abnormality factor identification method according to claim 11,
Contribution of hoteling of each feature value on the feature value space to the T ² statistic with respect to the process data of the non-defective product, and T of hoteling of each feature value on the feature value space with respect to the process data of the defective product ^2. An abnormality factor identification method characterized by identifying the abnormality factor based on a significant difference between the contribution to the statistics. In the abnormality factor identification method according to claim 12,
The process variable to provide a certain process data, if the contribution to T ² statistic Hotelling of each feature amount on the feature space for the process data of the defective is smaller than the predetermined threshold value, the process variable Is excluded from the abnormal factor candidates. An abnormality factor identification system for identifying a factor of defective product generation in a process of performing one or more processes on an object and performing an inspection process on each object,
In the above process, one or more kinds of process data including process execution conditions for each target object are acquired in association with the inspection result of the inspection step, and the object is determined to be non-defective or defective by the inspection result. It is determined whether it is a non-defective product,
A storage unit for storing the process data and the inspection result in the inspection process in association with each object,
An abnormality factor identification system comprising: an abnormality factor identification unit that executes the abnormality factor identification method according to any one of claims 1 to 13 using the storage contents of the storage unit.   An abnormality factor identification program for causing a computer to execute the abnormality factor identification method according to claim 1.   A computer-readable recording medium on which the abnormality factor identification program according to claim 15 is recorded.

Images