JP2009037306A - Monitoring method, monitoring program and monitoring system for production process - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a monitoring method, monitoring program and monitoring system for production process which have high prediction accuracy and can support identification of the cause of abnormality. <P>SOLUTION: In a monitoring system 1 for a production process, a production device 2 for producing a product and a monitoring device 4 in which a multiple regression model for predicting the performance of a product on the basis of production condition data of the production device 2 is stored are provided. The monitoring device 4 collects production condition data from the production device 2, applies wavelet transformation to the collected production condition data to obtain a pair of wavelet coefficients, and applies a part of the pair of wavelet coefficients to the multiple regression model to predict the performance of a product. If it is predicted that the performance of the product is abnormal, inverse wavelet transformation is applied to the part of the wavelet coefficients and the result is output to a terminal computer 5. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a manufacturing process monitoring method, a monitoring program, and a monitoring system, and more particularly, a manufacturing process monitoring method for predicting product performance by collecting manufacturing condition data from a product manufacturing apparatus and applying it to a multiple regression model. The present invention relates to a monitoring program and a monitoring system.

  In recent years, semiconductor manufacturing lines have introduced an FD (Fault Detection) system that collects manufacturing condition data during a manufacturing process and predicts the finished product. In this FD system, manufacturing condition data is collected from a manufacturing apparatus at regular time intervals, for example, every second, feature values such as average values in a predetermined time range are extracted from the collected data, and this feature value is subjected to multiple regression. It is applied to predictive models such as models to predict product performance. Then, an alarm is issued when the expected value of the finished product is out of a certain management range.

  The waveform of the manufacturing condition data, that is, the shape of the profile created by plotting the manufacturing condition data against time may greatly affect the product performance. However, in the above-described conventional method, even if the folding angle and manufacturing condition data are collected at regular intervals, time information is lost at the stage of dropping into feature values such as an average value. For this reason, there is a problem that the waveform information of the manufacturing condition data cannot be effectively used for prediction of workmanship and the prediction accuracy is low.

  On the other hand, Patent Document 1 discloses a technique for calculating wavelet coefficients by performing wavelet transform on operation chart data, creating a regression model using the wavelet coefficients, and performing analysis. However, this method has a problem that even if it is predicted that the product performance is abnormal, the cause of the abnormality cannot be specified.

JP 2004-288144 A

  An object of the present invention is to provide a manufacturing method monitoring method, a monitoring program, and a monitoring system, which have high prediction accuracy and can support specification of the cause of an abnormality.

  According to one aspect of the present invention, a step of collecting manufacturing condition data from a product manufacturing apparatus, a step of performing wavelet transform on the collected manufacturing condition data, obtaining a set of wavelet coefficients, Applying a part of the wavelet coefficients of the set to the multiple regression model to predict the product performance, and when the product performance is predicted to be abnormal, the set of wavelets The manufacturing process characterized by comprising: a step of performing an inverse wavelet transform with the values of some of the wavelet coefficients as they are, the rest of the wavelet coefficients being 0, and outputting the result thereof A method is provided.

  According to another aspect of the present invention, a computer collects manufacturing condition data from a product manufacturing apparatus, and performs a wavelet transform on the collected manufacturing condition data to obtain a set of wavelet coefficients. A procedure, applying a part of the wavelet coefficients of the set of wavelet coefficients to a multiple regression model to predict the product performance, and when the product performance is predicted to be abnormal, A step of performing an inverse wavelet transform on the set of wavelet coefficients, leaving the values of some of the wavelet coefficients as they are, setting the values of the remaining wavelet coefficients to 0, and outputting the results; A manufacturing process monitoring program is provided.

  According to still another aspect of the present invention, a manufacturing apparatus that manufactures a product, and a monitoring device that stores a multiple regression model that predicts the performance of the product based on manufacturing condition data of the manufacturing apparatus are provided. The monitoring device collects manufacturing condition data from the manufacturing device, performs wavelet transform on the collected manufacturing condition data to obtain a set of wavelet coefficients, and a part of the set of wavelet coefficients When the product performance of the product is predicted by applying the wavelet coefficients of the product to the multiple regression model and the product performance is predicted to be abnormal, the values of the partial wavelet coefficients Is left as it is, the value of the remaining wavelet coefficients is set to 0, inverse wavelet transform is performed, and the result is output. Process monitoring system is provided.

  According to the present invention, it is possible to realize a manufacturing process monitoring method, a monitoring program, and a monitoring system that have high prediction accuracy and can support specification of the cause of an abnormality.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.
FIG. 1 is a block diagram illustrating a manufacturing process monitoring system according to this embodiment.
As shown in FIG. 1, in the monitoring system 1 according to the present embodiment, a manufacturing apparatus 2 for manufacturing a product is provided. The manufacturing apparatus 2 is an apparatus that constitutes a part of a semiconductor manufacturing line, for example, and is, for example, a plasma CVD (Chemical Vapor Deposition) apparatus that forms a thin film on a silicon wafer. In the present specification, the term “product” refers to a member after a predetermined process is performed in the manufacturing apparatus, and includes an intermediate product in addition to a so-called finished product. For example, in the above example, a silicon wafer on which a thin film is formed is also included in the “product”.

  The monitoring system 1 is provided with a host computer 3. The host computer 3 outputs a command for causing the manufacturing apparatus 2 to execute the manufacturing process. Furthermore, the monitoring system 1 is provided with a monitoring device 4. The monitoring device 4 stores a multiple regression model that predicts the product performance based on the manufacturing condition data of the manufacturing device 2. The creation method and contents of the multiple regression model will be described later. Furthermore, the monitoring system 1 is provided with a terminal computer 5. The terminal computer 5 receives an alarm or the like output from the monitoring device 4 and displays it.

  Furthermore, the monitoring system 1 is provided with an MES (Manufacturing Execution System) LAN (Local Area Network) 6 and an FD (Fault Detection) system LAN 7. The manufacturing apparatus 2, the host computer 3, and the monitoring apparatus 4 are connected to the MES LAN 6. The monitoring device 4 and the terminal computer 5 are connected to the FD system LAN 7.

Next, a method and contents of creating a multiple regression model stored in the monitoring device 4 will be described.
FIG. 2 is a flowchart illustrating a method for creating a multiple regression model in the present embodiment.
FIG. 3 is a flowchart illustrating the waveform data shaping method according to this embodiment.
FIG. 4A is a graph illustrating waveform data with time on the horizontal axis and the value of manufacturing condition data on the vertical axis, and FIG. 4B is a graph illustrating time on the horizontal axis. Is a graph illustrating the result of wavelet transform of the waveform data shown in (a) by taking the value of the wavelet coefficient in FIG.
FIG. 5 is a flowchart illustrating a method for constructing a multiple regression model in the present embodiment.
FIG. 6 is a flowchart illustrating a method for narrowing explanatory variables by the stepwise method in the present embodiment.
FIG. 7 is a graph illustrating the influence of the number of explanatory variables in the multiple regression equation on the determination coefficient, with the number of explanatory variables on the horizontal axis and the determination coefficient on the vertical axis.

  First, as shown in step S <b> 1 of FIG. 2, the monitoring device 4 shown in FIG. 1 collects manufacturing condition data from the manufacturing device 2 via the MES LAN 6. The production condition data is, for example, temperature, pressure, current value, etc. in the plasma CVD process. The monitoring device 4 reads these manufacturing condition data at regular time intervals, for example, every second, over a time range from the start to the end of a series of processes in the manufacturing device 2.

  At this time, if the time interval for reading the manufacturing condition data is set sufficiently short with respect to the significant fluctuation period of the manufacturing conditions, the collected data constitutes a data group that is substantially continuous in time, and the manufacturing condition data The profile created by plotting the time with respect to time forms a certain shape. Hereinafter, such a profile shape is referred to as a “waveform”, and a group of manufacturing condition data that is substantially continuous in time is referred to as “waveform data”. Thereby, the continuously collected manufacturing condition data includes a plurality of waveform data corresponding to the processing chances of the manufacturing apparatus 2.

  Next, as shown in step S2, the waveform data is shaped and aligned to a certain length. The time required for a series of processes in the manufacturing apparatus 2 may differ for each processing chance. For example, in the above example, the processing time required to form a thin film with a predetermined thickness may vary from wafer to wafer. However, if the data amount of each waveform data is different, subsequent handling is difficult. Therefore, in step S2, the processing shown in FIG. 3 is executed to keep the amount of waveform data constant. FIG. 3 explains the process shown in step S2 of FIG. 2 in detail.

  That is, as shown in step S21 in FIG. 3, the standard data length of the waveform data is set to P pieces. However, P is an even number. In the present embodiment, P is set to 32, for example. Next, as shown in step S22, it is determined whether or not the data length of the collected waveform data exceeds P (32). If it exceeds, the process proceeds to step S23, and the (P + 1) th and subsequent data are discarded. As a result, the waveform data is corrected so that the portion exceeding the standard data length P is deleted and the data length becomes P pieces. Thereafter, the process proceeds to step S26.

  On the other hand, if the data length of the collected waveform data does not exceed P in step S22, the process proceeds to step S24, and it is determined whether the data length of the waveform data is less than P. If the number is less than P, the process proceeds to step S25 to copy the last data up to the Pth. That is, a copy (copy) of the end of the waveform data is added to the end of the waveform data, and the length of the added waveform data is P. As a result, the data length is corrected to be P pieces. Thereafter, the process proceeds to step S26. On the other hand, in step S24, if the data length of the waveform data is not less than P, the data length is P, so the process proceeds to step S26 without doing anything. Thereafter, in step S26, the original waveform data is replaced with the corrected waveform data. Thereby, the shaping of the waveform data is completed. FIG. 4A shows an example of waveform data after shaping. As shown in FIG. 4A, the waveform data after shaping is composed of 32 pieces of data collected every second.

  Next, as shown in step S3 of FIG. 2, wavelet transform is performed on the manufacturing condition data, that is, the waveform data shown in FIG. At this time, the mother wavelet is arbitrary, but for example, a Haar wavelet is used. Thereby, as shown in FIG. 4B, a set of wavelet coefficients is obtained.

  D1 in FIG. 4B indicates a wavelet coefficient having the highest frequency. These wavelet coefficients are calculated based on manufacturing condition data at two adjacent time points, and there are 16 in total. D2 indicates a wavelet coefficient having the next highest frequency. These wavelet coefficients are calculated based on manufacturing condition data at four consecutive time points, and there are a total of eight wavelet coefficients. Further, D3 and D4 indicate 4 and 2 wavelet coefficients calculated based on the production condition data at 8 and 16 consecutive points. Thus, each time the frequency level is lowered by one stage, the number of wavelet coefficients is substantially halved. And A1 has shown the remaining part remove | excluding the part corresponded to the above-mentioned wavelet coefficient from the original waveform data as two wavelet coefficients.

  In this way, a total of 32 wavelet coefficients are calculated from one waveform data consisting of manufacturing condition data at 32 time points. The 32 wavelet coefficients include all information of the original waveform data, and the original waveform data shown in FIG. 4A can be completely restored by performing inverse wavelet transform. is there.

  On the other hand, as shown in step S <b> 4, the monitoring device 4 collects product performance data from the manufacturing device 2. The workmanship data is an index characteristically representing the product manufacturing result, and is, for example, the film thickness of the film formed on the silicon wafer by the manufacturing apparatus 2.

  Next, as shown in step S5, a data set is formed by associating the wavelet coefficients calculated in step S3 with the performance data collected in step S4. That is, the workmanship data of a certain product is associated with the wavelet coefficient of the manufacturing condition data when this product is manufactured. For example, about several hundreds of such data sets are formed. Each data set includes, for example, 32 wavelet coefficients obtained from first manufacturing condition data, for example, temperature waveform data, and 32 manufacturing data obtained from second manufacturing condition data, for example, pressure waveform data. , The third manufacturing condition data, for example, 32 wavelet coefficients obtained from the current value waveform data, and one film thickness data.

Next, as shown in step S6, a multiple regression model is constructed using a plurality of, for example, several hundred data sets formed in step S5. The multiple regression model is constituted by a multiple regression equation as shown in the following Equation 1. In the following formula 1, x _k (k is an integer of 1 to n) is an explanatory variable, and a wavelet coefficient is used in the present embodiment. Further, a _k is a coefficient. On the other hand, y is an objective variable, and in the present embodiment, workmanship data is used.

  In the multiple regression model, the smaller the number of explanatory variables, the more stable the prediction model. However, in the above example, there are 96 explanatory variables (wavelet variables) for one objective variable (performance data), and the number of explanatory variables is excessive. Therefore, the number of explanatory variables is narrowed down by selecting wavelet coefficients having high correlation with workmanship by the following method. FIG. 5 explains the process shown in step S6 of FIG. 2 in detail.

  First, as shown in step S61 in FIG. 5, the number of explanatory variables is narrowed down in consideration of the characteristics of the manufacturing process. For example, in regard to temperature, when it is known in advance that the fluctuation of high frequency is noise and the influence on the film thickness is small, the wavelet coefficient D1 having the highest frequency among the wavelet coefficients of the temperature waveform data. Exclude coefficients. However, if there is no such knowledge, step S61 is omitted.

  Next, as shown in step S62, the number of explanatory variables is specified by the stepwise method. The stepwise method is a method of finding an optimum multiple regression equation by sequentially creating multiple regression equations while decreasing (or increasing) the number of explanatory variables one by one. Hereinafter, the processing content of step S62 will be described with reference to FIG. FIG. 6 explains step S62 of FIG. 5 in detail.

  First, as shown in step S621 in FIG. 6, the plurality of data sets are grouped into N groups. N is an integer of 2 or more. The grouping method is arbitrary. For example, five data sets that are temporally continuous are set as one set. That is, in the above-described example, one data set relating to the five silicon wafers processed subsequently is taken as one set.

  Next, as shown in step S622, an initial value of the number M of explanatory variables used in the multiple regression equation is set. The initial value of M is the maximum number of explanatory variables. For example, in the above example, M is set to 96 when the number of explanatory variables is not narrowed down in step S61. On the other hand, if some explanatory variables are excluded in step S61, the remaining number after the exclusion is used.

  Next, as shown in step S623, one group is selected from the N groups as a model evaluation group, and the remaining (N-1) groups are used as a model creation group.

  Next, as shown in step S624, a multiple regression equation is created using (N-1) sets of model creation groups.

Next, as shown in step S625, the multiple regression equation is created in step S624, by applying the wavelet coefficients of a group model evaluation, determine the coefficient of determination R ^2. Incidentally, the coefficient of determination R ² is an index indicating the degree of true multiple regression equation, a value of 0 to 1, the more true is good close to 1. R is a multiple correlation coefficient.

Next, as shown in step S626, the model evaluation group is changed to another group. Then, in step S627, it is determined whether the calculated coefficient of determination R ² all groups as a group for model evaluation. Then, when there are still groups that do not calculate the coefficient of determination R ² as a group for model evaluation, the process returns to step S624, the group as a group model evaluation, model preparation and the remaining (N-1) sets of the group as a group, calculates the coefficient of determination R ² creates a multiple regression equation.

Create a multiple regression equation all groups as a group for model evaluation, After calculating the coefficient of determination R ^2, the process proceeds from step S627 to step S628, the smallest value among the N coefficient of determination R ² calculated, this Let the coefficient of determination R be ² for M. Thus, for one value of the number M of explanatory variables, one decision factor R ² are determined.

Next, as shown in step S629, the value of the number M of explanatory variables is decreased by 1. And it progresses to step S630, and if M is not 0, it will return to step S623 and will repeat the process shown to step S623-S629. Thus, for all values of the number M of explanatory variables, the coefficient of determination R ² one each is determined.

As shown in FIG. 7, the relationship between the number M of explanatory variables and the determination coefficient R ² is a monotonically increasing function, and the value of the determination coefficient R ² increases as the value of M increases. However, as described above, when the value of M increases, the multiple regression model becomes unstable. For this reason, it is necessary to determine the optimum value of M. Therefore, as shown in step S631, the value of the coefficient of determination R ² identifies the number M of the smallest explanatory variables exceeds the reference value. This reference value is determined by the prediction accuracy required in the manufacturing process. In the example shown in FIG. 7, the reference value is set to 0.6. In this case, the value of the minimum of M, such as the coefficient of determination R ² exceeds the reference value is 18. In this way, the number M of explanatory variables is specified by the stepwise method.

  Then, as shown in step S63 of FIG. 5, the multiple regression equation constituted by the number M of explanatory variables specified in step S631 is adopted as the optimum multiple regression model. Thereby, a multiple regression model is constructed. This multiple regression model is stored in the monitoring device 4 (see FIG. 1).

Next, the operation of the monitoring system according to this embodiment shown in FIG. 1, that is, the manufacturing process monitoring method according to this embodiment will be described.
FIG. 8 is a flow chart illustrating a manufacturing process monitoring method according to this embodiment.
FIGS. 9A and 9B are graphs illustrating waveform data, with time on the horizontal axis and the value of manufacturing condition data on the vertical axis. FIG. 9A is collected from the manufacturing apparatus. (B) shows a waveform that has been wavelet transformed and then inverse wavelet transformed,
FIG. 10A illustrates the wavelet coefficients calculated by the wavelet transform, and FIG. 10B illustrates the wavelet coefficients used for the inverse wavelet transform.

  When manufacturing a product, the host computer 3 outputs an instruction for executing a manufacturing process to the manufacturing apparatus 2 via the MES LAN 6. Thereby, the manufacturing apparatus 2 performs a manufacturing process. For example, the manufacturing apparatus 2 forms a thin film on a silicon wafer by plasma CVD processing. At this time, the monitoring device 4 monitors the manufacturing process. Specifically, the following processing is executed.

  First, as shown in step S81 of FIG. 8, the monitoring apparatus 4 collects manufacturing condition data from the manufacturing apparatus 2 by intervening in communication between the manufacturing apparatus 2 and the host computer 3 via the MES LAN 6. To do. The manufacturing condition data is, for example, the temperature, pressure, current value, etc. in the plasma CVD process as described above. This data collection is performed at regular intervals, for example, every second. Thereby, manufacturing condition data is collected at a plurality of time points. These manufacturing condition data are waveform data as illustrated in FIG. In FIG. 9A, the manufacturing condition data of a product predicted to be normal (OK) and the manufacturing condition data of a product predicted to be abnormal (NG) in the subsequent process are shown.

  Next, as shown in step S82, the monitoring device 4 performs wavelet transformation on the waveform data of the collected manufacturing condition data to obtain a set of wavelet coefficients. The relationship between the waveform data and the wavelet coefficients is as shown in FIGS. 4 (a) and 4 (b). The wavelet coefficients are as shown in FIG. 10A, for example. In the example shown in FIG. 10A, a set of wavelet coefficients is calculated for one silicon wafer, and the set of wavelet coefficients includes 32 pieces for each manufacturing condition such as temperature, pressure, and current value. Wavelet coefficients (D1-1), (D1-2), ..., (D1-16), (D2-1), ..., (D2-8), (D3-1), ... , (D3-4), (D4-1), (D4-2), (A4-1), and (A4-2).

  Next, as shown in step S83, the monitoring device 4 selects a part of the wavelet coefficients narrowed down when creating the above-described multiple regression model from the set of wavelet coefficients obtained in step S82, and performs multiple regression. Apply to model. As a result, the workmanship is predicted. For example, the film thickness of a thin film formed by plasma CVD processing is predicted.

  Next, as shown in step S84, SPC (Statistical Process Control) is executed using the predicted value of the workmanship. That is, if the predicted performance value is within the control limit range, it is determined that the product performance is normal, and the process proceeds to step S88.

  On the other hand, if the predicted value of performance is outside the control limit range, it is determined that the performance of the product is abnormal, and the process proceeds to step S85. Then, the monitoring device 4 transmits a warning mail to the terminal computer 5 via the FD system LAN 7. As a result, a warning is output.

  Thereafter, the process proceeds from step S85 to step S86, and the monitoring device 4 performs inverse wavelet transform on the set of wavelet coefficients related to the product determined to be abnormal as described above. At this time, the value of each wavelet coefficient is used as it is for the wavelet coefficient applied to the multiple regression model, and the value is set to 0 for the wavelet coefficient not applied to the multiple regression model. For example, in the example shown in FIG. 10A, the wavelet coefficient (D1-1) for temperature, the wavelet coefficient (D1-2) for pressure, and the wavelet (D1-1) for current value are selected as multiple regression models. If the wavelet coefficient (D1-2) for temperature, the wavelet coefficient (D1-1) for pressure, and the wavelet coefficient (D1-2) for current value are not selected, the inverse wavelet When performing the conversion, as shown in FIG. 10B, the value of the selected wavelet coefficient is left as it is, and the value of the wavelet coefficient that is not selected is set to 0. Thereby, the waveform data is reproduced based only on the wavelet coefficient having high correlation with the workmanship. Hereinafter, such waveform data is referred to as “reproduced waveform data”.

  Next, as shown in step S87, the monitoring device 4 transmits the reproduction waveform data of the abnormal product to the terminal computer 5 via the FD system LAN 7. At this time, the reproduction waveform data of the normal product is also transmitted for comparison. The reproduced waveform data of the normal product is the result of performing the inverse wavelet transform on a part of the wavelet coefficients that are predicted to have a normal product performance. As the reproduction waveform data of the normal product, the data of the normal product manufactured most recently may be used, or data prepared and stored in advance may be used. The terminal computer 5 simultaneously displays the reproduction waveform data of the abnormal product and the reproduction waveform data of the normal product. For example, as shown in FIG. 9B, the reproduction waveform data profile of the abnormal product and the reproduction waveform data profile of the normal product are superimposed and displayed.

  FIG. 9B shows an example of reproduced waveform data displayed by the terminal computer 5. The reproduced waveform data shown in FIG. 9B is obtained by performing wavelet transformation on the waveform data shown in FIG. 9A and further performing inverse wavelet transformation, but only wavelet coefficients having high correlation with the workmanship are obtained. Since the inverse transformation is performed by using, the portion of the waveform data that has a great influence on the product performance is emphasized.

Then, the engineer in charge looks at the warning mail received by the terminal computer 5, recognizes the occurrence of the abnormality, and identifies the cause of the abnormality with reference to the reproduced waveform data shown in FIG. 9B. For example, in the example shown in FIG. 9B, it can be seen that the manufacturing condition data of the abnormal product is too low in the first half of the process and too high in the second half compared to the manufacturing condition data of the normal product. Note that, in the original waveform data shown in FIG. 9A, since the portion that greatly affects the workmanship is not emphasized, it is difficult to derive the above view from the original waveform data.
Then, it progresses to step S88 and returns to step S81 when monitoring is continued.

  The monitoring device 4 can carry out the series of monitoring operations shown in FIG. 8 by executing, for example, a program. In this case, the monitoring device 4 stores a program that causes a computer built in the monitoring device 4 to execute the following procedures (1) to (5).

(1) Procedure for collecting manufacturing condition data from the manufacturing apparatus 2 (step S81)
(2) Procedure for performing wavelet transform on the waveform data of the collected manufacturing condition data to obtain a set of wavelet coefficients (step S82)
(3) A procedure for predicting the product performance by applying some of the wavelet coefficients of the set of wavelet coefficients to the multiple regression model (step S83).
(4) Procedure for outputting an alarm when it is predicted that the product performance is abnormal (step S85)
(5) When it is predicted that the product performance is abnormal, inverse wavelet transform is performed on some wavelet coefficients to obtain the reproduced waveform data of the abnormal product, and output together with the reproduced waveform data of the normal product (Steps S86, S87)
This program outputs the data to the monitoring device 4 in the format in which the reproduced waveform data of the abnormal product and the reproduced waveform data of the normal product are superimposed and displayed on the terminal computer 5 in the procedure (5). You may let them.

Next, the effect of this embodiment will be described.
As described above, in this embodiment, the wavelet coefficient is calculated by performing wavelet transform on the manufacturing condition data without dropping the manufacturing condition data into the feature value such as an average value, and the wavelet coefficient is converted into a multiple regression model. Since the performance is predicted by application, the waveform information of the original manufacturing condition data can be used for the prediction of the performance. For this reason, the monitoring method according to the present embodiment has high prediction accuracy of workmanship.

  In the present embodiment, when the multiple regression model is constructed, the number of wavelet coefficients used as explanatory variables is narrowed down in order to improve the stability of the multiple regression model. At this time, by using the stepwise method, wavelet coefficients that have a small influence on the performance are excluded in order, so that the stability of the multiple regression model can be improved without greatly reducing the prediction accuracy.

Hereinafter, experimental results indicating the high prediction accuracy will be described.
First, a multiple regression model was constructed using the method of this embodiment for a manufacturing process in which three types of manufacturing condition data are output. At this time, the number of explanatory variables was 18. That is, 18 coefficients used for the multiple regression equation were selected from 96 wavelet coefficients by the above-described method. In addition, a multiple regression equation was created by (N-1) sets of model creation groups. Then, the determination coefficient R ² was obtained by using each of the (N-1) groups for model creation and the remaining one group for model evaluation.

On the other hand, as a comparative example, a multiple regression model was also constructed by a method using a conventional average value. At this time, in order to set the number of explanatory variables to 18, each of the three types of waveform data was divided into six time regions, and an average value was obtained in each region, and these were used as explanatory variables. Then, similarly to the method of the present embodiment described above, a multiple regression equation is created by the (N-1) sets of model creation groups, and the determination coefficient R is determined using each of the model creation group and the model evaluation group. ² was obtained. The calculation results of the determination coefficient R ² of the embodiment and the comparative examples shown in Table 1 below.

The coefficient of determination R ² obtained by the group for modeling shows the construction accuracy of the regression model. Further, the determination coefficient R ² obtained by the model evaluation group indicates the performance prediction accuracy, that is, the generalization ability. As shown in Table 1, the multiple regression model constructed in the present embodiment was superior in both construction accuracy and workmanship prediction accuracy compared to the multiple regression model of the comparative example. In particular, performance prediction accuracy was extremely excellent.

  Furthermore, according to this embodiment, when it is predicted that the workmanship is abnormal, the reproduction waveform data is output together with an alarm. Since the reproduced waveform data is generated by inversely transforming only wavelet coefficients having high correlation with the performance, the reproduced waveform data is waveform data in which a portion that greatly contributes to the performance is emphasized. For this reason, it is possible to effectively support the identification of the cause of the abnormality by the engineer in charge. As a result, the engineer in charge can easily identify the cause of the abnormality and take early measures.

  Furthermore, according to the present embodiment, the monitoring device 4 is connected to the MES LAN 6 and collects manufacturing condition data by intervening in communication between the host computer 3 and the manufacturing device 2. Thereby, it is not necessary to provide a dedicated terminal for connecting to the monitoring apparatus 4 in the manufacturing apparatus 2, and it is not necessary to install a dedicated program for communicating with the monitoring apparatus 4 in the manufacturing apparatus 2. As a result, the manufacturing process can be monitored without greatly changing the existing system. The monitoring device 4 transmits an alarm or the like to the terminal computer 5 via the FD system LAN 7. Thereby, the line capacity of the MES LAN 6 is not compressed.

  The present invention has been described above with reference to the embodiment. However, the present invention is not limited to this embodiment. For example, as long as a person skilled in the art adds, deletes, or changes a design, or adds, deletes, or changes a process as appropriate to the above-described embodiment, the gist of the present invention is included. , Within the scope of the present invention.

  For example, in the above-described embodiment, the monitoring apparatus 4 is connected to the manufacturing apparatus 2 via the MES LAN 6 and is connected to the terminal computer 5 via the FD system LAN 7. For example, the monitoring device 4 may be connected to these devices through a dedicated line. Further, in the above-described embodiment, an example in which the monitoring device 4 outputs alarm and reproduction waveform data to the terminal computer 5 and the terminal computer 5 displays such information has been shown, but the present invention is limited to this. For example, the monitoring device 4 may directly display the alarm and the reproduction waveform data, or may cause the printing device to print them out.

  Furthermore, in the above-described embodiment, an example in which the monitoring device 4 collects manufacturing condition data at regular time intervals has been described. However, the timing at which the manufacturing condition data is collected is not necessarily constant. Furthermore, in the above-described embodiment, an example in which a program for executing a series of monitoring operations is stored in the monitoring device 4 has been described. However, the monitoring device 4 has a dedicated program for performing the monitoring operations. Hardware may be provided. Furthermore, in the above-described embodiment, the example of monitoring the manufacturing process of the semiconductor manufacturing line has been shown. However, the present invention is not limited to this, and can be applied to the manufacturing process of any product.

It is a block diagram which illustrates the monitoring system of the manufacturing process which concerns on embodiment of this invention. It is a flowchart figure which illustrates the creation method of the multiple regression model in this embodiment. It is a flowchart figure which illustrates the shaping method of the waveform data in this embodiment. (A) is a graph illustrating waveform data by taking time on the horizontal axis and the value of the manufacturing condition data on the vertical axis, and (b) taking time on the horizontal axis and wavelet on the vertical axis. It is a graph which takes the value of a coefficient, and illustrates the result of having performed the wavelet transform of the waveform data shown to (a). It is a flowchart figure which illustrates the construction method of the multiple regression model in this embodiment. It is a flowchart figure which illustrates the narrowing-down method of the explanatory variable by the stepwise method in this embodiment. It is a graph which illustrates the influence which the number of explanatory variables of a multiple regression equation has on the determination coefficient, taking the number of explanatory variables on the horizontal axis and the determination coefficient on the vertical axis. It is a flowchart figure which illustrates the monitoring method of the manufacturing process which concerns on this embodiment. (A) and (b) are graphs illustrating waveform data, with time on the horizontal axis and the value of manufacturing condition data on the vertical axis, and (a) remains collected from the manufacturing apparatus. A waveform is shown, and (b) shows a waveform after wavelet transform and then inverse wavelet transform. (A) illustrates the wavelet coefficients calculated by the wavelet transform, and (b) illustrates the wavelet coefficients used for the inverse wavelet transform.

Explanation of symbols

1 monitoring system, 2 manufacturing equipment, 3 host computer, 4 monitoring equipment, 5 terminal computer, 6 MES LAN, 7 FD system LAN

Claims (13)

A process of collecting manufacturing condition data from the product manufacturing equipment;
Performing a wavelet transform on the collected manufacturing condition data to obtain a set of wavelet coefficients;
Applying some of the wavelet coefficients of the set of wavelet coefficients to a multiple regression model to predict the performance of the product;
When the product performance is predicted to be abnormal, with respect to the set of wavelet coefficients, the value of the part of the wavelet coefficients is left as it is, the value of the remaining wavelet coefficients is set to 0, and inverse wavelet transform is performed. Outputting the result; and
A method for monitoring a manufacturing process, comprising:   2. The manufacturing process according to claim 1, wherein in the step of outputting the result, a result obtained by performing an inverse wavelet transform on the partial wavelet coefficients that are predicted to be normal in product is also output. Monitoring method.   In the step of outputting the result, the result of the inverse wavelet transform for the part of the wavelet coefficient that is predicted to be abnormal, and the part of the wavelet coefficient that is predicted to be normal 3. The method of monitoring a manufacturing process according to claim 2, wherein the result of the inverse wavelet transform is displayed in a superimposed manner.   When the length of the collected manufacturing condition data is shorter than the standard data length, a copy of the last data in the manufacturing condition data is added to the last of the manufacturing condition data, and the manufacturing after the addition If the length of the condition data is the standard data length, and the length of the manufacturing condition data is longer than the standard data length, a step of deleting a portion of the manufacturing condition data that exceeds the standard data length is further included The method for monitoring a manufacturing process according to any one of claims 1 to 3, further comprising: On the computer,
A procedure for collecting manufacturing condition data from the product manufacturing equipment;
A procedure for performing wavelet transform on the collected manufacturing condition data and obtaining a set of wavelet coefficients;
Applying a portion of the wavelet coefficients of the set of wavelet coefficients to a multiple regression model to predict the performance of the product;
When the product performance is predicted to be abnormal, with respect to the set of wavelet coefficients, the value of the part of the wavelet coefficients is left as it is, the value of the remaining wavelet coefficients is set to 0, and inverse wavelet transform is performed. The procedure to output the result,
A manufacturing process monitoring program characterized in that   6. The procedure for outputting the result, the computer causing the computer to also output a result of performing an inverse wavelet transform on the partial wavelet coefficients that are predicted to have a normal product performance. The manufacturing process monitoring program described.   In the procedure of outputting the result, the computer is configured to output a result of inverse wavelet transform for the part of the wavelet coefficients that are predicted to be abnormal, and the part of which the result is predicted to be normal. 7. The manufacturing process monitoring program according to claim 6, wherein the result of the inverse wavelet transform for the wavelet coefficients is superimposed and displayed. On the computer,
When the length of the collected manufacturing condition data is shorter than the standard data length, a copy of the last data in the manufacturing condition data is added to the last of the manufacturing condition data, and the manufacturing after the addition If the length of the condition data is the standard data length, and the length of the manufacturing condition data is longer than the standard data length, a step of deleting a portion of the manufacturing condition data that exceeds the standard data length is further included 8. The manufacturing process monitoring program according to claim 5, wherein the manufacturing process monitoring program is executed. Production equipment for producing products;
A monitoring device storing a multiple regression model for predicting the product performance based on the manufacturing condition data of the manufacturing device;
With
The monitoring device collects manufacturing condition data from the manufacturing device, performs a wavelet transform on the collected manufacturing condition data to obtain a set of wavelet coefficients, and a part of the set of wavelet coefficients. The wavelet coefficients are applied to a multiple regression model to predict the product performance, and when the product performance is predicted to be abnormal, the values of the partial wavelet coefficients for the set of wavelet coefficients are A manufacturing process monitoring system, characterized in that the wavelet coefficient is left as it is, the inverse wavelet transform is performed with the value of the wavelet coefficient being zero, and the result is output.   The said monitoring apparatus, when outputting the result, also outputs a result of performing an inverse wavelet transform on the partial wavelet coefficient that is predicted to have a normal product performance. 9. The manufacturing process monitoring system according to 9. A host computer that outputs instructions to the manufacturing apparatus;
A first communication network connecting the manufacturing apparatus and the host computer;
A terminal computer for displaying the result;
A second communication network connecting the monitoring device and the terminal computer;
Further comprising
The monitoring device collects the manufacturing condition data from the manufacturing device via the first communication network, and outputs the result to the terminal computer via the second communication network. The manufacturing process monitoring system according to claim 9 or 10.   The terminal computer includes a result of an inverse wavelet transform for the part of the wavelet coefficients that is predicted to be abnormal in performance, and an inverse wavelet for the part of the wavelet coefficients in which the result is predicted to be normal. 12. The manufacturing process monitoring system according to claim 11, wherein the conversion result is displayed in an overlapping manner.   The monitoring device further adds a copy of the last data in the manufacturing condition data to the last of the manufacturing condition data when the length of the collected manufacturing condition data is shorter than a standard data length. If the length of the manufacturing condition data after addition is the standard data length and the length of the manufacturing condition data is longer than the standard data length, the length of the manufacturing condition data exceeds the standard data length. 13. The manufacturing process monitoring system according to claim 9, wherein the part is deleted.

Images