JP2009187175A - Analysis device of batch process data, and abnormality detection/quality estimation device using the same - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a data analysis device which can accurately check correlation of variables which differ in sampling periods or variables spreading over processes by handling data collectively and in a unified manner even if batches are mutually correlated and the data are different in sampling periods or spread over different processes. <P>SOLUTION: Batch process data 1 and quality data 4 are input into a model creation means 10 from a data storage means 3. The batch process data 1 and quality data 4 which have been input are input into a data analysis means 6 through a data input means 5. In the data analysis means 6, modeling is carried out by application of a multi-way multivariate analysis method. In a data determination/prediction means 9, operation is conducted and based on a "model" (coefficient matrix) stored in a model storing means 8 and the batch process data 1 which are collected online by a data collection means 2, alarm and a predicted value are output based on a conducted result of the mathematical operation, and abnormality detection or quality estimation is performed. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to an apparatus for analyzing batch process data and an abnormality detection / quality estimation apparatus using the same.

  FIG. 10 is a diagram showing an outline of the configuration of a manufacturing data management system in a conventional product manufacturing system using a general plant. For example, in food plants, sensors, etc. are installed to measure the temperature, pressure, flow rate, etc. of tanks, and the information measured by the sensors etc. is distributed via DCS (Distributed Control System) (distributed control system) via controller 500 and network 510. Control system) 520 for control. The upper level of DCS520 acquires and stores operation data and quality data such as sensor information, manages operation results, and makes production plans, etc. (Manufacturing Execution System) ) 530 is provided. In MES530, each batch / lot (hereinafter also simply referred to as 'batch') manages which device is passed through and the route time, etc., thereby linking batches across processes ( Non-patent document 1).

  When quality data is acquired, depending on the type of plant, it is brought into a so-called “lab” facility equipped with analytical equipment, and quality measurement 540 by measuring the number of microorganisms and concentration measurement 550 using an analyzer Inspection is often performed. In many cases, quality measurement and inspection in a “lab” are time consuming and costly, and only a part of the sampling inspection is performed.

  FIG. 11 is a diagram showing a conventional manufacturing process of a certain product (food). This manufacturing process includes a plurality of processes, and two processes, for example, “process A” 560 and “process B” 570 are selected and described in detail. In the manufacturing process shown in FIG. 11, it is assumed that products are manufactured and processed in “batch” units. That is, one batch is input into each process shown in FIG. 11, and the result (intermediate product) is output after the processing / processing in the process is completed. The output intermediate product is also input to the next process in the same batch unit, and this is repeated for each process.

  In step A560, the intermediate product manufactured in the previous step is processed in the tank for about 1 hour, for example. Next, the intermediate product, which is the manufacturing result of step A560, is input to step B570. In the process B570, another processing for about 2 hours is performed in another tank. The processing time in each of the process A560 and the process B570 is determined by the processing amount of the intermediate product. The tank used in both the process A560 and the process B570 is provided with a temperature sensor and a pressure sensor, and the temperature and pressure during the processing time of each batch in each process are measured at a constant sampling cycle. The measurement data is stored in the database 580. Further, the number of microorganisms in the intermediate product obtained as a result of the process A560 and the process B570 is also measured, and similarly stored in the database 580 as an element for evaluating the intermediate product. In addition, FIG. 11 also shows temperature and pressure data of the process A560 and the process B570 when a certain batch (referred to as batch 1) is passed through the manufacturing process shown in FIG. These temperature and pressure data are referred to as “batch process data”. The count of the number of microorganisms measured to obtain quality data for the batch is also shown here only at one location.

  Time series data of temperature and pressure measurement values of process A560 and process B570 of many batches (batch 1 to N) and microbial count measurement values as quality data are obtained. Defect (normal / abnormal) is judged. However, since it takes time to measure the number of microorganisms, it also takes time to determine whether it is abnormal or normal. Therefore, a device (system) that determines (detects) the normality / abnormality of the intermediate product of the batch from the temperature and pressure measurement data at step A560 and step B570 before actually counting the number of microorganisms is used. Has been. Such a device (system) is referred to herein as an “abnormality detection device”.

  Anomaly detection and quality estimation by applying multivariate analysis methods such as Principal Component Analysis (PCA) and Partial Least Squares (PLS) to the so-called continuous data of the plant It has been performed conventionally (see Non-Patent Document 2).

  Continuous data is data collected at a fixed timing for a set of constant variables (for example, a set of temperature, pressure, etc.) as a set of samples, and “variables” x “samples” in which multiple sets of samples are collected 2D data structure. In general, the sample collection timing is often the same for all variables, but for some variables, a time with a certain time difference (delayed) may be used. In general, the collection timing does not have to be equal (sampling period), and it is sufficient that each sample is prepared.

  In the following, a conventional analysis procedure (model creation) will be described by taking multivariate analysis by principal component analysis (PCA) as an example. Note that x and y are data after normalization (standardization) of measurement data to an average value of 0 and a standard deviation of 1, in general. Assuming that the original two-dimensional data is X, the two-dimensional data X is represented by a two-dimensional determinant, and for each element X (i, j) of the matrix, i represents a sample number and j represents a variable number. Then, the original data X is represented by the following equation 1 by principal component analysis (PCA).

X = T * P ^T + E (1)
Here, T is a principal component score, P is a loading (loading amount) matrix, E is a modeling error matrix, and T on the right shoulder of P represents transposition of the matrix. The principal component score T is obtained by collecting the original variables X into variables that are correlated with each other and collecting them as a smaller number of “principal components”. Is required.

T = X * P (2)
In this way, the original variables can be aggregated into principal components using the matrix P. The matrix P itself is generally called “anomaly detection model”.

The principal component score T is an aggregated principal component, and in detail, a principal component vector (principal component score vector) t calculated from a horizontal vector x _k corresponding to each sample (each row of X) of the data matrix X _k (horizontal vector) is calculated by t _k = x _k P.

Using the principal component vector t _k (lateral vector), the T ² statistic (T ² value) for each sample x _k of the data matrix X is calculated as follows.
The singular value of X is s _i (i = 1 ... N, N is the number of columns of P), and t _k is an N-dimensional horizontal vector, so its i-th element is t _{k, i}

The Q statistic for x _k for (Q value) is calculated by the following from x _k. The Q statistic represents how far the correlation between variables in the original data x deviates.

Q = x _k (I-PP ^T ) (I-PP ^T ) ^T x _k ^T (4)
T ² statistic and Q statistic always takes a positive value, but normal (usually) when often takes a large value at the time of abnormality despite falls within a certain range. Therefore, the threshold value that defines the boundary between normal and abnormal when the T ² statistic and Q statistic for each sample (corresponding to each row of X) calculated from the existing data matrix X is determined, and x (known measurement for data) and new measurement data (described later x _{new new),} it determines that an abnormality has occurred in the case where T ² statistic and Q statistic has taken a value that exceeds the threshold value (T ² statistic and Q statistic For the amount, see Non-Patent Document 2). As this determination method, for example, a value including 95% of the data can be automatically set for each of the T ² statistic and the Q statistic for each sample of X calculated from X.

Here, X is assumed to accumulate past data, and P is a matrix calculated from it, but new data (the same length as the number of X columns (horizontal length)) using the matrix P with respect to the vector side) x _new, t _new calculating the t _new = x _new P represents a main component that aggregates the contrast x _{new new.}

The T ² statistic and Q statistic for x _new calculated from x _new and t _new indicate how far x _new is from the past data group X. If this value is large, abnormal data and Abnormality detection can be performed by determining.

A vector x _new (I-PP ^T ) obtained in the process of calculating the Q statistic represents a “contribution plot”. This represents the “degree of abnormality” (how far away from normal time) each element of this vector corresponds to the corresponding element of the original vector x _new . It is also possible to specify input data to be

  Independent Component Analysis (ICA) is a new multivariate analysis method that has been studied recently. Independent component analysis (ICA) can be applied to anomaly detection as well as principal component analysis (PCA). (See Non-Patent Document 4). In this way, the management of the process abnormality detection using the statistics of the plant process measurement data and the upper and lower limits thereof is called statistical process management (SPC). In particular, the statistical process management by multivariate analysis as in this example is performed. It has come to be used.

  For each sample (row) x of X, if there is quality data corresponding to this, set it as a vertical vector y, and apply partial least squares to make x and y as follows: Thus, the matrices P, Q, and W related to minimize the error E can be obtained (see Non-Patent Document 3).

Y = XW (P ^T W) ^-1 Q + E (5)
(Where E is the error vector)
Partial Least Squares (PLS) is known as an influential method that can create a stable model even if there is multi-co-linearity between each variable of input data X. ing.

If new data x _new is obtained from these matrices, an estimated value y _e of quality data corresponding to the new data x _new can be obtained as follows.
y _e = x _new W (P ^T W) ^-1 Q (6)
(Where y _e is the quality data estimate for x _new )
Quality estimation can be performed based on Equation 6 above.

  Discriminant analysis is a method for creating a model that estimates output data y from input data X when the quality variable y is expressed as a binary value, such as good / bad or ○ ×. Are known.

  The above explanation is about the multivariate analysis method by Principal Component Analysis (PCA) for 2D data of "sample" x "variable", but the multivariate analysis method is applied to batch process data. It is also known that batch / lot abnormality detection and quality estimation are performed (see Patent Document 1). Batch / lot data has multiple variables for each of multiple batches, and further includes time-series data for each variable, so the data structure is three-dimensional: "batch" x "variable" x "time" It becomes data.

  In general, multivariate analysis methods such as principal component analysis (PCA) and partial least square method (PLS) are basically applied to two-dimensional data of “variable” × “sample”. Therefore, when applying a multivariate analysis method to three-dimensional data such as batch / lot data, a method called a “multiway method” (see FIG. 12) is conventionally used. This is because, for 3D data such as batch / lot, “variables” × “time” in each batch are rearranged in one dimension, and the entire data is set as a new “sample”. The entire variable included in “” is defined as “variable” and converted into “variable” × “sample” two-dimensional data, and then the multivariate analysis method is applied.

FIG. 12 is a diagram for explaining a conventional multiway method. As shown in FIG. 12, the three-dimensional data of the batch process is expressed as follows.
x3 (i, j, k) (7)
Here, i = 1... I is a batch, j = 1... J is a variable in the batch, k = 1... K is a time series (time) index of the variable in the batch.

  In general, in a batch process, not all batches have the same length in time series. Therefore, as shown in Non-Patent Document 5, a process called “data length adjustment” or “data alignment” is used to process the time series lengths of all batches. Yes.

3D data x3 (i, j, k) for 3D data x3 (i, j, k) (time in batch, variable number j, batch number i) after aligning batch time series length The procedure for making the two-dimensional data will be described below.
Focusing on one batch i of the batch process data shown in the left part of FIG. 12, the sets of variable J of each time K of three-dimensional data: x3 (i, j, k) are rearranged one-dimensionally (lower part of FIG. 12) (Refer downward from the top of the two-dimensional data shown in the above). When the rearranged result is x2 (i, l), i = 1... I, l = j + (k-1) * J or l = k + (j-1) * K. Here, when l = 1... L, it can be expressed as L = J * K.

  When l = j + (k-1) * J, the integer part of l / J (the largest integer not exceeding l / J) is k-1 for l (1≤l≤L) L− (k−1) * J is j. If l = k + (j-1) * K, the integer part of l / K (the largest integer not exceeding l / K) is j- with respect to l (1 ≤ l ≤ L). 1, l− (j−1) * K is k. This gives a one-to-one correspondence between l and (j, k).

  For each batch, that is, for the batch i = 1... I, the above-described one-dimensional rearrangement procedure is executed to generate a “sample” (see the lower part of FIG. 12). As a result, x2 (i, l) becomes two-dimensional data in which i is an index of “sample” and 1 is an index of “variable”.

The above-mentioned principal component analysis (PCA) is used to create a feature quantity calculation model from only (input) data. Using this method, a feature quantity calculation model is obtained from the two-dimensional data x2 (i, l). Will be calculated. (In other words, anomaly detection can be performed using the principal component analysis described above and the T ² and Q statistics calculated from it.)
On the other hand, the partial least square method (PLS) is used to create an estimation model of output data from input data. When this method is used, input data is generally three-dimensional data as described above, but output data. In many cases, one-dimensional or two-dimensional data is originally determined for one batch. Specifically, when the quality data for one batch is one type, it is one-dimensional data, and when the quality data for one batch is two or more types (for example, the number and concentration of microorganisms), it is two-dimensional data. An estimation model based on the partial least square method, that is, an estimation model of output data, can be obtained based on the input data and the output data two-dimensionalized as described above. Further, when the output data is three-dimensional data, it is possible to obtain an estimated model of the output data by converting the data into two dimensions and applying the partial least square method in the same manner as described above.

  The method of applying multivariate analysis methods such as Principal Component Analysis (PCA) and Partial Least Squares (PLS) after converting 3D data to 2D data in this way is "Multiway Principal Component Analysis" It is called a “way partial least square method” or the like (hereinafter referred to as a multi-way multivariate analysis method) and is known to those skilled in the art.

FIG. 13 is a diagram illustrating an example of abnormality detection using a conventional multiway multivariate analysis technique. The conventional method based on this multi-way multivariate analysis can be applied only to data having a common sampling period in one process. That is, as shown in FIG. 13, a statistical process is performed by calculating indices such as a Q statistic and a T ² statistic by multiway principal component analysis (PCA) for only the data of process A560 or process B570. Anomaly detection was performed using statistical process control (SPC) 590, but the correlation between both process A560 and process B570 (for example, “Production conditions (temperature and pressure) in process A560 to product quality If there is a correlation between the two, there is no relationship between the “impact” and the “influence between the manufacturing conditions (temperature and pressure) in the process B570 on the product quality”. (Abnormality detection) could not be performed.
Japanese Patent Laid-Open No. 10-228312 Naoki Higashiya, Mitsuhiro Nakamura, “Traceability Support System for Batch Processes”, Vol.7, No.7, pp81-84, 2004 Manabu Kano "Statistical Process Management with Process Chemometrics" System / Control / Information, Vol. 7, No. 13, pp.1-6, 1996 Miyashita, Sasaki “Chemometrics Chemical Pattern Recognition and Multivariate Analysis” Computer Chemistry Series 3 Kyoritsu Publishing 1995 Noboru Murata Introduction “Independent Component Analysis” Tokyo Denki University Press 2004 “Comparison of Methods for Handling Unequal Length Batches”, IFAC DYCOPS-5, 1998, pp.66-71,

  As in the conventional method shown in FIG. 13, it is not sufficient to consider only the production conditions in a single process and the influence on product quality, and the accuracy may be reduced by abnormality detection and quality estimation based on this. It has become clear in recent years that is high.

  Conventionally, each sample for obtaining time-series data of variables in each batch of batch process data handled in the multi-way multivariate analysis method (instead of “sample” in which one batch is 1 sample) Each sample) had to be collected at equal intervals (sampling period). However, not all variable data is collected in the same sampling period in a batch process. Generally, data collected at different sampling periods depending on the characteristics of variables, and data related to various "events". There is also data that can be obtained (determined) only for one batch, such as the processing amount of the batch and the attributes of the equipment that processed the batch.

  Further, in detecting abnormality of process data, estimating quality, etc., there is a demand to perform abnormality detection and quality estimation from as much information as possible, and to analyze / extract factors affecting abnormality and quality. Comprehensive analysis of process data based on a large amount of information reveals correlations (reciprocal relationships) between a wide range of variables, and abnormality detection and quality estimation are performed based on that correlation to detect the accuracy of abnormality detection and quality estimation. May be improved.

  Even if variables with the same sampling period are grouped and the multivariate analysis method is applied individually within each group, it will result in anomaly detection and quality estimation using only variables within the range handled individually within each group. The accuracy of abnormality detection and quality estimation is low, and the relationship between variables with different sampling periods is not understood even in the factor analysis that affects the abnormality of the range handled individually in each group and the quality of the quality, Furthermore, there are problems such as being complicated because the multivariate analysis method must be applied individually for each sampling period.

  Usually, batches are processed in a plurality of processes and processed into final products, and a mechanism (see Non-Patent Document 1) that can extract data across the processes for each batch by linkage is established. Qualitative analysis, such as displaying data across processes and viewing the relationship, is a factor that affects abnormalities and quality of each variable between processes A method for quantitatively evaluating the analysis of the causal relationship is not proposed.

  Therefore, the present invention has been made to solve the above problems, and batches are linked to each other, and even if the data has a different sampling cycle or spans different processes, they are unified. It is an object to provide a data analysis apparatus capable of accurately examining the correlation between variables with different sampling periods and variables across processes and an abnormality detection / quality estimation apparatus using the same. is there.

  The batch process data analysis apparatus of the present invention is a batch process data analysis apparatus that handles time series data with different sampling periods for each batch as a group of process data grouped in batch units. After performing the data alignment process, the batch data extraction means for extracting the batch data associated with each batch from each group, and the multi-way method for each group from the batch data extracted by the batch data extraction means 1-dimensionally by means of a group data one-dimensionalization means for the data extracted from all groups, and after all the groups are one-dimensionalized by the group data one-dimensionalization means, Combine these grouped one-dimensional data 1-dimensional data array means for making an array of 1-dimensional data as a whole, and each element of the 1-dimensional data for each batch obtained by the overall 1-dimensional data array means “variable”, 1-dimensional data (vector) A data two-dimensionalization means for obtaining two-dimensional data of “variable” × “sample” with one “sample”, and a multivariate analysis method for the two-dimensional data obtained by the data two-dimensionalization means And multivariate analysis processing means for applying the analysis processing.

  The batch process data analysis apparatus according to the present invention is a batch process data analysis apparatus that handles a plurality of process data related to each batch as a group of process data in batch units, After performing the data alignment process, batch data extraction means for extracting batch data associated with each batch from each group, and batch data extracted by the batch data extraction means, each group is subjected to a multi-way method. Group data one-dimensionalization means for performing one-dimensionalization on the data extracted from all groups, and one-dimensionalization for all groups by the group data one-dimensionalization means, and then all the groups Combining these one-dimensional data of An entire one-dimensional data array means for arraying dimensional data, and each variable of the one-dimensional data for each batch obtained by the entire one-dimensional data array means is “variable”, and one-dimensional data (vector) is one “ Applying a multivariate analysis method to the two-dimensional data obtained by the two-dimensional data obtained by the two-dimensional data obtained by the two-dimensional data of “variable” × “sample” as “sample”, Multivariate analysis processing means for executing analysis processing.

The batch process data abnormality detection apparatus of the present invention generates an abnormality detection model using the data analysis apparatus described above, and performs data alignment processing and two-dimensionalization processing on newly acquired batch process data. And performing a matrix operation with the abnormality detection model, outputting T ² statistics and Q statistics, and detecting whether these exceed a preset threshold value to detect an abnormality.

  The batch process data quality estimation apparatus of the present invention generates a quality estimation model using the above-described data analysis apparatus, and performs data alignment processing and two-dimensionalization processing on newly acquired batch process data. Then, a matrix operation with the abnormality detection model is performed, and a quality data estimated value is output to perform quality prediction.

  According to the present invention, for batch process data relating to a plurality of batches, even if the sampling period is different, or even if the data spans different processes, they are handled uniformly and the variables that have different sampling periods, or variables that span processes The correlation between each other can be examined accurately.

  In addition, according to the present invention, by using the above-described batch process data analysis apparatus, it is possible to accurately investigate the correlation between variables having different sampling periods or variables across processes, so that abnormality detection accuracy / quality prediction accuracy can be obtained. Can be greatly improved.

The best mode for carrying out the present invention will be described below with reference to the drawings.
FIG. 1 is a block diagram showing a schematic configuration of a batch process data analysis apparatus and an abnormality detection / quality estimation apparatus using the same according to an embodiment of the present invention.

  In FIG. 1, the batch process data analysis apparatus and the abnormality detection / quality estimation apparatus using the same according to the embodiment of the present invention store the batch process data 1 collected by the data collection means 2 in the data storage means 3. The data storage means 3 stores the quality data 4 if it exists, and then uses it for data for quality prediction. These data are input from the data storage means 3 to the model creation means 10. The model creation means 10 includes a data input means 5, a data analysis means 6 and a data output means 7. The batch process data 1 and quality data 4 input to the data input unit 5 are input to the data analysis unit 6 via the data input unit 5. The data input means 5 performs data alignment processing as appropriate and outputs it to the data analysis means 6. The data analysis means 6 performs modeling by applying the above-described multiway multivariate analysis method. This will be described later.

  The result analyzed by the data analysis means 6 is output as a “model” (coefficient matrix) from the data output means (result output means) 7 to the model storage means 8 where the result output is stored. The batch process data analysis apparatus is configured with the above configuration. Details will be described later.

  The data determination / prediction unit 9 performs an operation based on the “model” (coefficient matrix) stored in the model storage unit 8 and the batch process data 1 collected online by the data collection unit 2. Alarms and predicted values are output based on the results of the calculations performed, and abnormality detection or quality estimation is performed. Therefore, the batch process data abnormality detection apparatus or quality estimation apparatus is configured using the batch process data analysis apparatus described above.

  FIG. 2 is a flowchart showing the control flow and data flow of the batch process data analysis apparatus and the abnormality detection / quality estimation apparatus using the same according to the embodiment of the present invention shown in FIG. It shows an example when applied to. The same components as those in FIG. 1 will be described with the same numbers. Here, a case where principal component analysis (PCA) is used as multivariate analysis will be described.

  Plant data (batch process data) 1 collected by the data collection means 2 is stored in a data storage means (plant data database) 3. The (abnormality detection) model creation means 10 reads a plurality of batch process data sets (hereinafter referred to as “groups”) associated with each batch in a plurality of past batches from the plant data database 3 (see FIG. 5). ). Then, the data alignment processing 11, the two-dimensional processing 12, and the multivariate analysis processing 13 by principal component analysis are performed to create a (abnormality detection) model (specifically, coefficient matrix values and threshold values) 14, and model storage means Store in (model database) 8. The control flow and data flow thus far (that is, the left half of FIG. 2) are performed offline and are the control flow and data flow as the batch process data analysis apparatus according to the embodiment of the present invention.

When abnormality detection is performed using the model (coefficient matrix value and threshold value) stored in the model storage means 8, the abnormality detection means (for the newly acquired plant data (batch process data) 1) ( 20 corresponds to the data judgment / prediction means 9 in FIG. 1 and performs a data alignment process 21 and a two-dimensional process 22 and then performs a calculation with a matrix as a model in a matrix calculation process 23. ² Output 24 statistics and Q statistics, and perform abnormality detection based on this. That is, when the threshold value set in the model is exceeded, it is determined as “abnormal” and an alarm or the like is displayed on a display device (not shown). The control flow and data flow so far (that is, the right half of FIG. 2) are performed online, and the batch process data analysis apparatus according to the embodiment of the present invention shown in the left half of FIG. 2 is used. It is the flow of control and data flow which perform abnormality detection processing.

As described above, the anomaly detection from the T ² statistic and the Q statistic has already been described in the description of the conventional principal component analysis. The threshold is set manually by the user from existing data or, as described above, for example, a value that includes 95% of the data for each T ² statistic and Q statistic for each sample. Automatically done as a threshold.

  FIG. 3 is a flow diagram showing a control flow and a data flow of the batch process data analysis apparatus and the abnormality detection / quality estimation apparatus using the same according to the embodiment of the present invention shown in FIG. It shows an example when applied to. The same components as those in FIG. 1 will be described with the same numbers. Here, a case where a partial least square method (PLS) is used as multivariate analysis will be described.

  Plant data (batch process data) 1 collected by the data collection means 2 is stored in a data storage means (plant data database) 3. (Quality estimation) The model creation means 10 obtains a plurality of batch process data sets ("groups") associated with each batch in the past plurality of batches from the plant data database 3, and the quality corresponding to each batch. Data 4 is read (see FIGS. 5 and 11). Then, a data alignment process 11, a two-dimensional process 12, and a multivariate analysis process 13 using a partial least square method are performed to create a (quality estimation) model (specifically, coefficient matrix value) 15, and model storage means ( (Model database) 8. The control flow and data flow thus far (that is, the left half of FIG. 3) are performed offline and are the control flow and data flow as the batch process data analysis apparatus according to the embodiment of the present invention.

  When the quality estimation is performed using the model (coefficient matrix value) stored in the model storage unit 8, the quality estimation unit (batch process data) of FIG. (Corresponding to the data judgment / prediction means 9) 30 performs the data alignment process 31 and the two-dimensionalization process 32, and then performs an operation with the matrix as a model in the matrix operation process 33, and the quality data estimated value 34 Is output. The output quality data estimated value 34 can be displayed on a display device (not shown) or the like. The control flow and data flow so far (that is, the right half of FIG. 3) are performed online, and the batch process data analysis apparatus according to the embodiment of the present invention shown in the left half of FIG. 3 is used. It is the flow of control and data which perform quality estimation processing.

  Here, “data alignment” and “two-dimensionalization processing” in the description of FIGS. 2 and 3 will be described with reference to the flowchart of FIG. At the start of the description of the processing flow of FIG. 4, the data of each batch belonging to each group is data related to the same batch linked to each other (see FIG. 5), but the length of the time series is not uniform for the batch. In this case, “data alignment” is executed in order to align the lengths of the time series of all batches (see Non-Patent Document 5). This process first aligns the length of the batch time series.

After performing the data alignment process on each batch process data set (group) as described above, the following process is performed on each batch.
In batch loop processing (step S1), batch data associated with each batch is extracted from each group.

  Next, in the loop processing of each batch group (step S2), each group from the extracted batch data is made one-dimensional by a multi-way method as shown in FIG. 12 in step S3. This is performed on data extracted from all groups.

  Next, after all groups are made one-dimensional, in step S4, these one-dimensional data of all groups are combined to form one-dimensional data (array) as a whole (see FIG. 6).

By performing the above processing for all the batches (end of step S1 in FIG. 4), one-dimensional data can be obtained for each batch (see FIG. 7).
In this way, the “variable” × “sample” two-dimensionalization process in which each element of the one-dimensional data for each batch is “variable” and the one-dimensional data (vector) is one “sample” is completed.

  Then, as shown in Fig. 2 and Fig. 3, multivariate analysis methods such as principal component analysis and partial least squares are applied to the two-dimensional data that has been two-dimensionalized. can do. This is a processing outline of the batch process data analysis apparatus according to the embodiment of the present invention.

  Next, details of processing contents of the batch process data analysis apparatus according to the embodiment of the present invention will be described. Now, as shown in FIG. 5, for a plurality of batch data linked to each other, time series variables having different sampling periods of the same batch are grouped into a variable group. FIG. 5 illustrates a state in which the batches associated with the batch index i = 2 are linked to each other.

  This creates a group of variables for each of a plurality of sampling periods. Each data for the g-th group “group g” of variables is represented as four-dimensional data as follows.

x4 (i, g, j, k) (8)
Where g = 1 ... G is a group, i = 1 ... I is a batch, j = 1 ... Jg is a variable in the group in the batch, k = 1 ... Kg is a time series of variables in the group in the batch ( Each time). Although the number of batches is common to the groups, the number of variables and the length of the time series (sampling time points) are different for each group, so Jg and Kg and a subscript g are added respectively.

  Next, the four-dimensional data represented by the above equation 8 is converted into “samples” as shown in FIGS. Enable detection and quality estimation.

  Specifically, for each batch i, a vector in which a set of variables at each time is arranged one-dimensionally from each group is first created (see the lower part of FIG. 6). Furthermore, a vector in which this is arranged for all groups is created. This becomes one “sample” and is represented as two-dimensional data x2 (i, m) (see the lower part of FIG. 7). If it demonstrates in detail, said m will be represented like the following formula | equation 9.

Conversely, when m (1 ≦ m ≦ M) is given, g, j, and k are obtained as follows.

  Now, the above M is expressed by Equation 10,

For m = 1... M, an integer satisfying the following expression 11 is defined as g.

Next, an integer part (maximum integer not exceeding Expression 12) of Expression 12 below is defined as (k−1).

Equation 9 above is transformed into Equation 13 to obtain j from Equation 13.

By making this one-dimensional for each i, two-dimensional data x2 (i, m) is obtained (see FIG. 7). Here, i = 1... I is an index of “sample”, and m = 1... M (M is defined by Equation 10) is an index of “variable”.

  G, j, and k are uniquely determined by the above-described equations 9 to 13, and a one-to-one correspondence is attached between (g, j, k) and m. Thus, the multivariate analysis method is applied to the two-dimensional data x2 (i, m) where each i is an index of “sample” and m is an index of “variable”. As a means for applying the multivariate analysis method to the two-dimensional data x2 (i, m), a model creation means 10 including the data analysis means 6 shown in FIG. 1 is provided. Then, as the output of the model creation means 10, the created model is stored in the model storage means 8, the model stored in the model storage means 8, for example, the above-described abnormality detection model, and the actual batch process data collected online. 1 is performed, and the data determination / prediction means 9 performs abnormality detection by calculating the feature amount for each batch.

  When a quality estimation method such as a partial least square method is applied, the quality data 4 is stored in the data storage means 3 and the output is a variable (of a different sampling period), as in the case of the conventional technique described above. In the case of the above, the model (quality estimation model) is created by applying the partial least square method etc. by making the same two-dimensionally by the model creation means 10 shown in FIG. The model is stored in the model storage means 8, and calculation is performed on this model (quality estimation model) and the actual batch process data 1 collected online, and the data determination / prediction means 9 Quality estimation is performed by calculating feature values.

  By the way, when the data over the processes of each batch is obtained for a plurality of batches, these are the batch data for each process, and each is obtained as three-dimensional data. Here, there may be a variable acquired at a plurality of different sampling periods in each step.

  These are linked (linked) to batches (see Fig. 5), regardless of whether they crossed processes or within the same process or at different sampling periods. It can be expressed as data. In other words, the batch data obtained in each step for the common batch i can be defined as four-dimensional data x4 (i, g, j, k) as a “group”.

  Then, it is two-dimensionalized by the methods described in the above formulas 9 to 13 and FIGS. 6 and 7, and a multivariate analysis method such as principal component analysis or partial least square method is applied to the two-dimensional data as described above. can do. Specifically, the batch process data obtained in each step is set as each group in FIGS. 6 and 7, and the above-described processing can be performed by making the two-dimensional data as shown in FIG. Moreover, the concrete structure which performs these processes can use what was shown in FIGS.

  As described above, for multiple batch data, all of these data are related (linked) to batches even if they cross processes or have different sampling periods within the same process. Each group is represented as 3D data, and this is defined as 4D data x4 (i, g, j, k) as a “group” for the batch data obtained in each process for a common batch i. Therefore, a plurality of batch data can be collectively treated and converted into two-dimensional data of “sample” × “variable” by the above-described equations 9 to 13 and the method described with reference to FIGS. 6 and 7. . Apply various multivariate analysis methods such as principal component analysis, partial least squares, and independent component analysis to 2D data of "sample" x "variable" according to the nature of batch process data in the plant. Can do.

  In other words, if principal component analysis or independent component analysis is applied to 2D data of "sample" x "variable", anomaly detection is performed, and quality estimation is performed if partial least squares or discriminant analysis is applied. These are applied as appropriate according to the nature of the batch process data in the plant.

  FIG. 8 is a diagram showing an example when the method of the present invention is applied. In the example shown in FIG. 8, by taking both data of the process A60 and the process B70 as input data, it is not understood that only the process A60 and the process B70 are seen. Detection and quality estimation can be performed. Specifically, the data of the process A60 may be applied as “Group 1” and the data of the process B70 as “Group 2” shown in FIG. In the process of the abnormality detection apparatus according to the method of the present invention, a plurality of processes are combined and an SPC (statistical process management) 90 by the multiway method PCA (principal component analysis) is performed.

  FIG. 9 is a diagram comparing an example when the method of the present invention is applied and an example when the conventional method is applied. In the actual case where the conventional method shown in FIG. 9 is applied, in the normal batch, the vicinity of the end of the processing of the pressure data in the process A560 is flat, but sometimes it fluctuates. There may be a case where quality is deteriorated as a result of such fluctuation, and a normal product is produced. On the other hand, even near the end of processing of the pressure data in step B570, it is usually flat, but may occasionally change.

  The data of these parts of the process A560 and the process B570 are actually related to each other, and if they change in the process A560, if they also change in the process B570, the influence in the process A560 is canceled in the process B570 and is abnormal (defective). It has been confirmed that it will disappear.

  In such a case, since the conventional method can be applied only for each process, all batches that have fluctuations near the end of the process of step A560 or fluctuations near the end of the process of step B570 are determined to be “abnormal (defect)”. Will be.

  However, in reality, when both the fluctuation near the end of the process of step A560 and the fluctuation near the end of the process of step B570 occur in one batch, a non-defective product is obtained, and this is treated as a defect. This is clearly an erroneous detection, and the abnormality detection accuracy is remarkably deteriorated.

  On the other hand, in the example in the case where the method of the present invention shown in FIG. 9 is applied, it is possible to easily consider the correlation between the data across the processes by using the data of both the process A60 and the process B70. Therefore, it is possible to greatly improve the abnormality detection accuracy and the quality estimation accuracy.

It is a block diagram which shows the structure outline | summary of the analysis apparatus of batch process data which concerns on embodiment of this invention, and an abnormality detection / quality estimation apparatus using the same. It is a flowchart which shows the flow of control of the apparatus shown in FIG. 1, and the flow of data. It is a flowchart which shows the flow of control of the apparatus shown in FIG. 1, and the flow of data. FIG. 4 is a flowchart illustrating the “two-dimensionalization process” illustrated in FIGS. 2 and 3. It is a figure which shows the relationship of the some batch data linked | related with each other. It is a figure explaining 1-dimensionalization of a some group. It is a figure explaining two-dimensionalization of several batch data. It is a figure which shows the example at the time of applying this invention method. It is a figure which compares the example at the time of applying this invention method, and the example at the time of applying a conventional method. It is a figure which shows the structure outline | summary of the manufacturing data management system in the product manufacturing system by the conventional general plant. It is a figure which shows the manufacturing process of a certain certain conventional product (foodstuff). It is a figure explaining the conventional multiway method. It is a figure which shows the example of abnormality detection using the conventional multiway multivariate analysis method.

Explanation of symbols

1 Batch process data (plant data)
2 Data collection means 3 Data storage means (plant data DB)
4 quality data 5 data input means 6 data analysis means 7 data output means 8 model storage means (model DB)
9 Data determination / prediction means 10 Model creation means 11, 21, 31 Data alignment unit 12, 22, 32 Two-dimensional processing unit 13 Multivariate analysis processing unit 14 Anomaly detection model 15 Quality estimation model 20 Anomaly detection unit 23, 33 Matrix operation processor 24 outputs (T ² / Q)
30 Quality estimation means 34 Quality data estimated value 60 Step A
70 Process B
80 Plant data DB (database)
90 SPC (Statistical Process Management) by the method of the present invention

Claims (7)

A batch process data analysis device that handles time series data with different sampling cycles for each batch as a group of process data in batch units,
Batch data extraction means for extracting batch data associated with each batch from each group after performing a data alignment process for each of the groups;
A group data one-dimensionalization unit that performs one-dimensionalization for each group from the batch data extracted by the batch data extraction unit and performs this on data extracted from all groups;
After one-dimensionalization is performed on all groups by the group data one-dimensionalization means, the one-dimensional data of all the groups is combined to form a whole array of one-dimensional data. An array means;
Two-dimensional “variable” × “sample” where each element of the one-dimensional data for each batch obtained by the whole one-dimensional data array means is “variable” and one-dimensional data (vector) is one “sample”. Data two-dimensionalization means for obtaining digitized data;
A multivariate analysis processing means for executing an analysis process by applying a multivariate analysis method to the two-dimensional data obtained by the data two-dimensionalization means;
An apparatus for analyzing batch process data, comprising: A batch process data analysis device that handles a plurality of process data for each batch as a group of process data in batch units,
Batch data extraction means for extracting batch data associated with each batch from each group after performing a data alignment process for each of the groups;
A group data one-dimensionalization means for making one-dimensional data for each group from the batch data extracted by the batch data extraction means, and performing this for data extracted from all groups;
After one-dimensionalization is performed for all groups by the group data one-dimensionalization means, the one-dimensional data of all the groups are combined to form an overall array of one-dimensional data. Arrangement means;
Two-dimensional “variable” × “sample” in which each element of the one-dimensional data for each batch obtained by the whole one-dimensional data array means is “variable” and one-dimensional data (vector) is one “sample”. Data two-dimensionalization means for obtaining digitized data;
A multivariate analysis processing means for executing an analysis process by applying a multivariate analysis method to the two-dimensional data obtained by the data two-dimensionalization means;
An apparatus for analyzing batch process data, comprising:   The batch process data analysis apparatus according to claim 1 or 2, wherein any one of principal component analysis, independent component analysis, partial least squares, and discriminant analysis is used as the multivariate analysis processing means. An anomaly detection model is generated using the data analysis apparatus according to claim 1, and the anomaly detection is performed after data alignment processing and two-dimensionalization processing are performed on newly acquired batch process data. performing matrix operation of a model, and outputs the T ² statistics and Q statistic, the abnormality detecting device of a batch process data to be detected and the abnormality detecting whether more than a threshold value which is set in advance.   3. The data analysis apparatus according to claim 1 or 2, wherein when generating the abnormality detection model, the multivariate analysis processing unit uses either one of principal component analysis or independent component analysis. 4. The abnormality detection apparatus for batch process data according to 4.   A quality estimation model is generated using the data analysis apparatus according to claim 1, and the abnormality detection is performed after data alignment processing and two-dimensionalization processing are performed on newly acquired batch process data. A batch process data quality estimation device that performs matrix calculation with a model and outputs a quality data estimate value to perform quality prediction.   The data analysis apparatus according to claim 1 or 2 uses any one of a partial least square method and a discriminant analysis as the multivariate analysis processing means when generating the quality estimation model. Item 7. The batch process data quality estimation device according to Item 6.