JP4762088B2 - Process abnormality diagnosis device - Google Patents

Description

  The present invention relates to a process abnormality diagnosis apparatus for performing abnormality diagnosis of a process system such as a water and sewage treatment process, a wastewater treatment process, a water purification process, or a chemical process.

  In recent years, in process systems such as water and sewage water quantity processes, water quality processes, and chemical processes, PID control methods and sequence control methods that have been conventionally used as process control methods in accordance with the diversification and diversification of control purposes. In addition, various advanced control methods (highly improved control methods) have come to be used. For example, there is a method to which advanced automatic control represented by model predictive control or automatic control using a process simulator is applied.

  In order to realize these advanced control methods, it is necessary to detect not only the normal state (normal state) but also various abnormal states and clarify the procedure for dealing with them. For example, in process operation management, sensor failure for measuring process status and operation amount, failure of equipment such as actuator, sudden load fluctuation (sudden disturbance), process processing error, control system error It is necessary to appropriately deal with various abnormal situations such as That is, the process abnormality diagnosis apparatus is an important basic technology for supporting various advanced automatic controls and complete process automation in the process system.

  From such a viewpoint, plant-wide optimum process control having an abnormality diagnosis function has been proposed (see, for example, Patent Document 1). This abnormality diagnosis function has a minimum configuration necessary for realizing optimal control. On the other hand, as a more advanced process abnormality diagnosis method, a method called multivariate statistical process control (MSPC) using a multivariate statistical analysis method is known (for example, (Refer nonpatent literature 1). MSPC is sometimes called a chemometrics method, and is an abnormality diagnosis method developed in the field of chemical processes. Among the MSPCs, an abnormality diagnosis method based on principal component analysis (PCA) is widely used as the most basic and frequently used method.

  In abnormality diagnosis based on PCA, usually, (1) detection of process abnormality or situation change, (2) identification of measurement variable that causes it, (3) identification of true factor, (4) identification of the situation Take a procedure called countermeasures. In this series of procedures, it is a key point in practical use of this method to detect the abnormality of the process performed first and the change of the situation as quickly and accurately as possible. The reason for this is that (1) situation changes such as abnormalities are often meaningless even if detected after the passage of time, and there is no such monitoring system after the passage of time. Even humans can detect changes in the situation. (2) Although it is difficult to specify the cause of an abnormality or the like after an event such as an abnormality has occurred, it can be relatively easily performed immediately after the occurrence of an event such as an abnormality. . Furthermore, (3) a factor specifying process such as an abnormality can be performed by a series of mathematical processes if the abnormality can be detected correctly, but it is extremely difficult to accurately determine the difference between the abnormality and the normal near the limit.

Thus, although it is important to detect a process abnormality quickly and accurately, the conventional process abnormality diagnosis method often fails to perform a process abnormality accurately and quickly.
JP 2004-171531 A Manabu Kano, “principal component analysis”, May 2002, 2nd edition, Kyoto University Graduate School of Engineering, Department of Chemical Engineering, Process System Engineering Laboratory, [Search August 11, 2005], Internet <URL: http: // www-pse.cheme.kyoto-u.ac.jp/kano/document/text-PCA.pdf>

  Even in the abnormality diagnosis method based on PCA as described above, it is not always possible to accurately detect a process abnormality. There are two types of error detection errors as follows. That is, an error type (hereinafter referred to as A type) that is diagnosed as normal even though true (actual) is abnormal, and an error type that is diagnosed as abnormal even though true (actual) is normal (Hereinafter referred to as B type).

  In order to accurately detect abnormalities in process abnormality diagnosis, it is not sufficient to simply increase sensitivity to abnormality detection, and it is important to appropriately adjust the trade-off between type A errors and type B errors. is there. Although it is ideal to reduce both A-type errors and B-type errors as much as possible, in practice, there are many cases where there is an overlap between normal and abnormal at the boundary between normal and abnormal. For this reason, it has been confirmed that, normally, an attempt to reduce the A type error increases the B type error, and conversely an attempt to reduce the B type error increases the A type error. Yes.

  Therefore, an object of the present invention is to optimize the trade-off between an A type error and a B type error, and realize an abnormality diagnosis performance suitable for actual process monitoring together with accurate abnormality detection. Is to provide.

A process abnormality diagnosis apparatus according to an aspect of the present invention includes a measurement unit that measures a state or an operation amount of a target process, a data storage unit that collects and stores time-series data indicating measurement results from the measurement unit, and Process data extraction means for extracting process data indicating the state of the target process, used to create a process monitoring model from the time series data stored in the data storage means, and extracted by the process data extraction means A normalization model generation means for generating a normalization model for normalizing the process data, and a process monitoring model based on a principal component analysis method for the process data normalized by the normalization model is constructed, and the normalization process Q statistics or Hotelling T ² statistics statistics for data anomaly detection A process monitoring model unit that calculates data and calculates a contribution amount of a measurement item (measurement variable) measured by the measurement unit with respect to the statistical data; and abnormal data and normal data included in the statistical data a priori information about the percentage, the a priori information generation means for generating the prior information for the contribution amount, using the contribution of the statistic data which is calculated by the preliminary information and the process monitor model means , A threshold value determining means for determining a threshold value for determining an abnormal state and a normal state of the statistical data, and using the threshold value determined by the threshold value determining means and the statistical data, the target process abnormality And an abnormality diagnosis means for diagnosing the above.

  As another aspect of the present invention, there is a possibility that abnormal data is mixed with normal data, including at least half of normal data, and distinguishing abnormal data from normal data using a threshold value for the absolute value of the data. As a means of determining the ratio of abnormal data and normal data for time series data that can be determined, the normal data and clearly abnormal data are excluded in advance, and the existence range of the threshold of abnormal data and normal data is determined. A process abnormality diagnosis apparatus comprising: a threshold existence range limiting unit for limiting; and a threshold setting unit configured to set a threshold for the data of the threshold existence range determined by the threshold existence range limiting unit using a clustering method.

  In other words, this apparatus is a clustering determination apparatus for normal data and abnormal data. According to this device, a method that simply divides any time-series data in which normal data and abnormal data are mixed into two by clustering may not be considered an appropriate threshold depending on the nature of the target data to which clustering is applied. However, it is possible to set a threshold in an area where the threshold can exist even in the worst case by narrowing down the existing area of the threshold that separates normal data and abnormal data in advance. It is possible to cluster abnormal data and normal data with higher reliability.

  In particular, the setting of a threshold value for judging abnormality / normality and the setting of a normalization parameter (scaling parameter) for normalizing process data include anomaly detection data (eg, Q statistic, Hotelling T2 statistic, etc.) and scaling. Since the ratio of abnormal data and normal data is estimated from the data (eg process measurement data) to be performed and the abnormal data and normal data are appropriately clustered, The proportions can be appropriately clustered. More specifically, in an abnormality diagnosis system based on MSPC using PCA, (1) Q statistics for performing process abnormality diagnosis and situation determination with PCA and Hotling for adjusting sensitivity to abnormality / normality determination (2) A parameter setting method for normalizing process data for adjusting sensitivity to abnormal factors can be appropriately performed.

  According to the present invention, particularly in abnormality diagnosis based on PCA, the trade-off between an A-type error and a B-type error is optimized, and an abnormality diagnosis performance suitable for actual process monitoring along with accurate abnormality detection. Can provide a process abnormality diagnostic apparatus.

  Embodiments of the present invention will be described below with reference to the drawings.

[First Embodiment]
(System configuration)
FIG. 1 is a block diagram illustrating a configuration of a process abnormality diagnosis apparatus according to the first embodiment.

  The process abnormality diagnosis apparatus according to the present embodiment is used in a process monitoring system that monitors process 1 such as a sewage treatment process for the purpose of removing nitrogen and phosphorus, for example. Note that the process abnormality diagnosis apparatus of the present embodiment is not limited to the sewage treatment process as the process 1 to be monitored, and can be applied to other processes such as a chemical process.

  The sewage treatment process 1 includes a first sedimentation tank 101, a first anaerobic tank 102, a second aerobic tank 103, a third anaerobic tank 104, a fourth aerobic tank 105, and a final sedimentation tank 106. In addition, the sewage treatment process 1 includes an initial sedimentation pond excess sludge extraction pump 111 including an extraction flow sensor, an acetic acid input pump 112 that supplies acetic acid-based organic matter including an input amount sensor, a step inflow pump 113 that includes a step inflow sensor, and a supply A blower 114 for supplying oxygen to a second aerobic tank including an air flow sensor, a carbon source input pump 115 for supplying a carbon source (organic matter) including an input sensor, a circulation pump 116 including a circulation flow sensor, and a return flow sensor A return sludge pump 117 including a blower 118 for supplying oxygen to a fourth aerobic tank including a supply air flow rate sensor, a flocculant input pump 119 including an input amount sensor, and a final sedimentation tank surplus sludge extraction pump including an extraction flow rate sensor Each of 1110 has an actuator and its operation amount sensor group.

  Further, the sewage treatment process 1 includes a sewage inflow sensor 121 that measures the inflow sewage amount, an inflow TN sensor 122 that measures the total nitrogen amount contained in the inflow sewage, and an inflow TP sensor that measures the total phosphorus amount contained in the inflow sewage. 123, an inflow UV sensor or inflow COD sensor 124 that measures the amount of organic matter contained in the inflowing sewage, a phosphoric acid sensor 125 that measures the phosphoric acid concentration in the first anaerobic tank 102, and an ammonia concentration in the second aerobic tank 103 It has the ammonia sensor 126, the DO sensor 127 that measures the dissolved oxygen concentration in the second aerobic tank 103, and the nitric acid sensor 128 that measures the nitric acid concentration in the third anaerobic tank 104.

  The sewage treatment process 1 includes an MLSS sensor 129 that measures the amount of activated sludge in at least one of the first anaerobic tank 102, the second aerobic tank 103, the third anaerobic tank 104, and the fourth aerobic tank 105. A phosphoric acid sensor 1210 that measures the phosphoric acid concentration in the fourth aerobic tank 105, a DO sensor 1211 that measures the dissolved oxygen concentration in the fourth aerobic tank 105, and an ammonia sensor that measures the ammonia concentration in the fourth aerobic tank 105 1212, MLSS sensor 1213 that measures the MLSS concentration of the sludge amount drawn from the final sedimentation basin 106, SS sensor 1214 that measures the SS concentration of effluent discharged from the final sedimentation basin 106, and sewage discharge flow rate that measures the amount of sewage discharged Sensor 1215, discharge TN sensor 1216 for measuring the total amount of nitrogen contained in the discharged sewage, discharge TP cell for measuring the total amount of phosphorus contained in the discharged sewage Each of the sensor 1217 and the discharge UV sensor or the discharge COD sensor 1218 for measuring the amount of organic matter contained in the discharged sewage is a process sensor.

  Here, the above-described various actuators 111 to 1110 operate at a predetermined cycle. In addition, the operation amount sensor group of the various actuators 111 to 1110 and the various process sensors 121 to 1218 perform measurement at a predetermined cycle.

  The process abnormality diagnosis apparatus collects process data (time-series data), which is a measurement result of a process state or an operation amount, from various sensors provided in the process 1 at a predetermined cycle and stores it. A storage unit 2 is included.

  Furthermore, the process abnormality diagnosis apparatus includes a process data extraction unit for process monitoring model construction (hereinafter simply referred to as process data extraction part) 3, a process monitoring model construction / supply part (hereinafter simply referred to as model construction part) 4, An abnormality determination criterion setting / supply unit (hereinafter simply referred to as an abnormality determination criterion setting unit) 5, a process monitoring unit 6, a process abnormality diagnosis unit 7, a process abnormality factor sensor specifying unit 8, a process abnormality factor estimation unit 9, And a user interface unit 10.

  The process data extraction unit 3 mainly focuses on normal state (normal state) data in order to build a process monitoring model from various time series data stored in the process measurement data collection / storage unit 2. Process data over a predetermined period is extracted. The model construction unit 4 includes a process data normalization model generation unit 41 and an abnormality detection model generation unit 42. The process data normalization model generation unit 41 includes a process data normalization prior information input unit (hereinafter simply referred to as normalization prior information input unit) 411 and a parameter determination unit 412.

  The prior information input unit 411 for normalization uses various process data (xi (t), t = 1,2, --- l = predetermined period, i = 1,2,-) extracted by the process data extraction unit 3 --28), input prior information on the ratio of normal data and abnormal data included in these process data. In addition, the parameter determination unit 412 uses the equations [(xi (t) -ai) / bi, xi (t) for normalizing various process data according to the prior information input from the normalization prior information input unit 411. ) Determines the shift parameter (ai) and scaling parameter (bi) of the i-th process data]. The shift parameter (ai) is a constant representing a shift for the i-th process data. The scaling parameter (bi) is a constant representing scaling for the i-th process data.

The abnormality detection model generation unit 42 performs the above-described principal component analysis (PCA) on the normalized process data obtained by normalizing the process data using the normalization model defined by the process data normalization model generation unit 41. As a result, a PCA loading matrix (load matrix) and score matrix are calculated, and formulas (models) for calculating Q statistics and Hotelling T ² statistics defined using these are set. To do.

The abnormality determination criterion setting unit 5 includes a threshold determination prior information input unit 51 and a threshold determination unit 52. The threshold determination prior information input unit 51 uses the Q statistic generated by the anomaly detection model generation unit 42 and the hotelling T ² for the normalized process data normalized by the normalization model of the process data normalization model generation unit 41. By using a statistic calculation formula (model), prior information on the ratio of normal data and abnormal data included in each statistic is input. The threshold determination unit 52 sets normal and abnormal thresholds in the Q statistic and the hotelling T ² statistic according to the prior information input from the threshold determination prior information input unit 51.

The process monitoring unit 6 uses the normalized model and the process monitoring model supplied from the model construction unit 4 to, for example, Q statistics for the process data supplied online from the process measurement data collection / storage unit 2. And calculate the Hotelling T ² statistic to monitor the normal or abnormal state of process 1. Process error diagnostic portion 7, with respect to T ² statistic Q statistic and Hotelling monitored by the process monitoring unit 6, using a threshold of abnormality determination set by the failure determination reference setting unit 5, process measurement data Whether the process data supplied from the collection / storage unit 2 is normal or abnormal is determined.

  Furthermore, when the process abnormality diagnosis unit 7 diagnoses that the process 1 is not in a normal state (normal), the process abnormality factor sensor identification unit 8 identifies at least one or more sensors 11 to 1N that cause the process 1. The process abnormality factor estimation unit 9 estimates why such sensor information is output by performing knowledge processing or the like in advance based on the information specified by the process abnormality factor sensor specifying unit 8. The user interface unit 10 includes a display unit for displaying and outputting various types of information, and at least one of the process monitoring unit 6, the process abnormality diagnosis unit 7, the process abnormality factor sensor specifying unit 8, and the process abnormality factor estimation unit 9. The above information is notified to the operator.

(Operation of the first embodiment)
Hereinafter, the effect of this embodiment is demonstrated.

  First, in the process 1 such as a sewage treatment process, information on the state and operation of the process is measured at predetermined intervals by various sensors 11 to 1N. The process measurement data collection / storage unit 2 collects process data as measurement results from the various sensors 11 to 1N at a predetermined cycle, and stores the process data as time series data according to a predetermined format.

  The process data extraction unit 3 mainly focuses on normal state (normal state) data in order to build a process monitoring model from various time series data stored in the process measurement data collection / storage unit 2. A series of process data over a predetermined period is extracted. In other words, the process data extraction unit 3 extracts a series of process data so that, for example, data that belongs to the operation amount of the process 1 or the normal range of the process amount becomes a main component.

  Here, in the process data extraction unit 3, it is often impossible or difficult to extract only normal data in advance. That is, the fact that only normal data can be extracted in advance means that abnormality diagnosis has been completed for at least process monitoring model construction data and it is determined that the data is normal. Therefore, the process data extraction unit 3 obtains prior information preliminarily estimated with respect to the ratio between abnormal data and normal data, and extracts process data including a certain amount of abnormal data. The prior information is, for example, the following information.

(1) The extracted data contains almost no abnormal data. Even if it is included, there are only a few points.

(2) The extracted data includes a little abnormal data, which is about X%.

(3) The extracted data contains abnormal data of several percent or more, but the ratio is unknown.

  If the prior data (1) to (3) as described above cannot be acquired, the process data extraction unit 3 determines that the ratio of abnormal data to normal data in the extracted data is unknown. However, if the extracted data contains, for example, several tens of percent or more abnormal data, the process monitoring model itself cannot be constructed. Therefore, the process data extraction unit 3 is configured so that the main component is normal data. It is necessary to extract process data. In addition, it is preferable that as much data as possible within the normal operation range is mixed in the data to be extracted.

  Furthermore, if the process data includes missing data, abnormal data (including outliers), or data with noise, the process data extraction unit 3 preliminarily stores these data. It is preferable to have a preprocessing function for executing preprocessing. The outlier means abnormal data that occasionally occurs suddenly due to a transmission abnormality or the like. Specifically, the pre-processing function is, for example, a digital filter function such as an FIR filter function or an IIR filter function for removing noise. For outliers, the median filter function, wavelet filter function, etc. are used to extract data that is considered to be real process data from the data.

  The process data extraction unit 3 sends the extracted process data to the model construction unit 4. In the model construction unit 4, the process data normalization model generation unit 41 uses the process data extraction unit 3 to extract a calculation formula (normalization model) for normalizing the process data measured by the various sensors 11 to 1N. Determine using process data. At this time, in the normalization model generation unit 41, the normalization prior information input unit 411 first inputs prior information regarding various process data extracted by the process data extraction unit 3. This prior information is information on the ratio of normal data and abnormal data included in the process data. The normalization prior information input unit 411 displays screen information such as that shown in FIG. 2 on the display unit of the user interface unit 10, for example, to the user who builds the process monitoring model, for example, as shown in FIG. The two pieces of prior information (a) to (d) are selected. Here, when the user selects prior information (d), the user is further prompted to simultaneously input a rough ratio (%) of abnormal data mixed in the process data.

  The parameter determination unit 412 determines the values of the shift parameter (ai) and the scaling parameter (bi) necessary for normalization of various process data based on the prior information input from the normalization prior information input unit 411. It is determined by the following procedure.

  When the prior information (a) shown in FIG. 2 is input from the normalization prior information input unit 411, since there is no prior information, the parameter determination unit 412 is a robust normalization method that is as independent as possible from abnormal values. Determine the parameters when.

  Specifically, a robust sample average or median and robust sample standard deviation is used. Here, the robust sample average and the robust sample standard deviation are the sample average and the sample standard deviation obtained after removing, for example, about 0.5% of data in the vicinity of the maximum value and the minimum value of the process data in advance. is there. Therefore, the parameter determination unit 412 removes some data in the vicinity of the upper and lower limit values in advance and sets the shift parameter (ai) as an average value of process data “ai = 1 / N (Σxi (k))”, Alternatively, the median value (median) of all data is determined. Here, the median is used as it is because it is a robust statistic that is hardly affected by an abnormal value. Whether to use the robust sample average or the median is based on the determination of whether to give priority to setting the data reference value to 0 or to give priority to robustness.

  As a statistic, the sample mean is significantly less robust than the median. This is clear from the following simple example. For example, there is process data of “A = 1.5, 2.3, 1.3, 1.8, 2.0, 1.7, 1.9, 2.1, 2.2, 1.6, 1.4”, and abnormal data “10” is mixed for some reason, and “Ab = 1.5 , 2.3, 1.3, 1.8, 2.0, 1.7, 1.9, 2.1, 10, 1.6, 1.4 ”. In this case, the sample average of the true data A is “1.8”, and the median is also “1.8”. However, the sample average of the measured data Ab is “2.5”, but the median remains “1.8”. As described above, the statistical quantity called the sample average is less robust than the statistical quantity called the median, and greatly depends on the abnormal data. To avoid this, it is possible to use a robust sample average that is calculated by removing some data near the upper and lower limits of the process data. However, it is difficult to determine what percentage of data near the upper and lower limits is to be removed without prior information on abnormal data. In this respect, the median is not greatly affected by abnormal data unless there is an extremely large amount of abnormal data or an extremely large amount of abnormal data on the upper limit side or the lower limit side. However, since the purpose of scaling is to convert so that the average of the scaled data becomes approximately zero, it is preferable to use a robust sample average in this respect. For this reason, it is decided to use the median or the robust sample average depending on whether the designer gives priority to robustness or gives priority to making the scaled reference value close to zero.

On the other hand, as the scaling parameter, after removing some data in the vicinity of the upper and lower limits in advance, a sample standard deviation as shown in the following formula (1) is adopted.

  Here, when calculating the variance (square of the standard deviation), it is often set to “1 / (N−1)” instead of “1 / N” in order to obtain unbiased variance. However, since unbiased variance is not always the best estimate, it is set to “1 / N” or “1 / (N−1)”. Further, the square root of the unbiased variance may be multiplied by another coefficient in order to obtain an unbiased standard deviation instead of the unbiased standard deviation. However, if the number of data is sufficiently large, there is almost no difference in the resulting values.

  The reason for adopting the sample standard deviation is as follows. In the first place, the scaling is to align the operation range of data in order to handle different physical quantities as measured by various sensors 11 to 1N with the same scale. Therefore, if the operation range of the physical quantity measured by each sensor 11 to 1N is known, it is the most straightforward and best method to use the operation range (data range) as a scaling parameter as it is. it is conceivable that. However, the data range statistic is extremely sensitive to outliers. In other words, if abnormal data is mixed in even a little, the data range will change greatly depending on the value. For example, in the above example, the data range of the process data A is 1 (minimum value 1.3, maximum value 2.3), but the data range of Ab is 8.7 (minimum value 1.3, maximum value 10). For this reason, as described above, it may be possible to calculate by removing some data near the upper and lower limits of the process data, but there is no prior information on abnormal data as to what percentage of data near the upper and lower limits is removed. It is difficult to determine.

  Compared to such a data range, the sample standard deviation is a robust statistic and is an index indicating the operating range of data, so it is often used as a normal scaling parameter. Here, if the data follow a normal distribution, 99.7% of the data is covered by 3 times the sample standard deviation, so about 3 times the sample standard deviation corresponds to the data range. However, since it is not known whether the data follows a normal distribution, the correspondence between the data range and the sample standard deviation is generally unknown. However, since the scaling is performed in order to see a plurality of different physical quantities with the same index, it is not particularly necessary to associate the sample standard deviation with the data range. In addition, the sample standard deviation is a robust statistic compared to the data range. However, since it depends to some extent on abnormal data, the robust sample standard deviation is preferably used.

  Further, if the prior information (b) shown in FIG. 2 is input from the normalization prior information input unit 411, the parameter determination unit 412 knows that almost all data is normal in advance. The parameter is set using the information (b). Specifically, the parameter determination unit 412 uses a sample average for the shift parameter because it is known in advance that almost all data is normal for the reasons described above. However, assuming that some abnormal values are mixed, a robust sample average excluding data of about 0.5% near the upper and lower limits may be used.

  On the other hand, since the parameter determination unit 412 knows that almost all data is normal in advance for the scaling parameter, for example, the vicinity of the maximum value and the minimum value of the process data, for example, 0.5 After removing about% data, the operation range of the process data is directly set as a scaling parameter.

  Further, if the prior information (c) shown in FIG. 2 is input from the normalization prior information input unit 411, the parameter determination unit 412 has a mixture of abnormal data and normal data, and the ratio of abnormal data is unknown. Therefore, using the information (c), threshold values for normal data and abnormal data are determined in advance using a clustering method. Clustering is a general term for a method for determining which class each data belongs to when data considered to belong to a plurality of different sets (classes) are mixed. In this case, since it is assumed that data belonging to two classes of normal data and abnormal data are mixed, a two-class clustering method is used.

  Here, even in the case of prior information (a), it seems that such a two-class clustering method can be applied, but this is not the case. This is because the two-class clustering method works effectively in a situation where normal data and abnormal data are mixed, so if there is no prior information, 2 This is because when the class clustering method is applied, normal data is forcibly divided into two when abnormal data is hardly mixed in the data.

Since various methods have been developed for such a clustering method, any appropriate method may be used. For example, a kind of discriminant analysis method called “Otsu's threshold determination method” can be used. The “Otsu's threshold determination method” is a method of setting a threshold value so that the inter-class variance between two classes is maximized when data belonging to two different classes are mixed. In other words, the threshold value is set so that the average value of the variance within each class is minimized. That is, this is a method of determining a threshold value so that data belonging to two classes are as far as possible from each other. Specifically, a threshold value that minimizes the following expression (2) is obtained.

  Here, ω1 and ω2 are the ratios (class occurrence probabilities) of data belonging to class 1 and class 2, respectively. σ 1 and σ 2 are the variances in class 1 and class 2, respectively. This is shown in FIG.

  This Otsu threshold determination method is based on the condition that the two classes of data each follow a normal distribution, the condition that there is no significant difference in the proportion of data belonging to the two classes, and the condition that the variances of the normal distributions of the two classes are equal. It is shown that this is the maximum likelihood estimate of the threshold below. Here, the maximum likelihood estimated value is an estimated value by the maximum likelihood estimation method included in the statistical estimation method.

Usually, since the ratio of normal data is considerably larger than the ratio of abnormal data, the condition that there is no significant difference in the ratio of data belonging to the two classes is not satisfied. Thus, when there is a significant difference in the ratio of data belonging to the two classes, it is preferable to minimize the following expression (3) instead of the expression (2).

  Note that the minimization of “ω1 · σ1 + ω2 · σ2” is the same as the minimization of “log (ω1 · σ1 + ω2 · σ2)”. Thus, a correction term “-ω1 · log (ω1) -ω2 · log (ω2)” is added. Hereinafter, this is referred to as a modified Otsu threshold determination method.

  In this way, by applying a two-class clustering method such as the Otsu threshold determination method or the modified Otsu threshold determination method, the threshold values for normal data and abnormal data can be determined. The parameter determination unit 412 performs this determination for all measurement items measured by the various sensors 11 to 1N. FIG. 4 is a diagram showing the concept of determining the threshold value 410 for classifying normal data and abnormal data in this way. As shown in FIG. 4, since the process data for constructing the process monitoring model can be provisionally classified into normal data and abnormal data by the clustering method, the sample average of normal data is calculated using only the normal data portion 400. A shift parameter is used, and a sample standard deviation of normal data is used as a scaling parameter.

  If the process data is data that may cause anomalies in both the upper limit and lower limit directions, calculate the absolute value of the process data in advance, and then determine the threshold value for Otsu and the threshold value for modified Otsu. Two classes of clustering methods such as the method are applied. In this case, as a result, normal data and abnormal data can be classified with an image as shown in FIG.

  Furthermore, when the prior information (d) shown in FIG. 2 is input from the normalization prior information input unit 411, the parameter determination unit 412 can recognize the approximate proportion of abnormal data in which normal data is mixed. . In this case, as described above, the ratio of abnormal data is additionally input. For example, when the ratio of abnormal data is entered as 5%, if the process data is such that an abnormality occurs only in one direction, the maximum (or minimum) value of the process data is 5 The threshold value is a point including% data (see 410 in FIG. 4).

  In addition, when the process data is data that may be abnormal in both directions, the threshold value is a point that includes 2.5% of data from the maximum value and the minimum value of the process data (510 in FIG. 5). 511). As a result of such an operation, the normal data range 500 can be determined. Therefore, the sample average of data regarded as normal data is used as a shift parameter, and the sample standard deviation is used as a scaling parameter.

  According to the procedure described above, the parameter determination unit 412 performs the shift parameter (ai) and scaling parameter (bi) necessary for normalization of various process data based on the prior information input from the normalization prior information input unit 411. Determine the value of. That is, the process data normalization model generation unit 41 normalizes the process data with the normalization model based on the prior information of the process data.

  Here, the relationship between appropriate normalization of process data and abnormality diagnosis will be briefly described. The means of normalization is usually means for converting a plurality of physical quantities having completely different values so that they can be evaluated on the same scale. If this operation is not appropriate, a plurality of physical quantities cannot be evaluated on the same scale, and as a result, the detection of an abnormality or a sensor (= physical quantity) that becomes an abnormal factor after detecting an abnormality is erroneous. For example, when the operation range of the physical quantity A is 0 to 10 and the operation range of the physical quantity B is 0 to 100, both can be evaluated on the same scale by setting B to “1/10”. However, if normalization is mistaken and B is used as it is, A will be evaluated very small with respect to B. Therefore, even when an abnormality occurs in the amount of A, the abnormality is often not detected. Even when an abnormality is detected, since B is a larger operating range, the abnormality factor is determined to be B in most cases. For this reason, proper normalization of process data is extremely important in improving the abnormality detection accuracy and the abnormality factor identification accuracy.

  Next, in the model construction unit 4, when the parameter for normalizing the process data for constructing the process monitoring model is determined by the process data normalization model generation unit 41, the abnormality detection model generation unit 42 uses the parameter. Normalize process data for building process monitoring models. Then, the abnormality detection model generation unit 42 constructs an abnormality detection model using the normalized data.

Specifically, the abnormality detection model generation unit 42 sets variables corresponding to the various sensors 11 to 1N in the horizontal (row) direction and sets the normalized data to a matrix X that considers time in the vertical (column) direction. Set. Next, data decomposition is executed using PCA as shown in the following equation (4).

Equation (4) is equivalent to decomposing the original data matrix as shown in FIG. For the data thus decomposed, the loading matrix P becomes the original model for calculating the Q statistic and the Hotelling T ² statistic. The loading matrix P, may be considered as a process monitor model, in practice, T ² statistic Q statistic and Hotelling performing abnormality detection by the following equation (5) is calculated by (6).

  Here, ∧ is a matrix having the variance of each principal component by PCA (principal component analysis) as a diagonal element, and means that the variance is normalized. I is a unit matrix of an appropriate size. X (t) is the t-th element of the matrix X. In actual process monitoring, this x (t) is calculated by replacing it with process data that is measured online, so the above equations (5) and (6) are models generated by the abnormality detection model generation unit 42. is there.

Here, the meaning of the Q statistic defined by the equations (5) and (6) and the Hotelling T ² statistic will be briefly described with reference to FIG.

  The abnormality diagnosis method using PCA is a method of performing abnormality detection using correlation information included in a plurality of measurement variables. Normally, there is some correlation between the variables measured in the process, and there is little information that is completely independent of each other. FIG. 7 shows a case where two variables X1 and X2 are correlated for convenience, and the measurement data of X1 and the measurement data of X2 are plotted on the X1-X2 plane.

  Here, for example, if the X1 sensor fails, the correlation is lost, so data is plotted at points as indicated by triangles. Thus, the quantity representing the deviation from the correlation axis representing the correlation is the meaning of the Q statistic. On the other hand, for example, when the process load suddenly increases, X1 and X2 have a correlation, so that the values of X1 and X2 both increase (or decrease). In such a case, the data is plotted at points as represented by squares.

That is, although the correlation does not collapse, an index for measuring the amount by which the value increases or decreases beyond the normal operating range is the Hotelling T ² statistic. In other words, the Q statistic is an index representing a deviation in a direction in which the data distribution of a plurality of items is not distributed. Further, the amount of distribution in the direction in which the data of the data distribution of a plurality of items is distributed is the Hotelling T ² statistic.

In short, in the model construction unit 4, the process normalization model generation unit 41 determines the values of the normalization shift parameter and scaling parameter. The anomaly detection model generation unit 42 determines a calculation model necessary for performing anomaly detection, and generates a normalization calculation formula, a Q statistic, and a hotelling T ² statistic calculation formula.

Next, the abnormality determination criterion setting unit 5 determines a threshold value for performing abnormality determination using the equations (5) and (6) which are abnormality detection models. At this time, using the normalized data X which is the process monitoring model construction data, the Q statistic of the process monitoring model construction data and the hotelling T ² statistic calculated by the formulas (5) and (6) are used. To do. Hereinafter, a method of setting an abnormality determination criterion by the abnormality determination criterion setting unit 5 will be described.

First, the threshold determination prior information input unit 51 uses the calculation formula (model) of the Q statistic and the hotelling T ² statistic generated by the abnormality detection model generation unit 42 to be included in each statistic. Enter prior information for the ratio of normal data to abnormal data. Since the Q statistic and the hotelling T ² statistic are statistics for detecting anomalies, if prior information about this cannot be acquired, prior information on process data may be substituted. In the case of substitution, the prior information input from the prior information input unit 411 is used as it is.

Further, instead of the process data directly if you want to enter a pre-information on T ² statistic Q statistic and Hotelling is similar input screen information to that shown in FIG. 2 Q statistic and Hotelling A T ² statistic is also created, and the prior information is input in the same manner as the prior information input unit 411. Unlike process data, in general, prior information on Q statistics and hotelling T ² statistics is rarely known, but each piece of statistical information can be measured regardless of how much data is measured in the process. since become a single total two data respectively, it may be determined by plotting the T ² statistic calculated Q statistic and Hotelling using normalized data X of the data for the construction process monitoring model.

If such visualization is performed, it is often possible to acquire qualitative information such as whether or not abnormal data is mixed. Although prior information may be acquired by executing visualization on process data in the case of normalization, the process data is usually at least several tens according to items measured by various sensors 11 to 1N. The data is more than a variable. Therefore, it is unrealistic to obtain prior information by performing visualization on such data. On the other hand, the Q statistic and the Hotelling T ² statistic are always one data regardless of the number of process data measurement variables, and thus it is easy to obtain prior information by visualization. If the prior information can be acquired, the threshold determination prior information input unit 51 performs the same processing as the prior information input unit 411.

  Next, the threshold determination unit 52 determines a threshold for abnormality determination according to the prior information input from the threshold determination prior information input unit 51.

First, when there is no prior information (information regarding the Q statistic and the Hotelling T ² statistic is unknown), the threshold determination unit 52 uses the statistical confidence limit value of the Q statistic as a default setting method. statistical confidence limits about the T ² statistic Hotelling a (calculated by a known calculation formula) is used. Further, when information indicating that almost all data is normal is input as prior information, the threshold value determination unit 52, in principle, calculates using the normalized data X of the process monitoring model construction data. The maximum value of the Q statistic and the hotelling T ² statistic is determined as a threshold value. However, since there is a possibility that anomalous data is included in a slight amount, for example, the concept of 99% confidence interval is adopted, and the top 1% value from the maximum value of the Q statistic and the hotelling T ² statistic is selected. The method of determining the threshold value is preferable.

Further, when information indicating that abnormal data and normal data are mixed but the ratio of abnormal data is unknown is input as prior information, the threshold determination unit 52, for example, similar to the parameter determination unit 412, The threshold value of the Q statistic and the hotelling T ² statistic is determined by using a modified Otsu threshold determination method or the like. This simply changes the data to which a clustering algorithm such as the modified Otsu's threshold determination method is applied to Q statistics calculated using normalized data X and hotelling T ² statistics data instead of process data. It can be executed with.

In addition, when information indicating that abnormal data and normal data are mixed and the approximate ratio of abnormal data is known is input as prior information, the threshold determination unit 52 is the same as the parameter determination unit 412. Execute the process. That is, the threshold determination unit 52, a respective value corresponding to the upper alpha% of the data of the calculated Q statistic and Hotelling T ² statistic using normalized data X (alpha is entered as a priori information value), It is determined as a threshold value for each statistic.

  In short, in the abnormality determination criterion setting unit 5, the threshold determination unit 52 determines a threshold for determining whether the process 1 is normal or abnormal according to the prior information input from the threshold determination prior information input unit 51. The above series of operations is for constructing a process monitoring model and setting a criterion for abnormality determination, and is normally performed offline.

On the other hand, the following actions are executed online. First, the process monitoring unit 6 monitors the process 1 using the process monitoring model constructed by the model construction unit 4. Process monitoring unit 6, a process variable to be monitored, various sensors 11~1N not itself a T ² statistic Q statistic and Hotelling newly generated from each of these sensors. Here, at least one of these is monitored. Usually, both are monitored, and if necessary, other variables, for example, the sensors 11 to 1N themselves and the score calculated by the PCA can be plotted on a plane and monitored.

The process monitoring unit 6 uses the values of the various sensors 11 to 1N that are transmitted online from the process measurement variable collection / storage unit 2 (from these sensor values at a certain time) as the above-described Q statistics and hotelling T ² statistics. And a process monitoring model constructed by the model construction unit 4 is used to perform calculations as shown in the following formulas (7) and (8).

The difference between these formulas (7) and (8) and the above formulas (5) and (6) is simply the difference in the data input to calculate the Q statistic and the Hotelling T ² statistic. In online process monitoring, it is only necessary to input process data measured online to a process monitoring model that has already been established. Since Q (Xnew (t)) and T ² (Xnew (t)) calculated in this way are scalar values, they can be monitored by displaying these values as a trend graph. In this case, failure determination reference Q set by the setting unit 5 statistic and Hotelling T ² statistic threshold (limit level: CL) also may be plotted simultaneously.

Next, the process error diagnostic portion 7, T ² statistic Q statistic and Hotelling exceeded the abnormal set by criterion setting unit 5 Q statistic and Hotelling T ² statistic limit level (CL) If it is judged abnormal. At this time, in addition to the abnormality determination Q statistic was set by the reference setting unit 5 and the Hotelling T ² statistic threshold level, for example, based on the abnormality diagnostic rule separately provided warning level (WL) and attention level (NL) Judgment may be made.

  Specifically, for example, the process abnormality diagnosis unit 7 sets the warning level (WL) to “2/3” of the limit level and sets the attention level (NL) to “1/3” of the limit level. The normality and the abnormality are judged by the following diagnostic rule. (1) When one point exceeds CL, (2) Of two consecutive points, two or more points exceed WL, (3) Four or more points among five consecutive points are 1σ (standard) When (deviation) is exceeded, (4) when 8 consecutive data monotonously increase or monotonically decrease, (5) when some pattern with non-randomness appears, the process abnormality diagnosis unit 7 to decide.

Here, if a complicated abnormality diagnosis rule as described above is not necessary, or if a B type error (see FIGS. 13 and 14 to be described later) increases due to this, simply set the abnormality criterion. the threshold T ² statistic Q statistic and Hotelling set in part 5 performs an abnormality diagnosis as a limit level (CL). In short, the process abnormality diagnosis unit 7 determines the normal state and abnormal state of the process 1 by the above procedure.

Next, when the process abnormality diagnosis unit 7 diagnoses that the process 1 is in an abnormal state, the process abnormality factor sensor specifying unit 8 determines the Q statistic or hoteling determined to be abnormal as shown in FIG. By performing a contribution plot with respect to the T ² statistic, the sensor that is the factor is specified. FIG. 8 shows a sensor (here, a nitric acid sensor) in which each bar graph corresponds to various sensors 11 to 1N, and the bar having the longest length is a cause of abnormality. However, only one factor is not necessarily increased here as shown in FIG. For example, as shown in FIG. 9, some sensors may be considered unusual. In such a case, the process abnormality factor sensor specifying unit 8 specifies a plurality of sensors as factors.

  Next, the process abnormality factor estimation unit 9 estimates the true abnormality factor of the process 1 based on the sensor information identified by the process abnormality factor sensor identification unit 8. Specifically, the process abnormality factor estimation unit 9 executes inference as an expert system constructed based on, for example, past experience or scenario simulation.

  Specifically, as shown in FIG. 8, the process abnormality factor estimation unit 9 is a failure of a sensor when only one sensor (here, a nitric acid sensor) shows an outstanding value. Estimated. In addition, some sensors are configured so that a failure signal can be output. Therefore, the abnormality of the sensor is estimated by taking a logical sum or logical product with the failure signal obtained therefrom. . FIG. 9 is a diagram in a case where a phenomenon called nitrification inhibition occurs in which a microorganism called nitrifying bacteria that removes ammonia does not work. In this case, since nitrification is not performed, the concentration of ammonia becomes high, and nitrifying bacteria become inactive, so that the amount of oxygen necessary for the activity becomes small, and as a result, the dissolved oxygen concentration becomes high. If such process knowledge is prepared in advance, the process abnormality factor estimation unit 9 is based on an expert rule such as “there is a possibility of nitrification inhibition when the ammonia concentration is high and the dissolved oxygen concentration is high”. To estimate the factors.

  The user interface unit 10 notifies the operator of at least one information from the process monitoring unit 6, the process abnormality diagnosis unit 7, the process abnormality factor sensor specifying unit 8, and the process abnormality factor estimation unit 9.

(Effects of the first embodiment)
First, as described above, problems in the conventional abnormality diagnosis method based on PCA will be specifically described with reference to FIGS.

  Error detection in an abnormality diagnosis method based on PCA is true (actual) is abnormal but type A of an error to diagnose as normal, and true (actual) is normal but abnormal There are two types of error B type. FIG. 13 is a diagram showing the relationship between the distribution of normal data and abnormal data and the error type (A, B). FIG. 14 is a diagram conceptually showing the characteristics of the error type. That is, the A type corresponds to type 2 error detection with insufficient accuracy. The B type corresponds to the first type error detection which is an abnormal error detection.

  FIG. 15 is a diagram showing time-series data of the inflow ammonia concentration when the abnormality diagnosis algorithm is applied, and is a B type first type error detection example. That is, the abnormal portion 190 shown in FIG. 15 is an example that is not actually abnormal but is erroneously determined to be abnormal by the abnormality diagnosis algorithm. FIG. 16 is an example of A type second type error detection in which the abnormality data 200 is overlooked when the abnormality diagnosis algorithm is applied. That is, this is an example in which the data 200 that can be clearly determined to be abnormal from the actual process data is determined to be normal, and the abnormality detection accuracy is insufficient.

  In the system according to the present embodiment, the threshold value for detecting anomalies in multivariate statistical process monitoring (MSPC) using multivariate statistical analysis is appropriately determined according to prior information for process data, and the above-described A It is possible to reduce two types of abnormality detection errors of type B and type B. Therefore, the process abnormality monitoring performance by MSPC can be improved.

  In addition, when building a process monitoring model with MSPC, by appropriately determining parameters for normalizing multiple process data according to prior information on the process data, the cause of the abnormality is identified It is possible to reduce the error of the factor variable when doing. Therefore, the process abnormality monitoring performance by MSPC can be improved.

  In other words, in process monitoring (process abnormality diagnosis) by MSPC, the threshold of abnormality detection is appropriately determined according to the amount and quality of prior knowledge of process data, thereby allowing trade between A type error and B type error. It is possible to optimize the off state and improve the abnormality detection performance.

[Second Embodiment]
The system according to the second embodiment is the same as that shown in FIG. 1 as a basic configuration. Compared with the system of the first embodiment described above, the system of this embodiment is an abnormality detection model generation unit 42, a threshold determination unit 52, a process monitoring unit 6, a process abnormality diagnosis unit 7, and a process abnormality factor sensor identification. Each operation of the part 8 is different. Hereinafter, the description will be focused on the operation of different parts, and description of the same operation will be omitted.

  First, in the system of the present embodiment, the processes of the process measurement data collection / storage unit 2, the process data extraction unit 3, and the process data normalization model generation unit 41 in the model construction unit 4 are the same as those in the first embodiment. It is the same as the case of the system.

  When the parameter for normalizing process data for constructing the process monitoring model is determined by the process data normalization model generating unit 41, the abnormality detection model generating unit 42 uses the parameter to process the process data for constructing the process monitoring model. Normalize. Then, the abnormality detection model generation unit 42 constructs an abnormality detection model using the normalized data.

Specifically, the abnormality detection model generation unit 42 sets variables corresponding to the various sensors 11 to 1N in the horizontal (row) direction and sets the normalized data to a matrix X that considers time in the vertical (column) direction. Set. Next, in the same manner as described above, the Q statistic and the hotelling T ² statistic are obtained by using PCA as shown in the equations (5) and (6). Furthermore, in this embodiment, the abnormality detection model generation unit 42 calculates the contribution amount of the various sensors 11 to 1N to the Q statistic and the hotelling T ² statistic by the following equations (9), (10), and (11). To do.

  Here, x (t, n) means the value at the time t of the nth element of the variable x, and n corresponds to the various sensors 11 to 1N. The notation F (:, n) means taking out the nth row of the matrix F. The same applies to P (:, n).

A calculation formula (model) for calculating the contribution amount corresponds to the process monitoring model in the present embodiment. In a normal MSPC, the equations for calculating the contribution amount as shown in the equations (10) and (11) are usually used as follows. First, the equation (5), and executes fault detection by T ² statistic Q statistic and Hotelling represented by (6), if the if abnormality has occurred, the equation (10), ( It is determined by which sensor process data measured by the contribution amount calculation as shown in 11) contributes to the abnormality.

  On the other hand, the present embodiment is a method for detecting an abnormality by setting a threshold value for the contribution amount itself. Hereinafter, a description will be given with reference to FIG.

FIG. 10 is a diagram showing the concept of the contribution amount to the Q statistic in the case of two variables. In FIG. 10, reference numeral 70 indicates a Hotelling T ² statistic, and reference numeral 71 indicates a Q statistic. As described above, the Q statistic is an index representing a deviation in a direction in which data is not distributed. As shown in FIG. 10, the Q statistic itself can be decomposed into a contribution amount 72 due to the X1 component and a contribution amount 73 due to the X2 component. In normal MSPC, when an abnormality is detected by the Q statistic of the equation (5), the anomaly factor is estimated by decomposing into each component by the equation (10), but the contribution amount itself detects the abnormality. It can be seen that the contribution amount represented by the equation (10) can always be calculated regardless of whether or not it is performed. Therefore, if the contribution amount is always calculated by the equation (10), it can be easily understood that the abnormality detection may be performed by setting the threshold value directly.

Furthermore, in the method of the present embodiment, the sum of the contribution amount of the Q statistic for all the variables becomes the Q statistic, so that it is configured rather than the summed Q statistic. More accurate abnormality detection can be performed by performing abnormality detection on elements. However, in this case, diagnosis is performed using the contribution amount, not the Q statistic or the hotelling T ² statistic. Therefore, unless an appropriate threshold is set for the contribution amount, the abnormality diagnosis performance is It may deteriorate. The method of the present embodiment aims to set an appropriate threshold value even in such a case.

In short, when the process data normalization model generation unit 41 determines the normalization shift parameter and scaling parameter values, the abnormality detection model generation unit 42 calculates a calculation model (P , ∧) is determined. Further, a normalization calculation formula, a Q statistic, and a hotelling T ² statistic calculation formula using these values are generated.

Next, the abnormality determination criterion setting unit 5 determines a threshold for performing abnormality determination using the equations (10) and (11) that are abnormality detection models. At this time, the contribution of the Q statistic of the process monitoring model construction data and the hotelling T ² statistic calculated by the above formulas (10) and (11) using the normalized data X that is the process monitoring model construction data Use quantity.

First, the threshold determination prior information input unit 51 inputs prior information regarding the contribution amount of the Q statistic and the hotelling T ² statistic. Incidentally, if it can not obtain a priori information about the contribution amount, and prior information about the process data, it may be replaced by a priori information for the T ² statistic Q statistic and Hotelling.

  Next, the threshold determination unit 52 determines a threshold for abnormality determination according to the prior information input from the threshold determination prior information input unit 51.

First, when there is no prior information (information regarding the Q statistic and the Hotelling T ² statistic is unknown), the threshold determination unit 52 uses the statistical confidence limit value of the Q statistic as a default setting method. statistical confidence limits about the T ² statistic Hotelling a (calculated by a known calculation formula) is used. In this case, the threshold values of the Q statistic and the Hotelling T ² statistic are calculated. In normal abnormality diagnosis by MSPC, an abnormality is detected when this threshold is exceeded, and the contribution amount is calculated only when the factor is specified. For example, (i) the one with the largest contribution amount is used as a factor candidate, (ii) the top three contribution amount values are used as factor candidates, or (iii) the contribution amount value is a certain percentage of the threshold value. Those that exceed are considered abnormal factor candidates.

Here, for example, 80% of the threshold values of the Q statistic and the Hotelling T ² statistic are determined as the threshold according to the concept of the method (iii) that is normally performed. In normal abnormality diagnosis, even the contribution of certain variables has exceeded 80% of the threshold, abnormality determination unless T ² statistic own Q statistic or Hotelling is the sum of the contribution amount exceeds the threshold value Is not. On the other hand, when performing an abnormality diagnosis based on the contribution amount, even if the Q statistic or the hoteling T ² statistic itself does not exceed the threshold, the contribution to the statistic exceeds a certain ratio of the threshold. Diagnosed as abnormal. For this reason, the merit which can raise the sensitivity with respect to the abnormality which has arisen with the variable which various sensors 11-1N measure is acquired.

  However, in normal MSPC abnormality diagnosis, even if each variable does not exceed a certain ratio such as 80% of the threshold value, it is determined as abnormal if the total statistic exceeds the threshold value. For this reason, the method described here may deteriorate the abnormality detection accuracy for a situation where each variable is slightly deviated from normal. In order to cope with such a situation, both methods may be used together.

Next, when information indicating that almost all data is normal is input as prior information from the threshold determination prior information input unit 51, the threshold determination unit 52 is in principle configured for process monitoring model construction. contribution amount and using the normalized data X of the calculated Q statistic data, determines the maximum value of T ² statistic contributions of Hotelling as a threshold value. However, there is a possibility that contains very little abnormal data, for example by adopting the idea of a 99% confidence interval, the maximum value of Q statistic contribution amount and Hotelling T ² statistic contribution amount The method of determining the upper 1% value as a threshold is preferable. In this case, the maximum value is obtained for each contribution amount of the various sensors 11 to 1N. However, since it is known in advance that almost all data is normal as prior information, the maximum value determined for each contribution amount is determined. The largest value is adopted as the threshold value.

Furthermore, when information indicating that abnormal data and normal data are mixed but the ratio of abnormal data is unknown is input as prior information, the threshold value determination unit 52, for example, a threshold value determination method for the modified Otsu, etc. Is used to determine the threshold for the contribution of the Q statistic and the contribution of the Hotelling T ² statistic. However, whether or not abnormal data and normal data are mixed for all measurement variables of the various sensors 11 to 1N, even if it is input as prior information that the abnormal data and normal data are mixed. I can't judge. For this reason, if there is a measurement variable in which abnormal data is not mixed, if the threshold value is determined by a clustering method such as the modified Otsu threshold value determination method, normal data will be divided into two, and an appropriate threshold value will be obtained. Not determined. However, since any of the measurement variables of the various sensors 11 to 1N should include abnormal data, any one of the threshold values determined by a method such as the threshold determination method of the modified Otsu for each measurement variable is used. The threshold value must always separate normal data and abnormal data. The maximum value among the threshold values determined for each measurement variable always separates normal data and abnormal data. Accordingly, the maximum threshold value determined by a method such as the modified Otsu threshold value determination method for each measurement variable is adopted as the threshold value.

Further, when information indicating that abnormal data and normal data are mixed and the approximate ratio of abnormal data is known is input as prior information, the threshold value determination unit 52 uses the normalized data X. Each value corresponding to the upper α% (α is a value input as prior information) of the data of the contribution amount of the Q statistic and the contribution amount of the Hotelling T ² statistic calculated in advance is calculated in advance. If the ratio of abnormal data to normal data is known for all measurement variables, a threshold value calculated for each measurement variable may be applied. If prior information has been obtained that not all measurement variables, but any measurement variable contains approximately α% of abnormalities, the threshold value determined for each measurement variable Is used as the threshold value.

  In short, in the abnormality determination criterion setting unit 5, the threshold determination unit 52 determines a threshold for determining whether the process 1 is normal or abnormal according to the prior information input from the threshold determination prior information input unit 51. The above series of operations is for constructing a process monitoring model and setting a criterion for abnormality determination, and is normally performed offline.

On the other hand, the following actions are executed online. First, the process monitoring unit 6 monitors the process 1 using the process monitoring model constructed by the model construction unit 4. Process monitoring unit 6, a process variable to be monitored, to monitor the contribution of the contribution amount and Hotelling T ² statistic Q statistic newly generated from the various sensors 11~1N directly. That is, the process monitoring unit 6, the following equation (12), (13), (14) is used to monitor the contribution of the contribution amount and Hotelling ^{T 2} statistic Q statistic.

That is, the process monitoring target is not a Q statistic or a Hotelling T ² statistic, but a contribution amount thereof.

Next, the process abnormality diagnosis unit 7 determines that there is an abnormality when the contribution amount of either the Q statistic or the Hotelling T ² statistic exceeds a threshold value. The process abnormality factor sensor specifying unit 8 is diagnosed on the basis of the contribution amount by the process abnormality diagnosis unit 7, so that each measurement variable of the various sensors 11 to 1N corresponding to the contribution amount exceeding the threshold is a factor sensor. Directly identify that there is. The operations of the process abnormality factor estimation unit 9 and the user interface unit 10 are exactly the same as those in the first embodiment.

  As described above, according to the second embodiment, usually, the method of firstly detecting an abnormality and secondly specifying the abnormality factor is changed, and the amount calculated to specify the abnormality factor ( (Contribution amount) is directly used for abnormality detection. In this case, errors in abnormality detection can be reduced by appropriately determining the threshold value for detecting the abnormality according to prior information on the process data. As a result, the performance of process abnormality monitoring by MSPC can be improved. In other words, it is possible to improve the abnormality detection performance that occurs particularly in a specific process measurement variable.

[Third Embodiment]
The system according to the third embodiment is the same as that shown in FIG. 1 as a basic configuration. Compared with the system of the first embodiment described above, the system of this embodiment is an abnormality detection model generation unit 42, a threshold determination unit 52, a process monitoring unit 6, a process abnormality diagnosis unit 7, and a process abnormality factor sensor identification. Each operation of the part 8 is different. Hereinafter, the description will focus on the action of different parts, and the description of the action of shaking will be omitted.

  First, in the system of the present embodiment, the processes of the process measurement data collection / storage unit 2, the process data extraction unit 3, and the process data normalization model generation unit 41 in the model construction unit 4 are the same as those in the first embodiment. It is the same as the case of the system.

  When the parameter for normalizing process data for constructing the process monitoring model is determined by the process data normalization model generating unit 41, the abnormality detection model generating unit 42 uses the parameter to process the process data for constructing the process monitoring model. Normalize. Then, the abnormality detection model generation unit 42 constructs an abnormality detection model using the normalized data.

Specifically, the abnormality detection model generation unit 42 sets variables corresponding to the various sensors 11 to 1N in the horizontal (row) direction and sets the normalized data to a matrix X that considers time in the vertical (column) direction. Set. Next, in the same manner as described above, the Q statistic and the hotelling T ² statistic are obtained by using PCA as shown in the equations (5) and (6). Thus, the process monitoring model construction is completed, but in this embodiment, the abnormality detection model generation unit 42 uses the Q statistic and the Hotelling T ² statistic shown in the equations (5) and (6) as discrete wavelets, respectively. Decompose data into multiple frequency bands using transformation. This can be achieved by reconstituting the decomposition of T ² statistic Q statistic and Hotelling by discrete wavelet transform, the decomposition respectively of data to the inverse discrete wavelet transform. As a result, this operation can be realized in the form of a digital filter. Therefore, when the filter is “Wi, i = 1, ---, p (p is the number of data divisions)”, each statistic decomposed by the wavelet Can be expressed as the following equations (15) and (16), for example.

  The calculation formula (model) decomposed for each frequency band corresponds to the process monitoring model in the present embodiment. FIG. 11 is a diagram illustrating a specific example in which the Q statistic is decomposed into data for each of a plurality of frequency bands by the discrete wavelet transform. FIG. 4A is a graph of Q statistics. FIGS. 7B to 7F are diagrams showing Q statistics divided into five frequency bands. Here, FIG. 11D is a diagram that approximates the original Q statistic, and represents the slowest change. The other figures (B), (C), (E), and (F) are diagrams showing the decomposition data of the Q statistics in the faster frequency band where the average becomes almost zero, respectively. In FIG. 11, the horizontal axis is assigned in order from the fast frequency band to the slow frequency band.

  The significance of decomposing data for each frequency band as described above is as follows.

FIG. 12 is a diagram illustrating a specific example in which certain original data is decomposed by applying a discrete wavelet transform, although it is not a Q statistic or a Hotelling T ² statistic. In the original data on the left side of the arrow in FIG. 12, it can be confirmed that the data includes a high-frequency abnormality. When this original data is decomposed by wavelet transform, it appears as an abnormality in the high frequency part as shown on the right side of the arrow.

  In general, it is relatively easy for a human to visually confirm a high-frequency abnormality regardless of whether or not wavelet transform is applied, but the original data as shown on the left side of the arrow in FIG. Is detected mechanically (automatically), it is impossible to detect an abnormality simply by setting a threshold value. On the other hand, in the decomposed data as shown on the right side of the arrow in FIG. 12, an abnormality can be detected by simply setting a threshold value. As described above, the wavelet transform has a function of detecting an abnormality at a higher level while inheriting a simple operation of detecting an abnormality based on a normal threshold.

This embodiment applies a wavelet transform, a system for performing abnormality detection by fragmenting the T ² statistic Q statistic and Hotelling. In this case, the present embodiment is configured to improve the accuracy of abnormality detection by setting an appropriate threshold value.

In short, when the process data normalization model generation unit 41 determines the normalization shift parameter and scaling parameter value, the abnormality detection model generation unit 42 converts the wavelet transform into a Q statistic and a hotelling T ² statistic. By applying to the above, the calculation model (P, ∧, Wi) necessary for detecting the abnormality is determined. Further, a calculation formula for normalization using these values, a Q statistic decomposed for each frequency band, and a calculation formula for Hotelling's T ² statistic decomposed for each frequency band are generated.

Next, the abnormality determination criterion setting unit 5 determines a threshold value for performing abnormality determination using the equations (15) and (16), which are abnormality detection models. At this time, using the normalized data X which is the process monitoring model construction data, the Q statistic decomposed for each frequency band of the process monitoring model construction data calculated by the above formulas (15) and (16), and Use Hotelling's T ² statistics broken down by frequency band.

First, the threshold value determination prior information input unit 51 inputs a priori information about the Q statistics and T ² statistic Hotelling decomposed for each frequency band. Incidentally, if it can not obtain a priori information decomposed per this frequency band, and prior information about the process data, it may be replaced by a priori information for the T ² statistic Q statistic and Hotelling.

  Next, the threshold determination unit 52 determines a threshold for abnormality determination according to the prior information input from the threshold determination prior information input unit 51.

First, if prior information is nothing (information about T ² statistic Q statistic and Hotelling unknown), the threshold value determining unit 52, the default setting process, to employ the following method.

As shown in FIG. 11, those of the slowest frequency in T ² statistic Q statistic and Hotelling decomposed for each frequency band, turned approximation of T ² statistic original Q statistic and Hotelling ing. Thus, Q statistic statistical confidence limits and statistical confidence limits about the T ² statistic Hotelling, the threshold i.e. Q statistic and Hotelling T ² statistic was degraded as it is the slowest frequency band Q It is used as a threshold for statistics and hotelling T ² statistics.

On the other hand, T ² statistic slowest decomposed into each frequency band other than the Q statistic and Hotelling, as shown in FIG. 11 shows the variation distribution in the positive and negative directions around the zero mean. For this reason, since there is no prior information, the sample standard deviation (σrob) or robust sample of the Q statistic decomposed for each frequency band and the Hotelling T ² statistic according to the method used in normal statistical process monitoring Using the standard deviation (σrob), 3σ or 3σrob, which is three times the value, is set as a threshold value.

Next, when information indicating that almost all data is normal is input as prior information from the threshold determination prior information input unit 51, the threshold determination unit 52 is in principle configured for process monitoring model construction. The Q statistic decomposed for each frequency band calculated using the data normalization data X and the maximum value of the Hotelling T ² statistic decomposed for each frequency band are determined as threshold values. However, since there is a possibility that anomalous data is included in a very small amount, the Q statistic decomposed for each frequency band and the hotelling decomposed for each frequency band, for example, using the concept of 99% confidence interval The method of determining the upper 1% value as the threshold value from the maximum value of the T ² statistic is more preferable. In this case, the maximum value is obtained for each frequency band, but since each frequency band is different, each value is determined as a threshold value for each frequency band.

Here, also in the case of abnormality detection by the contribution amount in the second embodiment described above, a plurality of threshold values are calculated, but the contribution amount is the contribution of each variable to the Q statistic and the Hotelling T ² statistic. Therefore, it is common to evaluate with the same standard (threshold value). However, the present embodiment, since the performing abnormality detection by components of different frequency band T ² statistic Q statistic and Hotelling, essentially should set different thresholds.

Furthermore, when information indicating that abnormal data and normal data are mixed but the ratio of abnormal data is unknown is input as prior information, the threshold value determination unit 52, for example, a threshold value determination method for the modified Otsu, etc. Is used to determine the threshold of the Q statistic decomposed for each frequency band and the hotelling T ² statistic decomposed for each frequency band.

However, even if it is input as prior information that the abnormal data and the normal data are mixed, the abnormality is not always mixed with the statistical amount decomposed for each frequency band. For this reason, when there is a frequency band in which abnormal data is not mixed, if the threshold is determined by a clustering method such as the modified Otsu's threshold determination method, normal data will be divided into two, and an appropriate threshold is set. Not determined. However, unlike the abnormality diagnosis by the contribution amount, the abnormality diagnosis by T ² statistic of degraded Q statistic and Hotelling for each frequency band, it can not be taken out typical threshold for each respective measurement variable A threshold must be set. To address this, a threshold determined by the clustering method such as modification Otsu threshold determination method, in (a) Q statistic and Hotelling T ² statistic related information in the case of unclear threshold determination method When the threshold value determined by the clustering method is extremely small by comparing the value with the determined threshold value, the method of adopting the threshold value in the case (a) can be adopted. For example, when a threshold value that is less than 50% of the threshold value in the case of (a) is obtained by the clustering method, the rule of adopting the threshold value in the case of (a) is dealt with.

Further, when information indicating that abnormal data and normal data are mixed and the approximate ratio of abnormal data is known is input as prior information, the threshold value determination unit 52 uses the normalized data X. Q statistic and higher alpha% of the data decomposed into each frequency band T ² statistic Hotelling Q statistic decomposed for each corresponding frequency band in (alpha input value as the prior information) calculated Te and Hotelling Each value of the T ² statistic is calculated in advance. If, for all frequency bands, the approximate proportion of abnormal data is unknown and only known for some frequency bands, the threshold of the unknown frequency band is the same as that described above (a ). The proportion of abnormal data is roughly known, but if it is not known belong to which frequency bands, on the basis of the respective values of the degraded Q statistics and T ² statistic Hotelling for each frequency band , Once the threshold value is determined and compared with the threshold value determined by the threshold value determination method in the case of (a) above, and when a threshold value less than 50% of the threshold value in the case of (a) is obtained by the clustering method, ( Deal with the rule of adopting the threshold in the case of a).

  In short, in the abnormality determination criterion setting unit 5, the threshold determination unit 52 determines a threshold for determining whether the process 1 is normal or abnormal according to the prior information input from the threshold determination prior information input unit 51. The above series of operations is for constructing a process monitoring model and setting a criterion for abnormality determination, and is normally performed offline.

On the other hand, the following actions are executed online. First, the process monitoring unit 6 monitors the process 1 using the process monitoring model constructed by the model construction unit 4. Process monitoring unit 6, a process variable to be monitored, the T ² statistic Q statistic and Hotelling newly generated from the various sensors 11 to 1N, for each frequency band decomposed by the wavelet transform Q statistic and Hotelling monitoring the T ² statistic. That is, the process monitoring unit 6 uses the following formulas (17) and (18) to monitor the Q statistic decomposed for each frequency band and the hotelling T ² statistic.

That is, the process monitoring target is not a Q statistic or a Hotelling T ² statistic, but a Q statistic for each frequency band or a Hotelling T ² statistic obtained by decomposing them by wavelet transform.

Next, the process abnormality diagnosis unit 7 determines that one of the Q statistic and the hoteling T ² statistic for each frequency band obtained by decomposing the Q statistic and the hotelling T ² statistic by the wavelet transform is a threshold value. If it exceeds, it is judged as abnormal. The operations of the process abnormality factor sensor specifying unit 8, the process abnormality factor estimation unit 9, and the user interface unit 10 are exactly the same as those in the first embodiment.

As described above, according to the third embodiment, in the process monitoring by MSPC, the Q statistic for each frequency band and the hotelling T ² statistic obtained by decomposing the data to be subjected to abnormality detection using the wavelet transform are used. Thus, it is possible to improve the function for abnormality detection. In addition, by appropriately determining a threshold value for abnormality detection according to prior information for process data, errors in abnormality detection can be reduced. As a result, the performance of process abnormality monitoring by MSPC can be improved.

  In short, according to each embodiment, the error detection accuracy and the error factor identification accuracy are improved by appropriately adjusting the trade-off between an error that misses an abnormal state and an error that is detected as an abnormality even in a normal state. A process abnormality diagnosis system can be provided. In addition, process abnormality diagnosis with an adaptive learning function that appropriately reflects the subjective judgments of process supervisors (process operators and process managers) on abnormality detection accuracy and abnormality factor identification accuracy A system can be provided.

  In addition, by applying the process abnormality diagnosis device of each of the above-described embodiments, it is possible to construct a process monitoring system that provides a service that accumulates and provides process data and abnormality diagnosis results of process 1 in a database managed by a server. This system provides process data and an abnormality diagnosis result to the user terminal operated by the operator who manages and operates the process 1 from the server via the Internet or a dedicated communication line.

  Note that the present invention is not limited to the above-described embodiment as it is, and can be embodied by modifying the constituent elements without departing from the scope of the invention in the implementation stage. In addition, various inventions can be formed by appropriately combining a plurality of components disclosed in the embodiment. For example, some components may be deleted from all the components shown in the embodiment. Furthermore, constituent elements over different embodiments may be appropriately combined.

The block diagram which shows the principal part of the process abnormality diagnostic apparatus regarding the 1st Embodiment of this invention. The figure for demonstrating the input method of the prior information for normalization regarding this embodiment. The figure for demonstrating the concept of the threshold value determination method regarding this embodiment. The figure which shows the concept for setting the threshold value which classify | categorizes normal data and abnormal data regarding this embodiment. The figure which shows the concept for setting the threshold value which classify | categorizes normal data and abnormal data regarding this embodiment. The figure which shows the concept of the data decomposition calculation of PCA regarding this embodiment. Diagram for explaining the concept of the Q statistic and T ² statistic of Hotelling according to the embodiment. The figure for demonstrating specification of the process abnormality factor at the time of a sensor failure regarding this embodiment. The figure for demonstrating specification of the process abnormality factor at the time of process abnormality regarding this embodiment especially. The figure which shows the concept of the contribution amount with respect to Q statistic regarding 2nd Embodiment. The figure which shows the specific example which decomposed | disassembled the Q statistic regarding 3rd Embodiment into the data for every several frequency band by discrete wavelet transform. The figure for demonstrating the effect of the abnormality detection rate regarding 3rd Embodiment. The figure for demonstrating the effect of 1st Embodiment, and the figure which shows the relationship between data distribution and an error type. The figure for demonstrating the effect of 1st Embodiment, and the figure for demonstrating the content of error type A and B. FIG. The figure for demonstrating the effect of 1st Embodiment, and the figure which shows the specific example for erroneously detecting normal data as abnormality. The figure for demonstrating the effect of 1st Embodiment, The figure which shows the specific example which overlooks abnormal data.

Explanation of symbols

1 ... Process, 2 ... Process measurement data collection / storage unit,
3 ... Process data extraction unit for process monitoring model construction,
4 ... Process monitoring model construction / supply unit, 5 ... Process abnormality judgment standard setting / supply unit,
6 ... Process monitoring unit, 7 ... Process abnormality diagnosis unit, 8 ... Process abnormality factor sensor identification unit,
9 ... Process abnormality factor estimation unit, 10 ... User interface unit,
41 ... Process data normalization model generation unit,
42: Anomaly detection model generation unit, 51: Threshold determination prior information input unit, 52: Threshold determination unit.

Claims (3)

A measuring means for measuring the state or operation amount of the target process;
Data storage means for collecting and storing time-series data indicating measurement results from the measurement means;
Process data extraction means used to create a process monitoring model from the time series data stored in the data storage means, and to extract process data indicating the state of the target process;
Normalization model generation means for generating a normalization model for normalizing the process data extracted by the process data extraction means;
A process monitoring model based on a principal component analysis method for the process data normalized by the normalization model is constructed, and Q statistics or hotelling T ² statistics statistics data is used as abnormality detection data of the normalized process data. Process monitoring model means for calculating the contribution amount of the measurement item (measurement variable) measured by the measurement means for the statistical data ,
Prior information relating to a ratio between abnormal data and normal data included in the statistical data, and prior information generating means for generating the prior information for the contribution amount ;
Threshold determination means for determining a threshold for determining an abnormal state and a normal state of the statistical data using the prior information and the contribution amount of the statistical data calculated by the process monitoring model means; ,
A process abnormality diagnosis device comprising: an abnormality diagnosis unit that diagnoses an abnormality of the target process using the threshold value determined by the threshold value determination unit and the statistical data. A measuring means for measuring the state or operation amount of the target process;
Data storage means for collecting and storing time-series data indicating measurement results from the measurement means;
Process data extraction means used to create a process monitoring model from the time series data stored in the data storage means, and to extract process data indicating the state of the target process;
Normalization model generation means for generating a normalization model for normalizing the process data extracted by the process data extraction means;
A process monitoring model based on a principal component analysis method for the process data normalized by the normalization model is constructed, and Q statistics or hotelling T ² statistics statistics data is used as abnormality detection data of the normalized process data. And process monitoring model means for calculating the statistical data decomposed for each frequency band by applying a discrete wavelet transform,
Prior information generation means for generating the prior information for the statistical data decomposed for each frequency band, which is prior information on the ratio between abnormal data and normal data included in the statistical data;
Using the prior information and the statistical data decomposed for each frequency band calculated by the process monitoring model means, a threshold for determining the abnormal state and the normal state of the statistical data is determined. Threshold determination means;
An abnormality diagnosis means for diagnosing an abnormality of the target process using the threshold value determined by the threshold value determination means and the statistical data;
A process abnormality diagnosis apparatus comprising: The process monitoring model means calculates a contribution amount of the statistic data decomposed for each frequency band by applying a discrete wavelet transform,
The prior information generating means generates the prior information for the contribution amount of the statistical data decomposed for each frequency band,
The threshold value determination means uses the contribution amount of the statistical data decomposed for each frequency band calculated by the prior information and the process monitoring model means to determine an abnormal state and a normal state The process abnormality diagnosis apparatus according to claim 1, wherein a threshold value is determined.

Images