JP2010271997A - Standard time series data calculation method, abnormality detection method, standard time series data calculation device, abnormality detection device, standard time series data calculation program, and abnormality detection program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a standard time series data calculation method, a standard time series data calculation device, and a standard time series data calculation program which calculate standard time series data representing a normal state, and to provide an abnormality detection method, an abnormality detection device, and an abnormality detection program which use the standard time series data to detect an abnormality. <P>SOLUTION: The standard time series data calculation method includes: a data acquisition step of acquiring a plurality of pieces of time series data about a plurality of variables in one system; a data normalization step of generating a plurality of pieces of normalized time series data from the plurality of pieces of time series data; a warping path calculation step of obtaining a warping path by selecting masses on the basis of a distance in a DTW description using two of the plurality of pieces of normalized time series data generated in the data normalization step; and a standard time series data calculation step of calculating standard time series data comprising time-series sets of a plurality of average values per mass on the warping path. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to a technique for calculating standard time series data representing a normal state of a system defined by a plurality of variables, and a technique for detecting a system abnormality using the standard time series data.

  In order to operate a production facility in a normal state, it is necessary to control various variables in the production facility, for example, various variables such as temperature, pressure, and flow rate according to the production facility.

  For example, in a vertical oxidation diffusion furnace for manufacturing a semiconductor, wafers are replaced and repeated processing is performed, and it is necessary to repeatedly control the temperature and pressure in the oxidation diffusion furnace for each processing. . However, since it is difficult to ideally control all the processes, it is normal to always have an error within a normal range. Furthermore, if the error increases, an abnormal state may occur.

  If the abnormality of the production equipment as in the above example can be checked, it can be used for omission of the inspection process, early detection of equipment failure, and the like. A method is known in which standardization is performed based on measurement values of a plurality of variables such as temperature and pressure measured at a production facility, and the result is used for finding an abnormal state. Here, normalization means creating data representing the standard behavior of these variables in the normal state. In this standardization, if the normal state is measured and standardized only once, there may be an error in the data measured due to the influence of noise or the like, and ideal standard data may not be obtained. Therefore, in order to reduce the influence of such errors, a method for standardizing by measuring the normal state a plurality of times has been proposed.

  Multivariate statistical process management called Multi-Way Principal Component Analysis (hereinafter referred to as multi-way PCA) has been proposed as one of the techniques for performing this standardization (see, for example, Patent Document 1).

  In multi-way PCA, normalization is performed using a plurality of time-series data obtained by sampling a plurality of variables such as temperature and pressure by a predetermined number of samples at a predetermined sampling interval. Is called. The standardized time series data is used for detecting an abnormal state.

JP 2007-65883 A

  Since the multi-way PCA does not support expansion and contraction of the time axis, standardization becomes difficult if there is a difference in sampling interval between a plurality of time-series data used for standardization or the number of samples is different. In addition, it is difficult to detect an abnormal state even when the sampling interval between the time-series data that is an abnormality detection target and the standardized time-series data is shifted or the number of samples is different. However, if the sampling intervals and the number of samples of a plurality of time-series data that are actually used for standardization and detection of an abnormal state are all made uniform, the degree of freedom of data collection is limited.

  In view of the above circumstances, the present invention provides highly accurate standard time-series data representing the normal state of a system defined by a plurality of variables, even if the sampling interval and the number of samples are different among a plurality of time-series data used for standardization. It is an object of the present invention to provide a standard time-series data calculation method and a standard time-series data calculation device that can be calculated. In addition, even if the sampling time interval and the number of samples are different between the standard time-series data and the time-series data to be detected, the present invention uses the standard time-series data, for example, a single control target such as a production facility. An object of the present invention is to provide an abnormality detection method and an abnormality detection device for detecting an abnormality in a system. Furthermore, the present invention is executed in an arithmetic processing device that executes a program, and is executed in a standard time series data calculation program that operates the arithmetic processing device as the standard time series data calculation device, and an arithmetic processing device that executes the program, It is an object of the present invention to provide an abnormality detection program for operating this arithmetic processing device as the abnormality detection device.

In order to achieve the above object, the standard time series data calculation method of the present invention is:
Time-series data composed of time-series sets of a plurality of data obtained by sampling a plurality of variables defining one system at the same sampling time, and sequentially sampling a plurality of variables of the one system A data acquisition step for acquiring a plurality of time series data collected in a time zone;
By multiplying multiple data for each variable across multiple time series data acquired in the above data acquisition step by a coefficient for each variable, the multiple data for each variable can be converted to each other without depending on the variable. A data normalization step for generating a plurality of normalized time series data obtained by normalizing each of the plurality of time series data by performing normalization distributed within the same numerical range,
The first normalized time series data of the plurality of normalized time series data generated in the data normalization step is arranged on one axis, and the second normalized time series of the plurality of normalized time series data is arranged. In DTW notation in which data is arranged on the other axis, a plurality of variables corresponding to the plurality of variables on one axis corresponding to each square divided for each sampling time of two normalized time series data arranged on the one axis and the other axis Calculate the total distance for each of the variables between the data and the data corresponding to the variables on the other axis corresponding to each cell, and calculate the mass based on this distance. A warping path calculation step for obtaining a warping path by selecting;
By calculating an average value for each variable between a plurality of data on one axis and a plurality of data on the other axis corresponding to each square on the warping path, each square on this warping path is calculated. A standard time-series data calculating step for calculating standard time-series data including a time-series set of a plurality of average values.

  Here, the “total distance” in the present invention is a distance based on the difference for each variable corresponding to each cell, and means, for example, the Euclidean distance, the Mahalanobis general distance, the Manhattan distance, or the like. Further, “based on distance” in the present invention means, for example, tracing a square having the smallest distance in time series or selecting a square having a distance equal to or less than a specified value.

  According to the standard time-series data calculation method of the present invention, high-precision standard time-series data is calculated even when the sampling intervals of the time-series data used for calculating the warping path are shifted or the number of samples is different.

  Here, the standard time-series data calculation method of the present invention uses the standard time-series data calculated in the standard time-series data calculation step as a new first normalized time-series data that replaces the first normalized time-series data. And the third normalized time series data of the plurality of normalized time series data normalized in the data normalization step is replaced with the second normalized normal time data instead of the second normalized time series data. In the DTW notation arranged on the other axis as the normalized time series data, the new third normalized time series data is sequentially used, and the new first time series arranged on the one axis when executing the standard time series data calculation step. Repeating the warping path calculation step and the standard time series data calculation step while performing averaging with weighting according to the number of repetitions in the past. More, it is preferable to calculate the final standard time series data.

  As described above, when calculation of standard time series data is repeated using the new third normalized time series data sequentially, at least three or more time series data are used to calculate more accurate standard time series data. Is done.

  In addition, the standard time series data calculation step includes an average value between sampling times and each variable between a plurality of data on one axis and a plurality of data on the other axis corresponding to each square on the warping path. It may be a step of calculating standard time-series data consisting of a time-series set of average values for each square on the warping path by calculating each average value for each cell.

  When the start times of a plurality of time series data are aligned and the sampling times measured from the start time are not aligned between the time series data, the sampling time is also regarded as one data and an average value is calculated. Thus, the standard time of sampling can be calculated, and highly accurate standard time series data including the sampling standard time can be calculated.

In order to achieve the above object, the abnormality detection method of the present invention includes:
An abnormality detection method for detecting an abnormality of the system using the standard time series data calculated by the standard time series data calculation method according to any one of claims 1 to 3, wherein the plurality of the one system Anomaly detection object data acquisition step for acquiring anomaly detection object time series data consisting of a set of time series of a plurality of data obtained by sampling the variables of
By multiplying the data for each variable constituting the abnormality detection target time series data by the coefficient for each variable adopted in the data normalization step, the abnormality detection target time series data is normalized at the time of abnormality detection target. Anomaly detection target data normalization step for converting to series data;
In the DTW notation in which the standard time series data is arranged on one axis and the abnormality detection target time series data normalized in the abnormality detection target data normalizing step is arranged on the other axis, two time series data arranged on one axis and the other axis Between a plurality of data corresponding to the plurality of variables on one axis corresponding to each square partitioned at each sampling time and a plurality of data corresponding to the plurality of variables on the other axis corresponding to each square The total distance for each of the variables is calculated for each cell, and the total distance between the standard time series data and the normalized target time series data is calculated based on the distance. A calculation step;
And an abnormality detection step of detecting that the system is abnormal when the total distance calculated in the total distance calculation step is larger than a predetermined threshold.

  According to the anomaly detection method of the present invention, the anomaly detection target data and the standard time series data have high accuracy of the above system even if the sampling time when the start time is aligned or the number of samples is different. Abnormality detection can be performed.

In addition, the standard time series data calculation apparatus of the present invention for achieving the above-described object is
Time series data composed of a set of time series of a plurality of data obtained by sampling a plurality of variables defining one system at the same sampling time, and sequentially differing for the plurality of variables of the one system A data acquisition unit for acquiring a plurality of time-series data collected in a sampling time period;
Multiply multiple data for each variable across multiple time-series data acquired by the above data acquisition unit by multiplying the coefficient for each variable, and convert the multiple data for each variable without depending on the variable A data normalization unit that generates a plurality of normalized time series data obtained by normalizing each of the plurality of time series data by performing normalization to distribute within the same numerical range,
The first normalized time series data of the plurality of normalized time series data generated by the data normalization unit is arranged on one axis, and the second normalized time series of the plurality of normalized time series data In DTW notation in which data is arranged on the other axis, a plurality of variables corresponding to the plurality of variables on one axis corresponding to each square divided for each sampling time of two normalized time series data arranged on the one axis and the other axis Calculate the total distance for each of the variables between the data and the data corresponding to the variables on the other axis corresponding to each cell, and based on this distance A warping path calculation unit for obtaining a warping path by selecting a square;
By calculating an average value for each variable between a plurality of data on one axis and a plurality of data on the other axis corresponding to each square on the warping path, each square on this warping path is calculated. And a standard time-series data calculating unit that calculates standard time-series data including a time-series set of a plurality of average values.

  Here, the standard time series data calculation apparatus of the present invention uses the standard time series data calculated by the standard time series data calculation unit as a new first normalized time series data to replace the first normalized time series data. And the third normalized time series data of the plurality of normalized time series data normalized by the data normalization unit is replaced with the second normalized time series data. In the DTW notation arranged on the other axis as the normalized time series data, the new third normalized time series data is sequentially used, and the new first time series data arranged on one axis is added to the standard time series data calculation unit. The final standard time system is obtained by alternately and repeatedly operating the warping path calculation unit and the standard time series data calculation unit while performing averaging with weighting according to the number of past repetitions. It is preferable data are those provided with an operation control unit for calculating a.

  Further, the standard time series data calculation unit is configured to calculate an average value and each variable between sampling times between a plurality of data on one axis corresponding to each square on the warping path and a plurality of data on the other axis. It is also possible to calculate standard time-series data consisting of a time-series set of average values for each square on the warping path by calculating each average value for each cell.

In order to achieve the above object, the abnormality detection device of the present invention comprises:
A standard time-series data calculation device according to any one of claims 5 to 7,
The data acquisition unit further comprises an abnormality detection comprising a time series set of a plurality of data obtained by sampling the plurality of variables of the one system at the same sampling time in the sampling time zone of the abnormality detection target. To obtain target time series data,
The data normalization unit further multiplies a plurality of data for each variable constituting the abnormality target time series data by a coefficient for each variable adopted in the data normalization unit, thereby detecting the abnormality detection target time series data. To normalization abnormality detection target time series data,
In the DTW notation in which the standard time series data is arranged on one axis and the abnormality detection target time series data normalized in the abnormality detection target data normalizing step is arranged on the other axis, two time series data arranged on one axis and the other axis Between a plurality of data corresponding to the plurality of variables on one axis corresponding to each square divided at each sampling time and a plurality of data corresponding to the plurality of variables on the other axis corresponding to each square The total distance of the plurality of variables is calculated for each square, and the total distance between the standard time series data and the normalized target time series data is calculated based on the distance. A calculation unit;
And an abnormality detection unit that detects an abnormality when the total distance calculated by the total distance calculation unit is greater than a predetermined threshold.

In addition, the standard time series data calculation program of the present invention for achieving the above object is
It is executed in an arithmetic processing unit that executes a program.
Time series data composed of a set of time series of a plurality of data obtained by sampling a plurality of variables defining one system at the same sampling time, and sequentially differing for the plurality of variables of the one system A data acquisition unit for acquiring a plurality of time-series data collected in a sampling time period;
Multiply multiple data for each variable across multiple time-series data acquired by the above data acquisition unit by multiplying the coefficient for each variable, and convert the multiple data for each variable without depending on the variable A data normalization unit that generates a plurality of normalized time series data obtained by normalizing each of the plurality of time series data by performing normalization to distribute within the same numerical range,
The first normalized time series data of the plurality of normalized time series data generated by the data normalization unit is arranged on one axis, and the second normalized time series of the plurality of normalized time series data In DTW notation in which data is arranged on the other axis, a plurality of variables corresponding to the plurality of variables on one axis corresponding to each square divided for each sampling time of two normalized time series data arranged on the one axis and the other axis Calculate the total distance for each of the variables between the data and the data corresponding to the variables on the other axis corresponding to each cell, and based on this distance A warping path calculation unit for obtaining a warping path by selecting a square;
By calculating an average value for each variable between a plurality of data on one axis and a plurality of data on the other axis corresponding to each square on the warping path, each square on this warping path is calculated. It is characterized by operating as a standard time-series data calculation device having a standard time-series data calculation unit for calculating standard time-series data consisting of a plurality of average values in time series.

Here, the standard time series data calculation program of the present invention is
The standard time-series data calculated by the standard time-series data calculating unit is arranged on one axis as new first normalized time-series data that replaces the first normalized time-series data, and is normalized by the data normalizing unit. In the DTW notation in which the third normalized time series data among the plurality of normalized time series data is arranged on the other axis as new second normalized time series data replacing the second normalized time series data. Then, using the new normalized third time series data sequentially, and averaging the new first time series data arranged on one axis with a weight corresponding to the number of past repetitions in the standard time series data calculation unit The operation control unit is configured to calculate final standard time series data by alternately and repeatedly operating the warping path calculation unit and the standard time series data calculation unit. Rukoto is preferable.

  Further, the standard time series data calculation unit is configured to calculate an average value and each variable between sampling times between a plurality of data on one axis corresponding to each square on the warping path and a plurality of data on the other axis. It is also possible to calculate standard time-series data consisting of a time-series set of average values for each square on the warping path by calculating each average value for each cell.

In order to achieve the above object, the abnormality detection program of the present invention provides:
An abnormality detection program including the standard time-series data calculation program according to any one of claims 9 to 11 as a program part, wherein the arithmetic processing unit is
The data acquisition unit further comprises an abnormality detection comprising a time series set of a plurality of data obtained by sampling the plurality of variables of the one system at the same sampling time in the sampling time zone of the abnormality detection target. To obtain target time series data,
The data normalization unit further multiplies a plurality of data for each variable constituting the abnormality target time series data by a coefficient for each variable adopted in the data normalization unit, thereby detecting the abnormality detection target time series data. To normalization abnormality detection target time series data,
In the DTW notation in which the standard time series data is arranged on one axis and the abnormality detection target time series data normalized in the abnormality detection target data normalizing step is arranged on the other axis, two time series data arranged on one axis and the other axis Between a plurality of data corresponding to the plurality of variables on one axis corresponding to each square divided at each sampling time and a plurality of data corresponding to the plurality of variables on the other axis corresponding to each square The total distance of the plurality of variables is calculated for each square, and the total distance between the standard time series data and the normalized target time series data is calculated based on the distance. A calculation unit;
It is operated as an abnormality detection device having an abnormality detection unit that detects abnormality when the total distance calculated by the total distance calculation unit is larger than a predetermined threshold value.

  According to the present invention, the normal state of a system defined by a plurality of variables even if the sampling time and the number of samples differ when the time points of the time series data are aligned among the plurality of time series data used for standardization. High-precision standard time-series data representing is calculated. Also, when detecting anomalies, even if the sampling time and the number of samples differ between the standard time-series data and the time-series data subject to abnormality detection, An abnormality is detected.

It is a figure which shows the computer with which one Embodiment of this invention is applied. It is a hardware block diagram of the computer shown in FIG. It is a conceptual diagram which shows CD-ROM in which the standard time series data calculation program was memorize | stored. It is a functional block diagram of a standard time series data calculation device. It is a flowchart which shows the flow of calculation of the standard time series data in a standard time series data calculation apparatus. It is a figure which shows the three time series data acquired by step S1. It is the graph which showed the change of each variable for every time about three time series data shown in FIG. FIG. 7 is a diagram illustrating three normalized time series data obtained by normalizing the three time series data of FIG. 6. It is a figure which shows DTW notation. It is a figure which shows the average value about the data regarding each square which forms a warping path | pass. It is a figure which shows DTW notation. It is a figure which shows the weighted average value about the data regarding each square which forms a warping path | pass. It is a figure which shows standard time series data denormalized with standard time series data. It is a conceptual diagram which shows CD-ROM with which the abnormality detection program was memorize | stored. It is a functional block diagram of an abnormality detection device. It is a flowchart which shows the flow about the abnormality detection in an abnormality detection apparatus. It is a figure which shows the time series data which are the objects of abnormality diagnosis. It is the graph showing the change of each variable by the time of three time series data of the normal state used for calculation of standard time series data, and the time series data which are the objects of abnormality diagnosis. It is the graph which showed each total distance of three time series data of a normal state with respect to standard time series data, and the time series data which are the objects of abnormality diagnosis.

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

  FIG. 1 is a diagram illustrating a computer to which an embodiment of the present invention is applied.

  A computer 100 shown in FIG. 1 constitutes an embodiment of a standard time series data calculation apparatus of the present invention. The computer 100 also constitutes an embodiment of the abnormality detection apparatus of the present invention described later, and is shared by both embodiments. In appearance, the computer 100 includes a main body 110, a display 120 that displays an image on the display screen 121 according to an instruction from the main body 110, a keyboard 130 that inputs information corresponding to key operations to the computer 100, and the display screen 121. The mouse 140 for inputting an instruction corresponding to the position by designating an arbitrary position is provided.

  The main body 110 has an FD loading port 111 and a CD-ROM loading port 112 for loading a flexible disk (hereinafter referred to as FD) and a CD-ROM in appearance.

  FIG. 2 is a hardware configuration diagram of the computer shown in FIG.

  1 includes a CPU 113, a main memory 114, a hard disk device 115, an FD drive 116, a CD drive 117, an input interface 118, and an output interface 119. The CPU 113 executes various programs. The main memory 114 reads the program stored in the hard disk device 115 and develops the program for execution by the CPU 113. Various programs and data are stored in the hard disk device 115. The FD drive 116 accesses the FD 300. The CD drive 117 is loaded with the CD-ROM 310 and accesses the loaded CD-ROM 310. The input interface 118 inputs data from an external device, and the output interface 119 outputs data to the external device. These hardware built in the main device 110, the keyboard 130, the mouse 140, and the image display device 120 are connected via a bus 150.

  The CD-ROM 310 stores a standard time series data calculation program (to be described later) that causes a computer to operate as an example of the standard time series data calculation apparatus according to the present invention. The CD-ROM 310 is loaded into the CD drive 117, and the standard time series data calculation program stored in the CD-ROM 310 is uploaded to the computer 100 and stored in the hard disk device 115. When the standard time series data calculation program is activated and executed, the computer 100 operates as an example of the standard time series data calculation apparatus according to the present invention.

  Next, the standard time series data calculation program will be described.

  FIG. 3 is a conceptual diagram showing a CD-ROM in which a standard time series data calculation program is stored.

  As shown in FIG. 3, the standard time series data calculation program 311 stored in the CD-ROM 310 includes a data acquisition unit 312, a data normalization unit 313, a warping path calculation unit 314, a standard time series data calculation unit 315, and an operation control unit. 316.

  The details of each part of the standard time series data calculation program 311 will be described together with the operation of each part of the standard time series data calculation apparatus 410 described later.

  FIG. 4 is a functional block diagram of the standard time series data calculation apparatus.

  The block diagram shown in FIG. 4 shows a case where the standard time series data calculation program 310 shown in FIG. 3 is installed in the computer 100 of FIG. 1, and the computer 100 operates as the standard time series data calculation apparatus 410 according to an embodiment of the present invention. It represents the function of.

  4 includes a data acquisition unit 500, a data normalization unit 501, a warping path calculation unit 502, a standard time series data calculation unit 503, and an operation control unit 504. The data acquisition unit 500, data normalization unit 501, warping path calculation unit 502, standard time series data calculation unit 503, and operation control unit 504 are mainly performed by the CPU 113 shown in FIG. On the other hand, the storage device 505 is connected to the data acquisition unit 500 and the standard time series data calculation unit 503 in the standard time series data calculation device 410, and is mainly handled by the hard disk device 115 shown in FIG. . Further, the temperature sensor 420 and the pressure sensor 421 are connected to the standard time series data calculation device 410, and directly connected to the input interface 118 shown in FIG.

  Here, in the present embodiment, there is a device (hereinafter referred to as “target device”) that repeatedly performs the same processing, and the temperature and pressure inside the target device change according to a predetermined sequence for each processing. It is assumed that The temperature sensor 420 and the pressure sensor 421 measure the temperature and pressure in the target device. Here, the temperature and pressure are repeatedly measured for each of a plurality of processes while the target device performs a repetitive process, one time series data is acquired for each process, and a plurality of times over a plurality of processes are obtained. Based on the series data, standard time series data per process is calculated.

  In the data acquisition unit 500, during one process in the target device, each value measured by the temperature sensor 420 and the pressure sensor 421 is sampled at the same time, and one time-series data is acquired. And this sampling is repeated for every process, and several time series data are acquired.

  Although details will be described later, the data normalization unit 501 generates a plurality of normalized time-series data by normalizing a plurality of time-series data acquired by the data acquisition unit 500.

  The warping path calculation unit 502 selects two normalized time-series data among a plurality of normalized time-series data normalized by the data normalization unit 501, and the selected 2 A warping path, which will be described later, is calculated by a so-called DTW method in which one of the normalized time series data is arranged on one axis and the other normalized time series data is arranged on the other axis. Is done.

  The standard time series data calculation unit 503 calculates standard time series data based on the calculated warping path, as will be described in detail later.

  As will be described in detail later, the operation control unit 504 adds a new one from the standard time series data calculated by the standard time series data calculation unit 503 and a plurality of normalized time series data normalized by the data normalization unit 501. Using the selected normalized time series data, the warping path calculation unit 502 and the standard time series data calculation unit 503 repeatedly perform a cycle for causing the warping path calculation unit 502 to calculate new standard time series data.

  The storage device 505 stores time series data acquired by the data acquisition unit 500 and standard time series data calculated by the standard time series data calculation unit 503.

  FIG. 5 is a flowchart showing a flow of calculation of standard time series data in the standard time series data calculation apparatus shown in FIG.

  In step S <b> 1 shown in FIG. 5, the temperature and pressure measured by the temperature sensor 420 and the pressure sensor 421 are sampled at the same timing during one process of the target device described above. In this way, one time series data is obtained. Further, this sampling is repeated for each process of the target device, and a plurality of time-series data is acquired. This step S1 corresponds to an example of a data acquisition step according to the present invention. Hereinafter, in the present embodiment, description will be made assuming that three time-series data are acquired in step S1 when the target device is in a normal state. In the following, description will be made mainly with reference to FIG. 5, and the description of FIG. 6 and subsequent drawings will be interwoven as appropriate.

  FIG. 6 is a diagram showing the three time-series data acquired in step S1. 6A, 6B, and 6C show time-series data. Each time-series data has two variable values of “value1” representing temperature and “value2” representing pressure at each sampling time represented by “time”. Each time-series data has the same sampling interval. However, the number of samples (A) is 11 pieces, (B) is 12 pieces, and (C) is 11 pieces.

  FIG. 7 is a graph showing changes in temperature and pressure over time for the three time-series data shown in FIG. Comparing these graphs, although there are some errors, there is a common change that the pressure rises after the temperature rises and the pressure falls after the temperature falls. These three graphs are time-series data measured in a normal state, although there are some differences.

  After a plurality of time series data is acquired as described above, the process proceeds to step S2 in FIG. In step S2, all time series data is normalized for each variable of temperature and pressure. This step S2 corresponds to an example of a data normalization step according to the present invention. In the present embodiment, the values of the variables are normalized so that they are distributed within the same range.

  Here, in order to normalize the three time series data shown in FIG. 6 according to this step S2, the minimum value and the maximum value of the temperature and pressure are examined from the whole over these three time series data. First, when the minimum value and the maximum value of the temperature are examined, “value 1: 0.0000” corresponding to “time: 0.0000” in FIG. 6A is the minimum value, and “time” in FIG. : 'Value1: 3.2000' corresponding to 1.0000 'is the maximum value. Further, when the minimum value and the maximum value of the pressure are examined, “value 2: 0.0000” corresponding to “time: 0.0000” in FIG. 6A is the minimum value, and “time” in FIG. : 'Value 2: 2.0000' corresponding to: 5.0000 'is the maximum value. Therefore, each value of “value1” is divided by “3.2000” which is the difference between the maximum value “3.200” and the minimum value “0.0000”, and each value of “value2” is similarly divided into the maximum value “ Divide by “2.000”, which is the difference between 2.0000 ”and the minimum value“ 0.0000 ”. As a result, the range from the minimum value to the maximum value of each variable is normalized to the range from “0” to “1”.

  If this normalization step is not performed, the distance calculation described later will be performed using non-normalized time series data, and the distribution of values will be wider than the variables distributed in a small range. The distance affected by the variable is calculated. Since standard time series data to be described later is calculated based on this distance, standard time series data influenced by variables whose values are distributed over a wide range is calculated. On the other hand, a variable distributed only in a small range is less likely to be reflected in the standard time series data. In order to eliminate this effect and evaluate each variable equally, the normalization described above is performed. In this embodiment, normalization is performed based on the minimum value and the maximum value of each variable. However, any standard may be used for equally evaluating each variable. For example, for temperature, from '0.0000' When it is known that the pressure fluctuates within the range of “3.5000” and the pressure fluctuates within the range of “0.0000” to “3.0000”, the three time series data shown in FIG. Without being constrained by the minimum and maximum values, the temperature is normalized by dividing by “3.5000”, and the pressure is normalized by dividing by “3.0000”. Good. However, here, the description will be continued assuming that normalization using the maximum value and the minimum value appearing on the three time-series data shown in FIG. 6 described above is adopted.

  FIG. 8 is a diagram showing three normalized time series data obtained by normalizing the three time series data of FIG. FIG. 8A shows time-series data obtained by normalizing the time-series data of FIG. 6A according to step S2. Similarly, FIGS. 8B and 8C are time-series data obtained by normalizing the time-series data of FIGS. 6B and 6C in accordance with step S2. Hereinafter, the time series data shown in FIGS. 8A, 8B, and 8C are referred to as normalized time series data A, normalized time series data B, and normalized time series data C, respectively.

  The range of “value1” spanning the three time series data of FIGS. 6A, 6B, and 6C is “0.0000” to “3.2000”, but is normalized by the step S2. The range is converted from '0.0000' to '1.000'. For example, the minimum value “0.0000” before normalization indicated by “value1” corresponding to “time: 0.0000” in FIG. 6A is the minimum value after normalization by normalization. Converted to '0.0000'. This value is the value of “value1” corresponding to “time: 0.0000” in FIG. The maximum value “3.2000” before normalization shown in “value1” corresponding to “time: 1.0000” in FIG. 6C is “1” which is the maximum value after normalization by normalization. Converted to .0000 '. This value is the value of “value1” corresponding to “time: 1.0000” in FIG.

  The same applies to the normalization of 'value2'. That is, the range of 'value2' straddling the three time-series data in FIGS. 6A, 6B, and 6C is “0.0000” to “2.0000”, but the normal value of step S2 The range is converted from '0.0000' to '1.0000' by conversion. For example, the minimum value “0.0000” before normalization indicated by “value2” corresponding to “time: 0.0000” in FIG. 6A is the minimum value after normalization by normalization. Converted to 0.0000 '. This value is the value of “value2” corresponding to “time: 0.0000” in FIG. The maximum value “2.0000” before normalization indicated by “value 2” corresponding to “time: 5.0000” in FIG. 6B is “1.0000, which is the maximum value after normalization by normalization. Converted to '. This value is the value of “value2” corresponding to “time: 5.0000” in FIG. In this way, the values of “value1” and “value2” of the three normalized time series data of FIG. 8 are converted so as to be distributed within the range of “0.0000” to “1.0000”, respectively. Yes.

  .., B1,..., C1,..., Etc., shown beside each time series data shown in FIG. When a set of variables of 'time', 'value1', and 'value2' is expressed as ('time', 'value1', 'value2'), for example, FIG. 8A is configured. ('time', ' Data whose values (value1 ',' value2 ') are (' 0.0000 ',' 0.0000 ',' 0.0000 ') is called data A1. The same applies to other codes other than A1.

  After the time series data is normalized as described above, the process proceeds to step S3 shown in FIG. In step S3, two of the three normalized time series data normalized in step S2 are selected. Then, a warping path in DTW notation is calculated for the two selected normalized time series data. This step S3 corresponds to an example of a warping path calculation step according to the present invention.

  Here, a method of using DTW notation using three or more time-series data at the same time for the calculation of the warping path can be considered, but the processing becomes complicated. At the same time, even with DTW notation using only two time-series data, it is possible to calculate a warping path using three or more time-series data by a method described later. Therefore, the present invention adopts DTW notation that uses two normalized time series data at the same time. Note that the normalized time series data that is not selected in step S3 is used in the processing described later. Hereinafter, the description will be continued assuming that the normalized time-series data A and the normalized time-series data B are selected from the three normalized time-series data in FIG.

  FIG. 9 is a diagram showing DTW notation.

  In FIG. 9, the normalized time series data A shown in FIG. 8A is arranged on the vertical axis, and the normalized time series data B shown in FIG. 8B is arranged on the horizontal axis.

  In addition, each of the squares divided for each sampling time of the time series data on the vertical axis and the horizontal axis shown in FIG. 9 includes 'value1' and 'value2' on the vertical axis corresponding to each square, and each square. The Euclidean distance between 'value1' and 'value2' on the horizontal axis corresponding to is described. Here, 'value1' and 'value2' on the vertical axis corresponding to a square with DTW notation are 'Va1' and 'Va2', respectively, and 'value1' and 'value2' on the horizontal axis are 'Vb1', respectively. And “Vb2”, the Euclidean distance d of the square is calculated by the following equation (1).

  This Euclidean distance corresponds to an example of the distance referred to in the present invention. For example, in the square corresponding to the data A2 and the data B4 in the DTW notation shown in FIG. 9, the Euclidean distance '0.2577' between the data A2 and the data B4 is described. This is obtained by subtracting the “value1” value “0.8750” of the data B4 from the “value1” value “0.9375” of the data A2 and squaring, and the “value2” value “0.0000” of the data A2. Is a square root obtained by subtracting the “value2” value “0.2500” of the data B4 and squared. Although the Euclidean distance is used in the present embodiment, the present invention is not limited to this. For example, the Mahalanobis general distance expressed by the following expression (2) and the Manhattan expressed by the expression (3) are shown below. It may be a distance.

  Here, VA and VB in the above equation (2) mean the following equations (4) and (5), respectively.

  In addition, when a matrix such as (VA-VB) in the above equation (2) is represented by X,

Represents a transposed matrix of the matrix X. Furthermore, in the formula (2)

Represents the inverse of the covariance matrix.

  Furthermore, in the formula (3)

Is a symbol representing an absolute value.

  Here, the above-mentioned warping path refers to a trajectory when a square is traced from the lower right corner to the upper left corner so that the sum of the distances written in each square is minimized. In FIG. 9, the square forming the warping path is surrounded by a thick black frame.

  When the above-described calculation of the warping path is completed, the process proceeds to step S4 shown in FIG. In this step S4, it is determined whether or not one of the time series data used for the calculation of the current warping path is provisional standard time series data. The provisional standard time series data will be described later, but the provisional standard time series data has not yet been calculated at this stage. Accordingly, the process proceeds to step S6 here. In step S6, a simple average is calculated for each of 'time', 'value1', and 'value2' of the data on each axis corresponding to each square forming the warping path calculated in step S3. Standard time-series data is obtained by arranging the calculated simple averages in time series along the warping path. This step S6 constitutes an example of the standard time series data calculation step referred to in the present invention.

  In the present embodiment, the case where there are three time series data used for the calculation of the standard time series data is described. However, when the standard time series data is calculated using two time series data, the following steps are performed. In S7, it is determined that all normalized time series data has been used, and the calculation of the standard time series data is terminated. When these two time series data are used, the standard time series data calculated in step S6 is output as it is as the final standard time series data. Even when only two pieces of time series data are used, it is possible to calculate standard time series data with higher accuracy using only one piece of time series data as standard time series data. In addition, high-precision standard time-series data is calculated even if the sampling interval between two time-series data used to calculate the warping path is different or the number of samples is different. Can be used. When standard time series data is calculated using three or more normalized time series data, the intermediate time standard time series data calculated as described above is referred to as 'provisional' standard time series data.

  FIG. 10 is a diagram showing two normalized time series data and provisional standard time series data used in the DTW notation of FIG.

  10A shows again the normalized time series data A shown in FIG. 8A, and FIG. 10B shows again the normalized time series data B shown in FIG. 8B. Has been. The time-series data shown in FIG. 10 (AB) is provisional standard time-series data calculated from the data on each axis corresponding to each square of the warping path shown in FIG. Here, the time-series data shown in FIG. 10 (AB) is referred to as temporary standard time-series data AB.

  Here, 'time', 'value1', and 'value2' of the temporary standard time series data AB shown in FIG. 10 (AB) are attached with a code that combines the codes of the data used for the calculation of each data. is doing. For example, the data A1 and B2 of the temporary standard time series data AB are respectively “time”, “value1”, and “value2” for the data A1 of the normalized time series data A and the data B2 of the normalized time series data B, respectively. This is data consisting of the average value of. Further, the data A11 and B12 of the temporary standard time series data AB are respectively “time”, “value1”, and “value2” of the data A11 of the normalized time series data A and the data B12 of the normalized time series data B. It represents that the data consists of average values. The same applies to other codes other than those described above.

  For example, data A1 and data B1 correspond to a square in the upper left corner of the warping path shown in FIG. The average values of 'time', 'value1', and 'value2' of these data A1 and data B1 are calculated. Data consisting of these average values is data A1 and B1 constituting temporary standard time series data AB. More specifically, for “time”, “value1”, and “value2”, the values of data A1 (0.0000, 0.0000, 0.0000) and the values of data B1 (0. 0000, 0.0000, 0.0000) (0.0000, 0.0000, 0.0000), which is the average value for each variable, becomes the data A1, B1 constituting the temporary standard time series data AB. . Further, the data A4 and the data B4 correspond to the fifth square following the warping path from the upper left corner. In the fifth cell, for “time”, “value1”, and “value2”, the values of the data A4 (3.0000, 0.9375, 0.2500) and the values of the data B4 (3 (.0000, 0.8750, 0.2500) and (3.0000, 0.9063, 0.2500) are calculated as the average value for each variable. The average value for each variable (3.0000, 0.9063, 0.2500) is the data A4, B4 constituting the temporary standard time series data AB. In this way, the tentative standard time series data AB is calculated by obtaining the average value for each variable for all the squares on the warping path.

  When the temporary standard time series data is calculated as described above in step S6 shown in FIG. 5, the process proceeds to step S7. In step S7, it is determined whether or not unused normalized time series data still remains among the time series data normalized in step S2.

  If it is determined in step S7 shown in FIG. 5 that unused time-series data remains among the time-series data normalized in step S2, the process proceeds to step S8. Here, since the normalized time-series data C shown in FIG. 8C is unused, the above-described warping path calculation method using the normalized time-series data C and the temporary standard time-series data AB A new warping path is calculated by the same calculation method.

  FIG. 11 is a diagram showing DTW notation.

  In FIG. 11, the normalized time series data C shown in FIG. 8C is arranged on the vertical axis, and the temporary standard time series data AB is arranged on the horizontal axis. Since the method for calculating the distance for each square and the method for obtaining the warping path have already been described, description thereof will be omitted.

  In step S4 shown in FIG. 5, it is determined whether or not one of the time series data used for calculating the warping path immediately before step S4 is temporary standard time series data. Here, since the time series data arranged on the horizontal axis among the time series data used for calculating the warping path is provisional standard time series data, the process proceeds to step S5. In step S5, each of 'time', 'value1', and 'value2' of temporary standard time series data AB and normalized time series data C related to each square forming the warping path calculated in step S8 is described below. Calculate the weighted average as described. By arranging the calculated weighted averages in time series along the warping path, new temporary standard time series data is calculated. This step S5 also constitutes an example of the standard time series data calculation step referred to in the present invention.

  The calculation of the weighted average is performed by weighting according to how many normalized time series data the temporary standard time series which is one of the used time series data is calculated. Here, for a certain square among the squares forming the warping path, the values of “value1” of the temporary standard time series data AB and the normalized time series data C corresponding to this square are “Vab1” and “Vc1”, respectively. If there is, the weighted average value “Vabc1” of “value1” of this square is obtained by the following equation (6).

  Since the temporary standard time-series data AB is calculated from two normalized time-series data, Vab1 is multiplied by “2” for weighting. Vc1 is used as it is. Then, a weighted average value is obtained by dividing by “3”, which is the total number of normalized time-series data used in the calculation. For “value2” and “time”, a weighted average value is obtained in the same manner.

  Further, when the temporary standard time series data which is one of the used time series data is obtained based on, for example, three normalized time series data, the value used for the above multiplication is set to '2'. A new provisional standard time series data is calculated by setting the value used for division to “4” instead of “3” instead of “3”. When the normalized time series data used in this way increases, the standard time series data is calculated by increasing the values used for multiplication and division.

  Here, the weighted average is used for the calculation of the standard time series data. In the case of a simple average, the time series data used in the calculation in the later order has a greater influence on the calculation of the final standard time series data. On the other hand, if the weighted average as described above is adopted, the influence of the order can be removed.

  FIG. 12 is a diagram showing two time-series data and standard time-series data used in the DTW notation of FIG.

12 (AB) again shows the temporary standard time series data AB shown in FIG. 10 (AB), and FIG. 12 (C) shows again the normalized time series data C shown in FIG. 8 (C). Has been. The time-series data shown in FIG. 12 (ABC) is standard time-series data calculated from the data on each axis corresponding to each square of the warping path shown in FIG. Hereinafter, the time series data shown in FIG. 12 (ABC) is referred to as standard time series data ABC.
Also, FIG. 12 (ABC) shows “time”, “value1”, and “temporary standard time series data AB and normalized time series data C related to each square forming the warping path shown in FIG. The respective weighted average values of value 2 ′ are shown.

  Since the temporary standard time-series data AB is based on two normalized time-series data of the normalized time-series data A and the normalized time-series data B, the weighted average is performed as follows. For example, “4.6667” in the data A6, B6, C5 (4.6667, 0.9375, 1.0000) of the standard time series data ABC shown in FIG. 12 (ABC) is calculated as follows. That is, the temporary standard time-series data AB includes two normalized time-series data A and B in “5.000” of the data A6B6 (5.0000, 0.9063, 1.0000) of the temporary standard time-series data AB. Is multiplied by “2”, and “4.000” of the data C5 (4.0000, 1.0000, 1.0000) is added, and this added value is normalized at the time of normalization related to them. Divide by “3”, which is the total number of series data A, B, and C, to calculate “4.6667”. Further, “0.9375” of the data A6, B6, C5 (4.6667, 0.9375, 1.0000) of the standard time series data ABC is also the data A6B6 (5.000) of the temporary standard time series data AB. , 0.9063, 1.0000) is multiplied by “2”, and “1.0000” of data C5 (4.0000, 1.000, 1.0000) is added. The value is divided by “3”, which is the total number of normalized time series data A, B, and C related thereto, to calculate “0.9375”. By calculating the weighted average in this way, the standard time series data ABC is calculated.

  Here, in the standard time series data ABC shown in FIG. 12 (ABC), each of “time”, “value1”, and “value2” is similar to the way of the provisional standard time series data AB in FIG. The code | symbol which combined the code | symbol of the data used for calculation of data is attached | subjected. For example, the data A1, B2, and C1 of the standard time series data ABC are “time” for the data A1 of the normalized time series data A, the data B2 of the normalized time series data B, and the data C1 of the normalized time series data C. , 'Value1', and 'value2'. Further, the data A11, B12, C11 of the standard time series data ABC are the “time”, the data A11 of the normalized time series data A, the data B12 of the normalized time series data B, and the data C11 of the normalized time series data C, It represents that it is each average value of 'value1' and 'value2'. The same applies to other codes other than those described above.

  After calculating the standard time series data ABC as described above, the process proceeds to step S7 shown in FIG. 5, and whether or not unused normalized time series data still remains among the time series data normalized in step S2. However, since all normalized time series data is used here, the last calculated standard time series data ABC is output as final standard time series data. When four or more normalized time series data are used for calculating the standard time series data, the standard time series data ABC is returned to step S8 as temporary standard time series data, and the same processing as described above is further repeated.

  The above is the flow up to the calculation of the standard time series data in the standard time series data calculation device 410.

  According to the standard time series data calculation method described above, highly accurate standard time series data is calculated even when the sampling intervals of the time series data used for calculating the warping path are different or the number of samples is different. It can be used for highly accurate detection of abnormal conditions.

  FIG. 13 is a diagram showing standard time-series data and time-series data obtained by denormalizing the standard time-series data.

  The standard time-series data ABC is calculated based on the normalized time-series data, and each variable of the standard time-series data ABC is distributed in the range from “0” to “1”. Therefore, the standard time series data ABC is not intuitive for humans such as operators. Therefore, when presenting this standard time-series data to the operator, such as when it is displayed on the display screen, the standard time-series data ABC is subjected to a reverse conversion to the normalization performed at the time of calculation, and the normal time shown on the right side of FIG. It is preferable to generate standard time series data that has not been normalized and present the standard time series data that has not been normalized. However, standard time-series data that has been normalized is used inside the abnormality detection apparatus described below.

  In the present embodiment, the case of two variables of temperature and pressure for a certain device has been described. However, the present invention is not limited to this, and for example, the temperature, pressure, and flow rate for a certain device. There may be three or more variables as in the case of calculating standard time-series data based on the measured values.

  Next, embodiments of the abnormality detection device, abnormality detection method, and abnormality detection program of the present invention will be described with reference to the drawings.

  A computer 100 shown in FIG. 1 constitutes an embodiment of an abnormality detection apparatus of the present invention by installing and executing an abnormality detection program (see FIG. 14) described later. Except for the difference in the program executed in the computer 100, the configuration of the computer 100 is the same as that described in the embodiment of the standard time-series data calculation apparatus, and therefore the description of the computer 100 itself is omitted. In this embodiment, the CD drive 117 shown in FIG. 2 is loaded with a CD-ROM 320 (see FIG. 14) in which an abnormality detection program is stored, and the loaded CD-ROM 320 is accessed. The abnormality detection program stored in the CD-ROM 320 is uploaded to the computer 100 and stored in the hard disk device 115. When the abnormality detection program is activated and executed, the computer 100 operates as an example of the abnormality detection device according to the present invention.

  Next, the abnormality detection program will be described.

  FIG. 14 is a conceptual diagram showing a CD-ROM in which an abnormality detection program is stored.

  As shown in FIG. 14, the abnormality detection program 321 stored in the CD-ROM 320 includes a data acquisition unit 312, a data normalization unit 313, a warping path calculation unit 314, a standard time series data calculation unit 315, an operation control unit 316, A distance calculation unit 327 and an abnormality detection unit 328 are included.

  Details of each part of the abnormality detection program will be described together with the operation of each part of the abnormality detection apparatus 411 described later.

  FIG. 15 is a functional block diagram of the abnormality detection device 411.

  The block diagram shown in FIG. 15 represents functions when the abnormality detection program 321 shown in FIG. 14 is installed in the computer 100 of FIG. 1 and operates as the abnormality detection device 411 according to an embodiment of the present invention. .

  15 includes a data acquisition unit 500, a data normalization unit 501, a warping path calculation unit 502, a standard time series data calculation unit 503, an operation control unit 504, a total distance calculation unit 515, an abnormality detection unit 516, And a display unit 517. The data acquisition unit 500, the data normalization unit 501, the warping path calculation unit 502, the standard time series data calculation unit 503, the operation control unit 504, the total distance calculation unit 515, and the abnormality detection unit 516 are the CPU 113 shown in FIG. Is mainly borne by. The display unit 517 is mainly carried by the image display device 120 shown in FIG. On the other hand, the storage device 505 is connected to the data acquisition unit 500 and the standard time series data calculation unit 503 in the standard time series data calculation device 410, and is mainly handled by the hard disk device 115 shown in FIG. . Further, the temperature sensor 420 and the pressure sensor 421 are connected to the abnormality detection device 411 and directly connected to the input interface 118 shown in FIG. The temperature sensor 420, pressure sensor 421, data acquisition unit 500, data normalization unit 501, warping path calculation unit 502, standard time series data calculation unit 503, operation control unit 504, and storage device 505 are shown in FIG. Since it is the same as the structure of each part of the same name, the same code | symbol is attached | subjected and description is abbreviate | omitted. In the present embodiment, an example will be described in which an abnormality is detected from the measured values of temperature and pressure for the same device as that described in the previous embodiment of the standard time-series data calculation device. Hereinafter, only the total distance calculation unit 515, the abnormality detection unit 516, and the display unit 517 will be described.

  The total distance calculation unit 515 is a target of abnormality diagnosis in which the standard time series data calculated by the standard time series data calculation unit 503 is arranged on one axis, acquired by the data acquisition unit 500, and normalized by the data normalization unit 501. A warping path is calculated in DTW notation in which time series data is arranged on the other axis, and a value obtained by summing the distances of the cells forming the warping path is calculated as a total distance.

  The abnormality detection unit 516 determines whether or not the total distance calculated by the total distance calculation unit 515 is greater than a predetermined threshold, and if this total distance is greater than the threshold, an abnormality has occurred in the target being measured. The display unit 517 performs a warning display indicating that.

  FIG. 16 is a flowchart showing a flow of abnormality detection in the abnormality detection device 411. In the following description, it is assumed that the standard time series data calculated by the standard time series data calculation device 410 in the previous embodiment is stored in the storage device 505 shown in FIG.

  In step S <b> 11 shown in FIG. 16, time series data to be subjected to abnormality diagnosis is acquired by the data acquisition unit 500.

  In step S12 shown in FIG. 16, the time series data acquired in step S11 is normalized. In the above-described calculation of the standard time series data, normalization is performed based on the minimum value and the maximum value of the time series data used for calculating the standard time series data. However, the normalization here does not use the minimum and maximum values of the acquired time series data of the abnormality diagnosis target, but performs conversion using the same values as when the standard time series data was calculated. For example, when calculating the standard time series data described above, the value of “value1” is divided by “3.2000”, the value of “value2” is divided by “2.0000”, and the minimum value of each variable is maximized. Is normalized to a range from “0” to “1”. Therefore, here, this numerical value is applied as it is to the time-series data that is the object of abnormality diagnosis, and normalization is performed. FIG. 17 shows an example of time-series data for abnormality diagnosis, which will be described later. For example, “3.0000” of “value1” in the data D5 of FIG. 17 is converted into “0.9375” by normalization. Then, “2.000” of “value2” is converted to “1.000” by normalization.

  In step S13 shown in FIG. 16, the warping path is calculated in the DTW notation in which the time series data of the abnormality diagnosis target normalized in step S12 is arranged on one axis and the standard time series data is arranged on the other axis. Since the DTW notation method and the warping path calculation method are the same as the method described in the embodiment of the standard time-series data calculation device, illustration is omitted.

  In step S14 shown in FIG. 16, the sum of the distances of the squares of the warping path calculated in step S13 is the total distance between the standard time series data and the normalized time series data of the abnormality diagnosis target. Steps S13 and S14 correspond to an example of a total distance calculating step according to the present invention.

  In step S15 shown in FIG. 16, it is determined whether or not the calculated total distance exceeds a threshold value. If the calculated total distance exceeds the threshold value, the process proceeds to step S16, and a warning display indicating that an abnormality has been detected is displayed on the display unit 517. Steps S15 and S16 correspond to an example of an abnormality detection step according to the present invention.

  FIG. 17 is a diagram showing time-series data D that is an abnormality diagnosis target acquired in step S11. Here, the number of samples and the sampling interval of the time series data D and the standard time series data ABC shown in FIG. 12 will be described. The number of samples of the time series data D is 12, and the number of samples of the standard time series data ABC is 14, which are different from each other. The sampling intervals of the time series data D are all “1.000”, but the sampling intervals of the standard time series data ABC are, for example, sampling intervals from the data A1, B1, C1 to the data A1, B2, C1 are “0.3333”. Is different from the sampling interval of the time-series data D.

  As described in the standard time-series data calculation, a warping path can be calculated even with time-series data having different sampling intervals and different numbers of samples as described above. Therefore, each square on the warping path can be calculated. The total distance obtained by summing the distances can be obtained, and the presence or absence of abnormality can be determined.

  FIG. 18 is a graph showing changes in each variable according to time of three time series data in a normal state used for calculation of standard time series data and time series data which is an object of abnormality diagnosis.

  In the graph of FIG. 18, in the three time-series data of (A) to (C) in the normal state, “value 2” indicating the pressure changes behind the change of “value 1” indicating the temperature. On the other hand, in the time series data (D) for abnormality diagnosis, the order of increase and decrease in temperature and pressure is reversed.

  That is, in the normal states (A) to (C), the pressure increases after the temperature rises, and the pressure decreases after the temperature drops, whereas in the abnormal state (D), the opposite is true. After the pressure increases, the temperature rises, and after the pressure decreases, the temperature drops.

  FIG. 19 is a graph in which the total distances of the time series data A, the time series data B, the time series data C, and the time series data D with respect to the standard time series data are calculated by the total distance calculation unit of the abnormality detection device and compared. It is. From this graph, the total distance of the time series data D is prominent compared to the others, and it is clearly shown that the time series data D is abnormal data.

  As described above, according to this abnormality detection device 411, even if the sampling interval between the standard time series data and the time series data of the abnormality diagnosis target is different or the number of samples is different, the apparatus of the abnormality diagnosis target is highly accurate. An abnormal state can be detected.

  In the present embodiment, the case of two variables of temperature and pressure for a certain device has been described. However, the present invention is not limited to this, and for example, measurement of temperature, pressure, and flow rate for a certain device. There may be three or more variables as in the case of detecting an abnormality based on the value.

100 Computer 110 Main Body 111 FD Loading Port 112 CD-ROM Loading Port 113 CPU
114 Main Memory 115 Hard Disk Device 116 FD Drive 117 CD Drive 118 Input Interface 119 Output Interface 120 Display 121 Display Screen 130 Keyboard 140 Mouse 140 Bus 150 300 FD
310 CD-ROM
320 CD-ROM
311 Standard Time Series Data Calculation Program 312 Data Acquisition Unit 313 Data Normalization Unit 314 Warping Path Calculation Unit 315 Standard Time Series Data Calculation Unit 316 Operation Control Unit 321 Abnormality Detection Program 327 Total Distance Calculation Unit 328 Abnormality Detection Unit 410 Standard Time Series Data Calculation Device 411 Abnormality detection device 420 Temperature sensor 421 Pressure sensor 500 Data acquisition unit 501 Data normalization unit 502 Warping path calculation unit 503 Standard time series data calculation unit 504 Operation control unit 505 Storage device 515 Total distance calculation unit 516 Abnormality detection unit 517 Display unit

Claims (12)

Time series data composed of a set of time series of a plurality of data obtained by sampling a plurality of variables defining one system at the same sampling time, and sequentially differing for the plurality of variables of the one system A data acquisition step for acquiring a plurality of time-series data collected in a sampling time period;
By multiplying a plurality of data for each variable across the plurality of time series data acquired in the data acquisition step by a coefficient for each variable and performing scale conversion, the plurality of data for each variable is not dependent on the variable. A data normalization step for generating a plurality of normalized time-series data obtained by normalizing the plurality of time-series data by performing normalization distributed within the same numerical range,
The first normalized time series data of the plurality of normalized time series data generated in the data normalization step is arranged on a single axis, and the second normalized time series of the plurality of normalized time series data In DTW notation in which data is arranged on the other axis, a plurality of variables corresponding to the plurality of variables on one axis corresponding to each square divided at each sampling time of two normalized time series data arranged on the one axis and the other axis A total distance for each of the plurality of variables between the data and a plurality of data corresponding to the plurality of variables on the other axis corresponding to each of the cells is calculated for each cell, and based on the distance A warping path calculating step for obtaining a warping path by selecting a mass;
By calculating an average value for each variable between a plurality of data on one axis corresponding to each square on the warping path and a plurality of data on the other axis, each square on the warping path is calculated. A standard time-series data calculation method, comprising: a standard time-series data calculation step for calculating standard time-series data including a time-series set of a plurality of average values.   The standard time series data calculated in the standard time series data calculation step is uniaxially arranged as a new first normalized time series data replacing the first normalized time series data, and is normalized in the data normalization step. In the DTW notation in which the third normalized time series data among the plurality of normalized time series data is arranged on the other axis as new second normalized time series data replacing the second normalized time series data. In addition, the new normalized first time series data is sequentially used, and the new first time series data arranged on one axis is weighted according to the number of past iterations when the standard time series data calculation step is executed. The final standard time series data is calculated by repeating the warping path calculation step and the standard time series data calculation step while performing averaging. Time series data calculation method according to claim 1, wherein.   The standard time-series data calculating step includes calculating an average value of sampling times and a variable for each variable between a plurality of data on one axis corresponding to each square on the warping path and a plurality of data on the other axis. 2. The step of calculating standard time-series data comprising time-series sets of average values for each square on the warping path by calculating average values. Or the standard time series data calculation method of 2 description. An abnormality detection method for detecting an abnormality in the system using the standard time series data calculated by the standard time series data calculation method according to any one of claims 1 to 3, wherein the plurality of the one system Anomaly detection object data acquisition step for acquiring anomaly detection object time series data consisting of a set of time series of a plurality of data obtained by sampling the variables of
By multiplying a plurality of data for each variable constituting the abnormality detection target time series data by a coefficient for each variable adopted in the data normalization step, the abnormality detection target time series data is normalized at the time of abnormality detection target. Anomaly detection target data normalization step for converting to series data;
In the DTW notation in which the standard time series data is arranged on one axis and the abnormality detection target time series data normalized in the abnormality detection target data normalizing step is arranged on the other axis, two time series data arranged on one axis and the other axis Between a plurality of data corresponding to the plurality of variables on one axis corresponding to each square divided at each sampling time and a plurality of data corresponding to the plurality of variables on the other axis corresponding to each square A total distance for calculating a total distance between the plurality of variables for each square and calculating a total distance between the standard time series data and the normalized target time series data based on the distance A calculation step;
An abnormality detection method comprising: detecting an abnormality in the system when the total distance calculated in the total distance calculation step is larger than a predetermined threshold. Time series data composed of a set of time series of a plurality of data obtained by sampling a plurality of variables defining one system at the same sampling time, and sequentially differing for the plurality of variables of the one system A data acquisition unit for acquiring a plurality of time-series data collected in a sampling time period;
Scale conversion is performed by multiplying a plurality of data for each variable across a plurality of time series data acquired by the data acquisition unit by a coefficient for each variable, and the plurality of data for each variable is not dependent on the variable. A data normalization unit that generates a plurality of normalized time-series data obtained by normalizing the plurality of time-series data by performing normalization distributed within the same numerical range,
The first normalized time series data of the plurality of normalized time series data generated by the data normalization unit is arranged on one axis, and the second normalized time series of the plurality of normalized time series data In DTW notation in which data is arranged on the other axis, a plurality of variables corresponding to the plurality of variables on one axis corresponding to each square divided at each sampling time of two normalized time series data arranged on the one axis and the other axis A total distance for each of the plurality of variables between the data and a plurality of data corresponding to the plurality of variables on the other axis corresponding to each of the cells is calculated for each cell, and based on the distance A warping path calculation unit for obtaining a warping path by selecting a square;
By calculating an average value for each variable between a plurality of data on one axis corresponding to each square on the warping path and a plurality of data on the other axis, each square on the warping path is calculated. A standard time-series data calculation device comprising: a standard time-series data calculation unit that calculates standard time-series data including a time-series set of a plurality of average values.   The standard time series data calculated by the standard time series data calculation unit is uniaxially arranged as new first normalized time series data to replace the first normalized time series data, and is normalized by the data normalization unit. In the DTW notation in which the third normalized time series data among the plurality of normalized time series data is arranged on the other axis as new second normalized time series data replacing the second normalized time series data. Then, averaging is performed by sequentially using new third normalized time series data, and adding a weight corresponding to the number of past iterations to the new first time series data arranged on one axis in the standard time series data calculation unit. An operation control unit that calculates final standard time series data by alternately and repeatedly operating the warping path calculation unit and the standard time series data calculation unit. Time series data calculation apparatus according to claim 5, symptoms.   The standard time-series data calculation unit calculates an average value of sampling times and a variable for each variable between a plurality of data on one axis corresponding to each square on the warping path and a plurality of data on the other axis. 6. Standard time-series data comprising time-series sets of average values for each square on the warping path is calculated by calculating average values. Or the standard time series data calculation apparatus of 6. A standard time-series data calculation device according to any one of claims 5 to 7,
The data acquisition unit further includes an abnormality detection composed of a set of time series of a plurality of data obtained by sampling the plurality of variables of the one system at the same sampling time in the sampling time zone of the abnormality detection target. To obtain target time series data,
The data normalization unit further multiplies a plurality of data for each variable constituting the abnormality target time series data by a coefficient for each variable adopted in the data normalization unit, thereby the abnormality detection target time series data. To normalization abnormality detection target time series data,
In the DTW notation in which the standard time series data is arranged on one axis and the abnormality detection target time series data normalized in the abnormality detection target data normalizing step is arranged on the other axis, two time series data arranged on one axis and the other axis Between a plurality of data corresponding to the plurality of variables on one axis corresponding to each square divided at each sampling time and a plurality of data corresponding to the plurality of variables on the other axis corresponding to each square A total distance for calculating a total distance between the plurality of variables for each square and calculating a total distance between the standard time series data and the normalized target time series data based on the distance A calculation unit;
An abnormality detection unit comprising: an abnormality detection unit that detects an abnormality when the total distance calculated by the total distance calculation unit is greater than a predetermined threshold. It is executed in an arithmetic processing unit that executes a program, and the arithmetic processing unit is
Time series data composed of a set of time series of a plurality of data obtained by sampling a plurality of variables defining one system at the same sampling time, and sequentially differing for the plurality of variables of the one system A data acquisition unit for acquiring a plurality of time-series data collected in a sampling time period;
Scale conversion is performed by multiplying a plurality of data for each variable across a plurality of time series data acquired by the data acquisition unit by a coefficient for each variable, and the plurality of data for each variable is not dependent on the variable. A data normalization unit that generates a plurality of normalized time-series data obtained by normalizing the plurality of time-series data by performing normalization distributed within the same numerical range,
The first normalized time series data of the plurality of normalized time series data generated by the data normalization unit is arranged on one axis, and the second normalized time series of the plurality of normalized time series data In DTW notation in which data is arranged on the other axis, a plurality of variables corresponding to the plurality of variables on one axis corresponding to each square divided at each sampling time of two normalized time series data arranged on the one axis and the other axis A total distance for each of the plurality of variables between the data and a plurality of data corresponding to the plurality of variables on the other axis corresponding to each of the cells is calculated for each cell, and based on the distance A warping path calculation unit for obtaining a warping path by selecting a square;
By calculating an average value for each variable between a plurality of data on one axis corresponding to each square on the warping path and a plurality of data on the other axis, each square on the warping path is calculated. A standard time-series data calculation program that operates as a standard time-series data calculation device having a standard time-series data calculation unit that calculates standard time-series data including a time-series set of a plurality of average values.   The standard time series data calculated by the standard time series data calculation unit is uniaxially arranged as new first normalized time series data to replace the first normalized time series data, and is normalized by the data normalization unit. In the DTW notation in which the third normalized time series data among the plurality of normalized time series data is arranged on the other axis as new second normalized time series data replacing the second normalized time series data. Then, averaging is performed by sequentially using new third normalized time series data, and adding a weight corresponding to the number of past iterations to the new first time series data arranged on one axis in the standard time series data calculation unit. An operation control unit that calculates final standard time series data by alternately and repeatedly operating the warping path calculation unit and the standard time series data calculation unit. Claim 9, wherein the standard time series data calculation program to symptoms.   The standard time-series data calculation unit calculates an average value of sampling times and a variable for each variable between a plurality of data on one axis corresponding to each square on the warping path and a plurality of data on the other axis. 10. The standard time-series data consisting of a time-series set of average values for each square on the warping path is calculated by calculating each average value. Or the standard time series data calculation program of 10 description. An abnormality detection program including the standard time-series data calculation program according to any one of claims 9 to 11 as a program part, wherein the arithmetic processing unit is
The data acquisition unit further includes an abnormality detection composed of a set of time series of a plurality of data obtained by sampling the plurality of variables of the one system at the same sampling time in the sampling time zone of the abnormality detection target. To obtain target time series data,
The data normalization unit further multiplies a plurality of data for each variable constituting the abnormality target time series data by a coefficient for each variable adopted in the data normalization unit, thereby the abnormality detection target time series data. To normalization abnormality detection target time series data,
In the DTW notation in which the standard time series data is arranged on one axis and the abnormality detection target time series data normalized in the abnormality detection target data normalizing step is arranged on the other axis, two time series data arranged on one axis and the other axis Between a plurality of data corresponding to the plurality of variables on one axis corresponding to each square divided at each sampling time and a plurality of data corresponding to the plurality of variables on the other axis corresponding to each square A total distance for calculating a total distance between the plurality of variables for each square and calculating a total distance between the standard time series data and the normalized target time series data based on the distance A calculation unit;
An abnormality detection program that operates as an abnormality detection device having an abnormality detection unit that detects an abnormality when the total distance calculated by the total distance calculation unit is greater than a predetermined threshold.

Images