Training data generation device, method, and program

Abstract

To appropriately generate training data.SOLUTION: A training data generation device according to an embodiment is provided with an acquisition unit, a division unit, a detection unit, and a generation unit. The acquisition unit acquires operation data concerning an operating state of a device over a predetermined period. The division unit divides the operation data into at least provisional training data and provisional test data. The detection unit detects abnormal values from the provisional test data based on the provisional training data. The generation unit generates training data from the operation data by excluding abnormal periods in which the abnormal values were detected from the predetermined period.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] FIELD Embodiments of the present invention relate to a training data generation device, method, and program. [Background technology]

[0002] Industrial systems (e.g., infrastructure facilities, manufacturing facilities) consist of multiple pieces of equipment, and each piece of equipment controls its own process. The shutdown of an industrial system can result in social and economic losses. To maintain the health of an industrial system, it is necessary to monitor the operating status of the industrial system.

[0003] For example, the operating status of an industrial system is monitored by a machine learning model. The machine learning model uses operating data (normal data) from a period when the equipment is in a normal operating state (normal period) as training data. When trained with this training data, the machine learning model can detect abnormal values ​​from any operating data of the equipment.

[0004] Conversely, a machine learning model may use operating data (abnormal data) from a period (abnormal period) in which the equipment is in an abnormal operating state as training data. When trained with this training data, the machine learning model has difficulty detecting abnormal values ​​from any operating data of the equipment. Therefore, it is required to appropriately generate training data so that the training data includes normal data but does not include abnormal data. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Japanese Patent Application Laid-Open No. 2013-218725 [Patent Document 2] Japanese Patent Publication No. 2022-074890 Summary of the Invention [Problem to be solved by the invention]

[0006] The problem that the present invention aims to solve is to generate training data appropriately. [Means for solving the problem]

[0007] A training data generation device according to an embodiment includes an acquisition unit, a division unit, a detection unit, and a generation unit. The acquisition unit acquires operating data related to the operating state of an apparatus over a predetermined period. The division unit divides the operating data into at least provisional training data and provisional test data. The detection unit detects abnormal values ​​from the provisional test data based on the provisional training data. The generation unit generates training data from the operating data by excluding, from the predetermined period, an abnormal period in which the abnormal value was detected. [Brief explanation of the drawings]

[0008] [Figure 1] FIG. 1 is a functional configuration diagram of a training data generation device according to a first embodiment. [Figure 2] FIG. 2 is an operational flow diagram of the training data generation device according to the first embodiment. [Figure 3] FIG. 3 is an explanatory diagram showing a training data generation process according to the first embodiment. [Figure 4] FIG. 10 is an explanatory diagram showing a process of dividing driving data according to a first modified example of the first embodiment. [Figure 5] FIG. 10 is an explanatory diagram showing a process of dividing driving data according to a second modification of the first embodiment. [Figure 6] FIG. 10 is an explanatory diagram showing a process of dividing driving data according to a third modified example of the first embodiment. [Figure 7] FIG. 10 is an explanatory diagram showing an abnormal value detection process according to a third modification of the first embodiment. [Figure 8] FIG. 10 is an explanatory diagram showing a process of dividing driving data according to a fourth modified example of the first embodiment. [Figure 9] FIG. 11 is a first explanatory diagram showing the abnormal value detection process according to the fourth modification of the first embodiment. [Figure 10] FIG. 20 is a second explanatory diagram showing the abnormal value detection process according to the fourth modification of the first embodiment. [Figure 11] FIG. 10 is a functional configuration diagram of a training data generation device according to a second embodiment. [Figure 12] FIG. 10 is an operational flow diagram of the training data generation device according to the second embodiment. [Figure 13] FIG. 10 is a frequency distribution diagram of provisional training data according to Modification 2 of the second embodiment. [Figure 14] FIG. 10 is a frequency distribution diagram according to Modification 3 of the second embodiment. [Figure 15] FIG. 10 is an explanatory diagram showing an abnormal value detection process according to a fourth modification of the second embodiment. [Figure 16] FIG. 11 is a functional configuration diagram of a training data generation device according to a third embodiment. [Figure 17] FIG. 11 is an operational flow diagram of the training data generation device according to the third embodiment. [Figure 18] FIG. 11 is an explanatory diagram showing an abnormal value detection process according to the third embodiment. [Figure 19] FIG. 11 is a frequency distribution diagram of operating data and reference data according to the third embodiment. [Figure 20] FIG. 2 is a hardware configuration diagram of a training data generation device according to each embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0009] Hereinafter, each embodiment will be described with reference to the drawings. In each embodiment, parts with the same reference numerals perform similar operations, and redundant description will be omitted as appropriate.

[0010] (First embodiment) 1 is a functional configuration diagram of a training data generation device 1 according to a first embodiment. The training data generation device 1 is a device that generates training data XD from driving data OD. The training data generation device 1 includes an acquisition unit 11, a division unit 12, a detection unit 13, and a generation unit 14.

[0011] The acquisition unit 11 acquires various types of data. The acquisition unit 11 acquires operating data OD related to the operating state of equipment over a predetermined period of time. The acquisition unit 11 acquires the operating data OD from outside the training data generation device 1. The acquisition unit 11 transmits the acquired operating data OD to the division unit 12 and the generation unit 14. The operating data OD is also referred to as "candidate data."

[0012] The dividing unit 12 divides various types of data. The dividing unit 12 divides the driving data OD into at least temporary training data RD and temporary test data SD. The dividing unit 12 transmits the divided temporary training data RD and temporary test data SD to the detection unit 13.

[0013] The detection unit 13 detects various types of data. The detection unit 13 detects abnormal values ​​AV from the provisional test data SD based on the provisional training data RD. The detection unit 13 transmits the detected abnormal values ​​AV to the generation unit 14. The abnormal values ​​AV are also referred to as "candidates for abnormality." The abnormal values ​​AV may include signs of abnormality.

[0014] The generation unit 14 generates various types of data. The generation unit 14 generates training data XD from the driving data OD by excluding a period in which an abnormal value AV is detected (an abnormal period) from a predetermined period of the driving data OD. The generation unit 14 outputs the generated training data XD to the outside of the training data generation device 1.

[0015] 2 is an operational flow diagram of the training data generation device 1 according to the first embodiment. The training data generation device 1 sequentially executes the processes of steps S11 to S14.

[0016] (Step S11) First, the training data generation device 1 acquires operation data OD through the acquisition unit 11. The operation data OD is time-series data relating to the operation state of a specific device in an industrial system. The operation data OD may be time-series data relating to each of a plurality of process variables indicating the operation state of the specific device. Each process variable may be a measurement value (e.g., temperature, pressure, speed, acceleration, current, voltage) from each sensor installed in the specific device.

[0017] If a given device is subject to periodic fluctuations, the period of the operating data OD for the given device may be set to one period or more. For example, if a given device is subject to seasonal fluctuations, the period of the operating data OD may be set to one year or more. The operating data OD may include values ​​sampled from each sensor at any time interval (e.g., one minute).

[0018] (Step S12) Next, the training data generation device 1 divides the driving data OD through the dividing unit 12. The dividing unit 12 divides the driving data OD acquired in step S11 into temporary training data RD and temporary test data SD. The period of the temporary training data RD does not overlap with the period of the temporary test data SD. The period of the temporary training data RD may be longer than the period of the temporary test data SD.

[0019] In particular, if a specific device is affected by seasonal variations, the period of the provisional test data SD for that specific device may be set to be less than the length of each season (e.g., less than three months in Japan). If the period of the provisional test data SD is set to be longer than the length of each season, seasonal variations in air temperature, water temperature, etc. may be erroneously detected as an abnormal value AV.

[0020] (Step S13) Subsequently, the training data generation device 1 detects an abnormal value AV through the detection unit 13. The detection unit 13 uses the temporary training data RD and temporary test data SD divided in step S12. The detection unit 13 assumes that the temporary training data RD is data in a normal state (normal data), and detects an abnormal value AV from the temporary test data SD.

[0021] The detection unit 13 may determine whether the value of the provisional test data SD is included in a range between the maximum and minimum values ​​of the provisional training data RD. The detection unit 13 detects a value of the provisional test data SD that is not included in the range as an abnormal value AV. The detection unit 13 may also detect a period (abnormal period) during which an abnormal value AV is detected in the provisional test data SD.

[0022] (Step S14) Finally, the training data generation device 1 generates training data XD through the generation unit 14. The generation unit 14 uses the abnormal value AV (or abnormal period) detected in step S13. The generation unit 14 generates training data XD from the driving data OD by excluding the period (abnormal period) in which the abnormal value AV was detected from the driving data OD.

[0023] FIG. 3 is an explanatory diagram showing the process of generating training data XD according to the first embodiment. FIG. 3 sequentially shows each process related to steps S11 to S14 in FIG. 2. The period of each data in each process is indicated by bands B1A to B4A. For ease of explanation, the period is expressed as [start time, end time] or (start time, end time). In a closed interval represented by [A, B], the start time and end time are included in the interval. In an open interval represented by (A, B), the start time or end time is not included in the interval.

[0024] As shown in band B1A in the first stage, the acquisition unit 11 acquires the driving data OD (see step S11). The period of the driving data OD is [t1, t4].

[0025] As shown in band B2A in the second row, the dividing unit 12 divides the driving data OD into two pieces of temporary training data RD1 and RD2 and one piece of temporary test data SD (see step S12). The period of the temporary training data RD1 is [t1, t2], and the period of the temporary training data RD2 is [t3, t4]. The period of the temporary test data SD is (t2, t3). There is no blank period G between the temporary training data RD1 or RD2 and the temporary test data SD.

[0026] As shown in band B3A in the third row, the detection unit 13 detects an abnormal value AV from the provisional test data SD based on the provisional training data RD1 and RD2 (see step S13). The abnormal period during which the abnormal value AV is detected is [tP1, tP2].

[0027] As shown in the fourth band B4A, the generation unit 14 generates two pieces of training data XD1 and XD2 by excluding the abnormal period [tP1, tP2] in which the abnormal value AV was detected from the period [t1, t4] of the operating data OD (see step S14). The period of the training data XD1 is [t1, tP1], and the period of the training data XD2 is (tP2, t4]).

[0028] According to the first embodiment, the training data generation device 1 detects an abnormal value AV from a period of the provisional test data SD. The training data generation device 1 generates training data XD by excluding an abnormal period in which an abnormal value AV is detected from a period of the operating data OD. Therefore, the training data generation device 1 can appropriately generate training data XD so that the training data XD includes normal data but does not include abnormal data. When trained with the training data XD, the machine learning model can detect an abnormal value AV from any operating data OD of the equipment. In other words, the training data generation device 1 can support anomaly detection by the machine learning model.

[0029] (Modification 1 of the first embodiment) Fig. 4 is an explanatory diagram showing the division process of the driving data OD according to the first modification of the first embodiment. Fig. 4 shows the steps S11 and S12 of Fig. 2 in order. The period of each data in each process is indicated by bands B1B and B2B.

[0030] As shown in band B1B in the first stage, the acquisition unit 11 acquires the driving data OD (see step S11). The period of the driving data OD is [t1, t6].

[0031] As shown in band B2B in the second row, the division unit 12 divides the driving data OD into two pieces of temporary training data RD1 and RD2 and one piece of temporary test data SD (see step S12). The period of the temporary training data RD1 is [t1, t2], and the period of the temporary training data RD2 is [t5, t6]. The period of the temporary test data SD is [t3, t4].

[0032] Between the temporary training data RD1 and the temporary test data SD there is a blank period G1 (t2, t3). Between the temporary training data RD2 and the temporary test data SD there is a blank period G2 (t4, t5). The two blank periods G1 and G2 do not belong to the period of the two temporary training data RD1 and RD2, nor to the period of the temporary test data SD.

[0033] Assume that an abnormal value AV exists in at least one of the two blank periods G1 and G2 and in the provisional test data SD. Furthermore, assume that no abnormal value AV exists in the provisional training data RD1 and RD2. In this case, the detection unit 13 can detect an abnormal value AV from the provisional test data SD based on the provisional training data RD1 and RD2, which do not include an abnormal value AV. On the other hand, assume that there is no blank period G between the provisional training data RD and the provisional test data SD, and that the blank period G1 is included in the provisional training data RD1 and the blank period G2 is included in the provisional training data RD2. In this case, since the abnormal value AV is included in the provisional training data RD, no abnormal value AV is detected from the provisional test data SD.

[0034] According to the first modification of the first embodiment, the training data generation device 1 divides the driving data OD so that there is a blank period G between the provisional training data RD and the provisional test data SD. Therefore, the training data generation device 1 can detect the abnormal value AV from the provisional test data SD even if the abnormal value AV is included in the blank period G adjacent to the period of the provisional test data SD.

[0035] (Modification 2 of the first embodiment) Fig. 5 is an explanatory diagram showing the division process of the driving data OD according to the second modification of the first embodiment. Fig. 5 shows the steps S11 and S12 of Fig. 2 in order. The period of each data in each process is indicated by bands B1C and B2C.

[0036] As shown in band B1C in the first stage, the acquisition unit 11 acquires the driving data OD (see step S11). The period of the driving data OD is [t1, t10].

[0037] As shown in band B2C in the second row, the dividing unit 12 divides the driving data OD into three pieces of temporary training data RD1 to RD3 and two pieces of temporary test data SD1 and SD2 (see step S12). The period of the temporary training data RD1 is [t1, t2], the period of the temporary training data RD2 is [t5, t6], and the period of the temporary training data RD3 is [t9, t10]. The period of the temporary test data SD1 is [t3, t4], and the period of the temporary test data SD2 is [t7, t8].

[0038] A blank period G1 (t2, t3) exists between the temporary training data RD1 and the temporary test data SD1. A blank period G2 (t4, t5) exists between the temporary training data RD2 and the temporary test data SD1. A blank period G3 (t6, t7) exists between the temporary training data RD2 and the temporary test data SD2. A blank period G4 (t8, t9) exists between the temporary training data RD3 and the temporary test data SD2. The four blank periods G1 to G4 do not belong to the period of the three temporary training data RD1 to RD3, nor to the period of the two temporary test data SD1 and SD2.

[0039] According to the second modification of the first embodiment, the training data generation device 1 divides the driving data OD into a plurality of provisional test data SD. The training data generation device 1 detects an abnormal value AV from each of the plurality of provisional test data SD based on the plurality of provisional training data RD. Therefore, the second modification can achieve the same effects as the first embodiment and the first modification.

[0040] (Modification 3 of the first embodiment) Fig. 6 is an explanatory diagram showing the division process of the driving data OD according to the third modification of the first embodiment. Fig. 6 shows the steps S11 and S12 of Fig. 2 in order. The period of each data in each process is indicated by a band B1D and six patterns P1 to P6.

[0041] As shown in band B1D in the first stage, the acquisition unit 11 acquires the driving data OD (see step S11). The period of the driving data OD is [t1, t7].

[0042] As shown in the patterns P1 to P6 in the second to seventh rows, the dividing unit 12 divides the driving data OD into six patterns P1 to P6 corresponding to combinations of temporary training data RD and temporary test data SD (see step S12). The dividing unit 12 may divide the driving data OD so that the durations of the six pieces of temporary test data SD1 to SD6 differ among the six patterns P1 to P6. The dividing unit 12 may divide the period [t1, t7] of the driving data OD into the durations of the six pieces of temporary test data SD1 to SD6 in the six patterns P1 to P6.

[0043] In pattern P1, the period of the temporary test data SD1 is [t1, t2], and the period of the temporary training data RD1 is (t2, t7]. In pattern P2, the period of the temporary test data SD2 is (t2, t3], the period of the temporary training data RD21 is [t1, t2], and the period of the temporary training data RD22 is (t3, t7]. In pattern P3, the period of the temporary test data SD3 is (t3, t4], the period of the temporary training data RD31 is [t1, t3], and the period of the temporary training data RD32 is (t4, t7].

[0044] In pattern P4, the period of the temporary test data SD4 is (t4, t5], the period of the temporary training data RD41 is [t1, t4], and the period of the temporary training data RD42 is (t5, t7]. In pattern P5, the period of the temporary test data SD5 is (t5, t6], the period of the temporary training data RD51 is [t1, t5], and the period of the temporary training data RD52 is (t6, t7]. In pattern P6, the period of the temporary test data SD6 is (t6, t7], and the period of the temporary training data RD6 is [t1, t6]. The periods of two adjacent temporary test data SD are consecutive at the boundary times (start time and end time).

[0045] The detection unit 13 detects an abnormal value AV from the temporary test data SD based on the temporary training data RD in each of the six patterns P1 to P6 (see step S13). In pattern P1, the detection unit 13 detects an abnormal value AV from the temporary test data SD1 based on the temporary training data RD1. In pattern P2, the detection unit 13 detects an abnormal value AV from the temporary test data SD2 based on the temporary training data RD21 and RD22. In pattern P3, the detection unit 13 detects an abnormal value AV from the temporary test data SD3 based on the temporary training data RD31 and RD32.

[0046] In pattern P4, the detection unit 13 detects an abnormal value AV from the temporary test data SD4 based on the temporary training data RD41 and RD42. In pattern P5, the detection unit 13 detects an abnormal value AV from the temporary test data SD5 based on the temporary training data RD51 and RD52. In pattern P6, the detection unit 13 detects an abnormal value AV from the temporary test data SD6 based on the temporary training data RD6. In this way, the detection unit 13 can detect an abnormal value AV from the entire period [t1, t7] of the driving data OD.

[0047] The periods of the six provisional test data SD1 to SD6 do not overlap with one another. In another example, the periods of the provisional test data SD may overlap with one another. If overlapping is permitted, the total period of the provisional test data SD may correspond to the entire period of the driving data OD. In this case, the detection unit 13 can detect the abnormal value AV from the entire period of the driving data OD.

[0048] The generating unit 14 generates training data XD by excluding abnormal periods in which abnormal values ​​AV are detected in the six patterns P1 to P6 from the period [t1, t7] of the operating data OD (see step S14).

[0049] FIG. 7 is an explanatory diagram showing the detection process of an abnormal value AV according to Modification 3 of the first embodiment. FIG. 7 shows a time series graph 200 related to the operating data OD of FIG. 6. The horizontal axis of the time series graph 200 represents time. The vertical axis of the time series graph 200 represents the value of the process variable included in the operating data OD. Time series data 210 shows the change over time in the value of the process variable using solid and dotted lines. Portions 211 and 212 of the time series data 210 during the abnormal period in which the abnormal value AV was detected are highlighted by dotted lines.

[0050] The period [t1, t7] of the time series data 210 is divided into six periods of provisional test data SD1 to SD6. The upper dashed line 220 and the lower dashed line 230 in the period [t1, t2] of provisional test data SD1 correspond to the maximum and minimum values ​​of provisional training data RD1, respectively (see pattern P1). The upper dashed line 220 and the lower dashed line 230 in the period [t2, t3] of provisional test data SD2 correspond to the maximum and minimum values ​​of provisional training data RD21 and RD22, respectively (see pattern P2). The upper dashed line 220 and the lower dashed line 230 in the period [t3, t4] of provisional test data SD3 correspond to the maximum and minimum values ​​of provisional training data RD31 and RD32, respectively (see pattern P3).

[0051] The upper dashed line 220 and the lower dashed line 230 in the period (t4, t5) of the provisional test data SD4 correspond to the maximum and minimum values ​​of the provisional training data RD41 and RD42, respectively (see pattern P4). The upper dashed line 220 and the lower dashed line 230 in the period (t5, t6) of the provisional test data SD5 correspond to the maximum and minimum values ​​of the provisional training data RD51 and RD52, respectively (see pattern P5). The upper dashed line 220 and the lower dashed line 230 in the period (t6, t7) of the provisional test data SD6 correspond to the maximum and minimum values ​​of the provisional training data RD6, respectively (see pattern P6).

[0052] The time series graph 200 may be displayed on a display device (see FIG. 20 ). A user can visually check the displayed time series graph 200 and recognize changes over time in the time series data 210. A user can visually check portions 211 and 212 in the time series data 210 and easily recognize the abnormal value AV and the abnormal period. Based on portion 211, the user can recognize that the value of the time series data 210 exceeds the upper dashed line 220 during the period (t4, t5) of the provisional test data SD4. Based on portion 212, the user can recognize that the value of the time series data 210 falls below the lower dashed line 230 during the period (t5, t6) of the provisional test data SD5.

[0053] According to the third modification of the first embodiment, the training data generation device 1 divides the driving data OD into a plurality of patterns corresponding to combinations of provisional training data RD and provisional test data SD. For each of the plurality of patterns, the training data generation device 1 detects an abnormal value AV from the provisional test data SD based on the provisional training data RD. Therefore, the training data generation device 1 can detect an abnormal value AV from a plurality of periods in the driving data OD. The training data generation device 1 excludes each abnormal period in which an abnormal value AV is detected from the driving data OD, and therefore can generate training data XD with less abnormal data compared to the first embodiment.

[0054] (Fourth modification of the first embodiment) Fig. 8 is an explanatory diagram showing the division process of the driving data OD according to the fourth modification of the first embodiment. Fig. 8 shows the steps S11 and S12 of Fig. 2 in order. The period of each data in each process is indicated by a band B1E and six patterns P1 to P6.

[0055] As shown in band B1E in the first stage, the acquisition unit 11 acquires the driving data OD (see step S11). The period of the driving data OD is [t1, t7].

[0056] As shown in the patterns P1 to P6 in the second to seventh rows, the dividing unit 12 divides the driving data OD into six patterns P1 to P6 corresponding to combinations of temporary training data RD and temporary test data SD (see step S12). In the six patterns P1 to P6, the periods of the temporary test data SD1 to SD6 are the same as those in Modification 3. In particular, the lengths of the periods of the temporary test data SD1 to SD6 are the same.

[0057] In pattern P1, there is a blank period G1 between the temporary training data RD1 and the temporary test data SD1. In pattern P2, there is a blank period G21 between the temporary training data RD21 and the temporary test data SD2, and there is a blank period G22 between the temporary training data RD22 and the temporary test data SD2. In pattern P3, there is a blank period G31 between the temporary training data RD31 and the temporary test data SD3, and there is a blank period G32 between the temporary training data RD32 and the temporary test data SD3.

[0058] In pattern P4, there is a blank period G41 between temporary training data RD41 and temporary test data SD4, and there is a blank period G42 between temporary training data RD42 and temporary test data SD4. In pattern P5, there is a blank period G51 between temporary training data RD51 and temporary test data SD5, and there is a blank period G52 between temporary training data RD52 and temporary test data SD5. In pattern P6, there is a blank period G6 between temporary training data RD6 and temporary test data SD6. In particular, the lengths of the ten blank periods G1 to G6 are all the same and correspond to half the lengths of the six periods of temporary test data SD1 to SD6.

[0059] The detection unit 13 detects an abnormal value AV from the provisional test data SD based on the provisional training data RD for each of the six patterns P1 to P6 (see step S13). The method for detecting an abnormal value AV is the same as in the third modification.

[0060] The generation unit 14 generates training data XD by excluding abnormal periods in which abnormal values ​​AV are detected in the six patterns P1 to P6 from the period [t1, t7] of the operating data OD (see step S14). The method of generating the training data XD is the same as in the third modification.

[0061] Fig. 9 is a first explanatory diagram showing the detection process of an abnormal value AV according to the fourth modification of the first embodiment. Fig. 9 shows a time series graph 300A relating to the operating data OD of Fig. 8. The time series graph 300A is similar to the time series graph 200 of Fig. 7. Portions 311A ​​and 312A of the time series data 310 during the abnormal period in which the abnormal value AV was detected are highlighted by dotted lines.

[0062] The period [t1, t7] of the time series data 310 is divided into six periods of provisional test data SD1 to SD6. The upper dashed line 320 and the lower dashed line 330 in the period [t1, t2] of the provisional test data SD1 correspond to the maximum and minimum values, respectively, of the provisional training data RD1 (see pattern P1). The upper dashed line 320 and the lower dashed line 330 are similar to the upper dashed line 220 and the lower dashed line 230, respectively, in FIG. 7.

[0063] Unlike Modification 3, Modification 4 has a blank period G between the provisional training data RD and the provisional test data SD. That is, the length of the period of the provisional training data RD differs between Modification 3 and Modification 4. Therefore, the positions of the upper dashed line 320 and the lower dashed line 330, which respectively indicate the maximum and minimum values ​​in the provisional training data RD, may differ from the positions of the upper dashed line 220 and the lower dashed line 230 in Figure 7.

[0064] The time series graph 300A may be displayed on a display device (see FIG. 20). A user can visually check the time series graph 300A and recognize changes over time in the time series data 310. A user can visually check portions 311A ​​and 312A in the time series data 310 and easily recognize the abnormal value AV and the abnormal period. In particular, the period corresponding to portion 311A ​​(abnormal period) is longer than the period of portion 211 in FIG. 7. On the other hand, the period corresponding to portion 312A (abnormal period) is the same as the period of portion 212 in FIG. 7.

[0065] FIG. 10 is a second explanatory diagram showing the abnormal value AV detection process according to the fourth modification of the first embodiment. A time series graph 300B in FIG. 10 is a graph in which an upper dash-dot line 350 and a lower dash-dot line 360 ​​are superimposed on the time series graph 300A in FIG. 9 . A portion 311B of the time series data 310 that exceeds the upper dash-dot line 350 indicates a true abnormal value AV. A portion 312B of the time series data 310 that falls below the lower dash-dot line 360 ​​indicates a true abnormal value AV. A true abnormal value AV is an ideal abnormal value that should be detected from the time series data 310. A period in which a true abnormal value AV is detected can also be referred to as a true abnormal period.

[0066] Compare the time series graph 200 in Fig. 7 with the time series graph 300B in Fig. 10. The period (abnormal period) of portion 211 in time series graph 200 is much shorter than the period (true abnormal period) of portion 311B in time series graph 300B, and they do not substantially match. That is, in the third modification in Fig. 7, the training data generation device 1 detects only a small portion of the total truly abnormal period, and does not detect the majority of the remaining truly abnormal periods.

[0067] Compare time series graph 300A in FIG. 9 with time series graph 300B in FIG. 10. The period (abnormal period) of portion 311A ​​in time series graph 300A is slightly longer than the period (true abnormal period) of portion 311B in time series graph 300B, and encompasses the period of portion 311B. That is, in Modification 4 in FIG. 9, the training data generation device 1 detects the entire true abnormal period. Compared to Modification 3, the training data generation device 1 in Modification 4 detects abnormal values ​​(or abnormal periods) more appropriately, and therefore can generate training data XD more appropriately.

[0068] According to Modification 4 of the first embodiment, the training data generation device 1 divides the driving data OD into a plurality of patterns relating to combinations of provisional training data RD and provisional test data SD. The training data generation device 1 divides the driving data OD so that, in each of the plurality of patterns, a blank period G exists between the provisional training data RD and the provisional test data SD. Therefore, Modification 4 achieves the same effects as Modifications 1 and 3. Compared to Modification 3, Modification 4 makes it less likely for abnormal values ​​AV to be mixed into the training data XD. As a result, the training data generation device 1 can generate training data XD of higher quality than Modification 3.

[0069] (Second embodiment) 11 is a functional configuration diagram of a training data generation device 1 according to the second embodiment. In the second embodiment, the training data generation device 1 further includes a display control unit 15 and a selection unit 16 in addition to an acquisition unit 11, a division unit 12, a detection unit 13, and a generation unit 14.

[0070] The display control unit 15 controls various displays. The display control unit 15 receives various data (e.g., driving data OD, temporary training data RD, temporary test data SD, and abnormal value AV) from the acquisition unit 11, the division unit 12, and the detection unit 13. The display control unit 15 generates image data GD based on the various received data. The display control unit 15 outputs the generated image data GD to an external device (e.g., a display device) outside the training data generation device 1.

[0071] The selection unit 16 performs various selections. In response to a selection operation SO from the user, the selection unit 16 generates a selection instruction SI corresponding to the selection operation SO. The selection unit 16 transmits the generated selection instruction SI to the generation unit 14. In response to a feedback operation from the user, the selection unit 16 may generate a feedback instruction corresponding to the feedback operation and transmit the feedback instruction to the generation unit 14.

[0072] 12 is an operational flow diagram of the training data generation device 1 according to the second embodiment. The training data generation device 1 sequentially executes the processes of steps S21 to S26.

[0073] (Step S21) First, the training data generation device 1 acquires driving data OD through the acquisition unit 11. Step S21 is similar to step S11. The acquisition unit 11 transmits the driving data OD to the division unit 12, the generation unit 14, and the display control unit 15.

[0074] (Step S22) Next, the training data generation device 1 divides the driving data OD through the dividing unit 12. Step S22 is similar to step S12. The dividing unit 12 transmits the provisional training data RD and the provisional test data SD to the detection unit 13 and the display control unit 15.

[0075] (Step S23) Subsequently, the training data generation device 1 detects an abnormal value AV through the detection unit 13. Step S23 is similar to step S13. The detection unit 13 transmits the abnormal value AV to the generation unit 14 and the display control unit 15.

[0076] (Step S24) Subsequently, the training data generation device 1 displays an image via the display control unit 15. The display control unit 15 generates image data GD based on the driving data OD, temporary training data RD, temporary test data SD, and abnormal value AV obtained in steps S21 to S23. The display control unit 15 displays an image based on the image data GD on the display device.

[0077] The display control unit 15 may display the time series graph 200 of FIG. 7 as the image, or may display the time series graph 300A or 300B of FIG. 9 or 10 on the display device. According to the time series graph 300A of FIG. 9, the abnormal value AV and the abnormal period in the time series data 310 are indicated by portions 311A ​​and 312A. As with portions 311A ​​and 312A, the abnormal portions may be indicated by different line types (e.g., dotted lines). Alternatively, the abnormal portions may be indicated by different colors. In other words, the abnormal portions may be indicated in a manner that allows them to be distinguished from normal portions. Such a display manner allows the user to easily recognize the abnormal portions.

[0078] When the detection unit 13 detects abnormal values ​​AV for multiple process variables included in the operation data OD, the display control unit 15 may display a time series graph for each of the multiple process variables. The display control unit 15 may display each of the time series graphs in parallel. The display control unit 15 may display a time series graph for the process variable having the first ID. The display control unit 15 may switch from the currently displayed time series graph to another time series graph in accordance with an instruction from the user.

[0079] Conversely, if the detection unit 13 does not detect an abnormal value AV for each of the multiple process variables included in the operating data OD, the display control unit 15 may display a string indicating that no abnormal value AV was detected.

[0080] (Step S25) Subsequently, the training data generation device 1 selects a process variable via the selection unit 16. The selection unit 16 receives a selection operation SO (or a feedback operation) from the user on the image displayed in step S24. The selection operation SO may be an operation in which the user selects a specific process variable that the user has determined to be truly abnormal. The selection operation SO may be an operation in which the user selects a desired abnormal period from among the abnormal periods in the time series graph of the process variable. The selection operation SO may be input via an input device (see FIG. 20). The selection operation SO may be an operation in which the user selects a specific process variable that the user has determined not to be abnormal, or an operation in which the user selects a specific period that the user has determined not to be abnormal. By determining in advance the items to be selected by the user, the training data generation device 1 can perform various processes (e.g., identify an abnormal period) in response to the user's operation.

[0081] When a user selects a specific process variable, check boxes corresponding to each of the process variables may be displayed on a display device as a GUI (Graphical User Interface). When the user selects a desired check box, a selection instruction SI for selecting the process variable associated with the selected check box is generated. The selection unit 16 transmits the selection instruction SI corresponding to each selection operation SO to the generation unit 14.

[0082] (Step S26) Finally, the training data generation device 1 generates training data XD through the generation unit 14. The generation unit 14 performs processing corresponding to the selection instruction SI transmitted in step S25. When the selection instruction SI is an instruction to select a specific process variable, the generation unit 14 generates training data XD by excluding an abnormal period from the time-series data corresponding to the specific process variable. When the selection instruction SI is an instruction to select a specific abnormal period, the generation unit 14 generates training data XD by excluding the specific abnormal period from the time-series data of the process variable.

[0083] When there is no process variable corresponding to the selection instruction SI, the generation unit 14 generates the operation data OD as is as training data XD. When there are multiple process variables corresponding to the selection instruction SI, the generation unit 14 excludes each of the abnormal periods detected in the time-series data of each of the multiple process variables from the operation data OD to generate the training data XD.

[0084] According to the second embodiment, the training data generation device 1 generates training data XD based on a selection operation SO from a user. For a process variable that the user has determined to be truly abnormal, the training data generation device 1 generates training data XD by excluding an abnormal period from the time-series data of the process variable. Alternatively, for an abnormal period that the user has determined to be truly abnormal, the training data generation device 1 generates training data XD by excluding the abnormal period from the time-series data of the process variable. Therefore, if the training data generation device 1 erroneously detects a period that is actually normal as an abnormal period, the user can prevent the period from being excluded by not selecting the period.

[0085] Furthermore, the training data generation device 1 displays abnormal periods in the operating data OD in a manner that allows them to be distinguished from normal periods. This display manner allows the user to easily recognize abnormal periods in the operating data OD and easily select the abnormal periods as periods to be excluded. The user selects process variables that exhibit abnormal behavior, and training data XD is automatically generated. Therefore, the training data generation device 1 can reduce the effort required for the user to manually generate training data XD.

[0086] (Modification 1 of the second embodiment) The detection unit 13 may detect an abnormal value AV from the provisional test data SD based on statistics of the process variables related to the provisional training data RD. The detection unit 13 may detect a value of the provisional test data SD that is not included in the range from the L percentile to the H percentile of the provisional training data RD as an abnormal value AV. For example, L=5 and H=95. When L=0, the 0th percentile corresponds to the minimum value. When H=100, the 100th percentile corresponds to the maximum value. When L=0 and H=100, Modification 1 corresponds to the second embodiment.

[0087] For example, L>0 and H<100. In this case, the training data generation device 1 can easily detect true outliers AV.

[0088] (Modification 2 of the second embodiment) The detection unit 13 may detect, from the provisional test data SD, a value whose frequency (or normalized frequency) is less than a threshold in the frequency distribution (or normalized frequency distribution) of the process variable related to the provisional training data RD as an abnormal value AV. The detection unit 13 detects, in the provisional training data RD, a value whose frequency is less than the threshold as an abnormal value AV in the provisional test data SD.

[0089] 13 is a frequency distribution diagram 400A of provisional training data RD according to Modification 2 of the second embodiment. Figures 13(A) to 13(F) show, from left to right, the normalized frequency distributions of provisional training data RD1 to RD6 corresponding to provisional test data SD1 to SD6, respectively. The frequency distribution diagram 400A corresponds to patterns P1 to P6 in FIG. 8 and may be displayed on a display device.

[0090] As shown in Figures 13(A) to 13(F), solid lines indicate normalized frequencies of the provisional training data RD1 to RD6. Dashed lines indicate thresholds for the normalized frequencies. In Figure 13(A), in provisional test data SD1, values ​​greater than VA1 and values ​​less than VA2 are detected as abnormal values ​​AV. In Figure 13(B), in provisional test data SD2, values ​​greater than VB1 and values ​​less than VB2 are detected as abnormal values ​​AV. In Figure 13(C), in provisional test data SD3, values ​​greater than VC1 and values ​​less than VC2 are detected as abnormal values ​​AV. In Figure 13(D), in provisional test data SD4, values ​​greater than VD1 and values ​​less than VD2 are detected as abnormal values ​​AV. In Figure 13(E), in provisional test data SD5, values ​​greater than VE1 and values ​​less than VE2 are detected as abnormal values ​​AV. In FIG. 13(F), in the provisional test data SD6, values ​​greater than the value VF1 and values ​​less than the value VF2 are detected as abnormal values ​​AV.

[0091] According to the second modification of the second embodiment, the same effects as those of the second embodiment can be obtained.

[0092] (Modification 3 of the second embodiment) The display control unit 15 may display the frequency distribution (or normalized frequency distribution) of the process variables related to the temporary training data RD. Furthermore, the display control unit 15 may display the frequency distribution (or normalized frequency distribution) of the process variables related to the temporary test data SD corresponding to the temporary training data RD.

[0093] 14 is a frequency distribution diagram 400B according to Modification 3 of the second embodiment. 14(A) to 14(F) show, from left to right, the normalized frequency distributions of provisional training data RD1 to RD6 and provisional test data SD1 to SD6. The frequency distribution diagram 400B corresponds to patterns P1 to P6 in FIG. 8 and may be displayed on a display device.

[0094] As shown in Figures 14(A) to 14(F), the solid line indicates the normalized frequency distribution (first distribution) of the temporary training data RD. The dashed-dotted line indicates the normalized frequency distribution (second distribution) of the temporary test data SD. HI (Histogram Intersection) indicates the similarity between the first distribution and the second distribution. Other indices (e.g., Kullback-Leibler Divergence, Jensen-Shannon Divergence, Wasserstein Metric) may be used instead of the similarity. The first distribution and the second distribution may be displayed in a distinguishable manner using different display colors, etc. At least one of the first distribution, the second distribution, and the similarity between the first distribution and the second distribution may be displayed.

[0095] The larger the value of HI, the more similar the first distribution and the second distribution are. For example, if the value of HI is less than a threshold, it is defined that the provisional test data SD is not similar to the provisional training data RD. If the threshold is 0.3, the distributions shown in Figures 14(C) to 14(E) satisfy HI<0.3. Therefore, the training data generation device 1 determines that the provisional test data SD3, SD4, and SD5 are not similar to the corresponding training data RD3, RD4, and RD5, respectively.

[0096] If HI is less than the threshold, HI may be colored or surrounded by a border. Alternatively, a border may be added to a distribution where HI is less than the threshold, or the distribution may be colored in a different color. In this way, the training data generation device 1 can highlight and display the provisional test data SD that includes an outlier AV. If the distribution dissimilarity is used, the training data generation device 1 may determine that the provisional test data SD includes an outlier AV if the distribution dissimilarity exceeds a threshold.

[0097] According to the third modification of the second embodiment, the training data generation device 1 displays the distribution of the provisional test data SD and the distribution of the provisional training data RD. The user can easily recognize the difference between the two distributions. Furthermore, the training data generation device 1 displays the similarity (e.g., HI) between the two distributions. The user can quantitatively evaluate the similarity between the two distributions and easily select a process variable or an abnormal period that is determined to be truly abnormal.

[0098] (Fourth modification of the second embodiment) The detection unit 13 may detect an abnormal value AV from the provisional test data SD using a machine learning model trained with the provisional training data RD (see step S23). A known method may be used as the machine learning model (see Patent Document 2). In the following, the provisional training data RD and provisional test data SD in FIG. 8 are used.

[0099] The detection unit 13 trains a machine learning model for temporary test data SD1 using temporary training data RD1 (see pattern P1). The detection unit 13 trains a machine learning model for temporary test data SD2 using temporary training data RD21 and RD22 (see pattern P2). The detection unit 13 trains a machine learning model for temporary test data SD3 using temporary training data RD31 and RD32 (see pattern P3).

[0100] The detection unit 13 trains a machine learning model for temporary test data SD4 using temporary training data RD41 and RD42 (see pattern P4). The detection unit 13 trains a machine learning model for temporary test data SD5 using temporary training data RD51 and RD52 (see pattern P5). The detection unit 13 trains a machine learning model for temporary test data SD6 using temporary training data RD6 (see pattern P6). In this way, the detection unit 13 trains a total of six machine learning models for the six patterns P1 to P6.

[0101] The detection unit 13 generates a predicted value for the period [t1, t2] of the provisional test data SD1 using a machine learning model trained for the provisional test data SD1. Similarly, the detection unit 13 generates a predicted value for the period of the provisional test data SD2 to SD6 using each machine learning model trained for the provisional test data SD2 to SD6. The detection unit 13 transmits the generated predicted values ​​to the display control unit 15.

[0102] The detection unit 13 calculates the deviation by subtracting the predicted value corresponding to the provisional test data SD from the value (measured value) of the process variable included in the provisional test data SD. The detection unit 13 transmits the calculated deviation to the display control unit 15.

[0103] The detection unit 13 detects the timing when the deviation exceeds the upper threshold and the timing when the deviation falls below the lower threshold. The detection unit 13 detects the value of the process variable at both timings as an abnormal value AV. The detection unit 13 transmits the abnormal value AV to the display control unit 15.

[0104] Fig. 15 is an explanatory diagram showing the abnormal value detection process according to Modification 4 of the second embodiment. A time series graph 300C in Fig. 15(A) is similar to the time series graph 300A in Fig. 9. A time series graph 500 in Fig. 15(B) shows deviation.

[0105] As shown in FIG. 15(A), time series data 340 indicating predicted values ​​is superimposed on the time series graph 300C. The predicted values ​​are values ​​predicted using the six machine learning models described above. The time series data 340 indicates the temporal changes in the predicted values ​​using a dashed dotted line. The values ​​of the time series data 310 deviate significantly from the values ​​of the time series data 340 during the periods corresponding to portions 311A ​​and 312A (particularly portion 311A).

[0106] As shown in FIG. 15(B), time series graph 500 shows the deviation between time series data 310 and time series data 340 in time series graph 300C. The horizontal axis of time series graph 500 represents time. The vertical axis of time series graph 500 represents deviation. Time series data 510 shows changes in the deviation over time using dotted and solid lines. Upper dashed line 520 represents the upper threshold of the deviation. Lower dashed line 530 represents the lower threshold of the deviation.

[0107] The dotted line portion of the time series data 510 indicates that the deviation falls within the range between the upper and lower thresholds. The solid line portion of the time series data 510 indicates that the deviation deviates from that range (i.e., an abnormal value AV). Portion 511 of the time series data 510 indicates that the deviation exceeds the upper threshold. Portion 512 of the time series data 510 indicates that the deviation falls below the lower threshold. Portions 511 and 512 correspond to portions 311A ​​and 312A, respectively. Portions 511 and 512 may be displayed in a manner that makes them distinguishable from normal portions by using different display colors, line thicknesses, etc. This display manner is user-friendly because it allows the user to easily recognize abnormal portions.

[0108] According to the fourth modification of the second embodiment, the training data generation device 1 calculates predicted values ​​for the values ​​(measured values) of the provisional test data SD using a machine learning model trained with the provisional training data RD. The training data generation device 1 displays a time series graph showing the deviation between the measured value and the predicted value. Therefore, the training data generation device 1 can present the abnormal values ​​AV and abnormal periods in the operating data OD in a manner that allows the user to easily recognize them.

[0109] (Fifth Modification of the Second Embodiment) In the deviation time-series graph 500 shown in FIG. 15(B), the user may determine that a portion 511 of the time-series data 510 is abnormal and a portion 512 is not abnormal. In this case, the user may perform a selection operation SO to select the portion 511. The selection unit 16 may generate a selection instruction SI corresponding to this selection operation SO and transmit the generated selection instruction SI to the generation unit 14. The generation unit 14 generates training data XD by excluding the abnormal period related to the selection instruction SI from the operation data OD.

[0110] The selection unit 16 may receive feedback indicating in which period of the process variable an abnormality was found. In this case, the display control unit 15 may display a GUI (e.g., a text box or a calendar) for inputting the start time and end time of the period. The user may specify, via the displayed GUI, the period of portion 511 or the period of portion 512 that the user determines to be abnormal. The user can specify the period in more detail via the displayed GUI.

[0111] The display control unit 15 may display a button for adding or deleting a GUI for the user to input an abnormal period, thereby allowing the user to specify multiple abnormal periods even when one process variable includes multiple abnormal periods.

[0112] (Third embodiment) 16 is a functional configuration diagram of a training data generation device 1 according to the third embodiment. As in the second embodiment, the training data generation device 1 includes an acquisition unit 11, a division unit 12, a detection unit 13, a generation unit 14, a display control unit 15, and a selection unit 16.

[0113] The acquisition unit 11 acquires reference data FD in addition to the operating data OD. The reference data FD is data from a different period from the operating data OD. The operating data OD and the reference data FD may be data related to the operating state of the same device but from different periods. The relationship between the operating data OD and the reference data FD is similar to the relationship between the provisional test data SD and the provisional training data RD. The acquisition unit 11 transmits the operating data OD and the reference data FD to the detection unit 13.

[0114] 17 is an operational flow diagram of the training data generation device 1 according to the third embodiment. The training data generation device 1 sequentially executes the processes of steps S31 to S38.

[0115] (Steps S31 to S33) First, the training data generation device 1 acquires driving data OD through the acquisition unit 11. Next, the training data generation device 1 divides the driving data OD through the division unit 12. Subsequently, the training data generation device 1 detects an abnormal value AV through the detection unit 13. Steps S31 to S33 are the same as steps S21 to S23, respectively.

[0116] (Step S34) Subsequently, the training data generation device 1 acquires the reference data FD through the acquisition unit 11.

[0117] (Step S35) Subsequently, the training data generation device 1 detects abnormal values ​​AV through the detection unit 13. The detection unit 13 assumes that the reference data FD acquired in step S34 is data in a normal state (normal data), and detects abnormal values ​​AV (other abnormal values ​​AV) from the operating data OD. Any of the methods described above can be used to detect abnormal values ​​AV.

[0118] The detection unit 13 may determine whether the value of the operating data OD is included in a range between a maximum value and a minimum value related to the reference data FD. The detection unit 13 may calculate a predicted value of the operating data OD using a machine learning model trained using the reference data FD. The detection unit 13 may determine whether the deviation between the predicted value and the value (measured value) of the operating data OD is included in a range between an upper threshold and a lower threshold.

[0119] (Steps S36 to S38) Next, the training data generation device 1 displays an image via the display control unit 15. Next, the training data generation device 1 selects a process variable via the selection unit 16. Finally, the training data generation device 1 generates training data XD via the generation unit 14. Steps S36 to S38 are similar to steps S24 to S26, respectively.

[0120] Figure 18 is an explanatory diagram showing the detection process of an abnormal value AV according to the third embodiment. The time series graph 600 in Figure 18(A) is similar to the time series graph 300C in Figure 15(A). The time series graph 700 in Figure 18(B) is similar to the time series graph 500 in Figure 15(B).

[0121] As shown in FIG. 18(A), a time series graph 600 is a graph showing changes in the operating data OD over time. The time series data 610 shows changes in the process variable values ​​of the operating data OD over time using solid and dotted lines. The upper dashed line 620 shows the maximum value of the reference data FD. The lower dashed line 630 shows the minimum value of the reference data FD. The time series data 640 shows a predicted value based on a machine learning model trained using the reference data FD using a dashed-dotted line. Portion 611 of the time series data 610 indicates a portion of the time series data 610 that exceeds the upper dashed line 620 (i.e., an abnormal value AV). Portion 612 of the time series data 610 indicates a portion of the time series data 610 that falls below the lower dashed line 630 (i.e., an abnormal value AV).

[0122] A user can visually check the portions 611 and 612 and easily recognize the abnormal values ​​AV in the time-series data 610. The portions 611 and 612 may be displayed in a distinguishable manner using different colors, different line thicknesses, or other display modes.

[0123] As shown in FIG. 18(B), a time series graph 700 shows the deviation between the value (measured value) of the process variable of the operating data OD and the predicted value. The time series data 710 shows the change in the deviation over time using dotted and solid lines. The upper dashed line 720 indicates the upper threshold for the deviation. The lower dashed line 730 indicates the lower threshold for the deviation. Portion 711 of the time series data 710 indicates the portion of the time series data 710 that exceeds the upper dashed line 720 (i.e., the abnormal value AV). Portion 712 of the time series data 710 indicates the portion of the time series data 710 that falls below the lower dashed line 730 (i.e., the abnormal value AV).

[0124] A user can visually check the portions 711 and 712 and easily recognize the abnormal value AV in the time-series data 710. The portions 711 and 712 may be displayed in a distinguishable manner using different colors, different line thicknesses, or other display modes.

[0125] 19 is a frequency distribution chart 800 of the operating data OD and the reference data FD according to the third embodiment. The frequency distribution chart 800 in FIG. 19 is similar to the frequency distribution chart 400B in FIG.

[0126] In the frequency distribution diagram 800, the dashed-dotted line indicates the normalized frequency distribution of the operating data OD. The solid line indicates the normalized frequency distribution of the reference data FD. The similarity between the two distributions is displayed as "HI=0.5."

[0127] A user can easily recognize a transition (shift) between the distribution of the operating data OD and that of the reference data FD by visually checking the frequency distribution chart 800. Instead of the normalized frequency distribution, a frequency distribution may be displayed. Instead of the HI, another index (described above) indicating the degree of difference between the two distributions may be displayed.

[0128] 20 is a hardware configuration diagram of a training data generation device 1 according to each embodiment. The training data generation device 1 includes, as its components, a CPU 111, a RAM 112, a ROM 113, a storage 114, a display device 115, an input device 116, and a communication device 117. The components are connected to each other via an internal bus so that they can communicate with each other. The training data generation device 1 may include at least some of the components.

[0129] The CPU 111 is a processor that executes various processes according to programs. The CPU 111 uses a predetermined area of ​​the RAM 112 as a work area. The CPU 111 realizes each unit (e.g., the acquisition unit 11, the division unit 12, the detection unit 13, the generation unit 14, the display control unit 15, and the selection unit 16) by reading and executing each program stored in the ROM 113 or the storage 114. The CPU 111 is an example of a processing unit.

[0130] The RAM 112 is a memory that stores various types of data in a rewritable manner. For example, the RAM 112 is a Synchronous Dynamic Random Access Memory (SDRAM). The RAM 112 is an example of a storage unit.

[0131] The ROM 113 is a memory that stores various data in a non-rewritable manner. The ROM 113 is an example of a storage unit.

[0132] The storage 114 is a storage medium of any kind. The storage 114 may be a drive device that writes or reads various data to or from the storage medium. The storage 114 may write or read various data to or from the storage medium under control of the CPU 111. The storage 114 is an example of a storage unit.

[0133] Display device 115 is a device that displays various types of data. Display device 115 may be an LCD (Liquid Crystal Display). Display device 115 displays various types of data (e.g., images, time series graphs, distribution maps) based on a display signal from CPU 111. Display device 115 is an example of a display unit.

[0134] The input device 116 is a device that accepts various input operations from the user. The input device 116 may be a mouse or a keyboard. The input device 116 accepts operations input by the user as instruction signals and transmits the instruction signals to the CPU 111. The input device 116 is an example of an input unit.

[0135] The communication device 117 communicates with external devices via a network under the control of the CPU 111. The communication device 117 is an example of a communication unit.

[0136] Although several embodiments of the present invention have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These novel embodiments can be embodied in various other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their modifications are included within the scope and spirit of the invention, and are also included in the scope of the invention and its equivalents as defined in the claims. [Explanation of symbols]

[0137] 1...training data generation device, 11...acquisition unit, 12...division unit, 13...detection unit, 14...generation unit, 15...display control unit, 16...selection unit, 111...CPU, 112...RAM, 113...ROM, 114...storage, 115...display device, 116...input device, 117...communication device, 200, 300A, 300B, 300C, 500, 600, 700...time series graph, 210, 310, 340, 510, 610, 640, 710...time series data, 211, 212, 311A, 311B, 312A, 312B, 51 1,512,611,612,711,712...parts, 220,320,520,620,720...upper dashed line, 230,330,530,630,730...lower dashed line, 350...upper dashed line, 360...lower dashed line, 400A,400B,800...frequency distribution map, AV...abnormal value, B...band, FD...reference data, G...blank period, GD...image data, OD...operation data, P...pattern, RD...temporary training data, SD...temporary test data, SI...selection instruction, SO...selection operation, V...value, XD...training data

Claims

1. an acquisition unit that acquires operation data relating to the operating state of the equipment for a predetermined period of time; a division unit that divides the driving data into at least provisional training data and provisional test data; a detection unit that detects an abnormal value from the provisional test data based on the provisional training data; a generation unit that generates training data from the driving data by excluding an abnormal period in which the abnormal value is detected from the predetermined period; A training data generation device comprising:

2. the dividing unit divides the driving data so that a blank period exists between the provisional training data and the provisional test data. The training data generating device according to claim 1 .

3. the dividing unit divides the driving data into a plurality of patterns relating to combinations of the provisional training data and the provisional test data; the detection unit detects the abnormal value from the provisional test data based on the provisional training data for each of the plurality of patterns; The training data generating device according to claim 1 .

4. the dividing unit divides the driving data such that the duration of each of the provisional test data differs among the plurality of patterns. The training data generating device according to claim 3 .

5. the dividing unit divides the predetermined period of the driving data into periods of the provisional test data in the plurality of patterns; The training data generating device according to claim 3 .

6. the dividing unit divides the driving data such that a blank period exists between the provisional training data and the provisional test data in each of the plurality of patterns. The training data generating device according to claim 3 .

7. a display control unit that displays a time series graph relating to the driving data on a display unit; the display control unit emphasizes the abnormal period in the time series graph. The training data generating device according to claim 1 .

8. a selection unit that selects at least one abnormal period from among the abnormal periods in the time series graph in response to an operation from a user, the generation unit generates the training data by excluding the selected abnormal period from the predetermined period. The training data generating device according to claim 7 .

9. the operational data includes time series data for each of a plurality of process variables; a display control unit that displays a time series graph relating to the time series data on a display unit; the display control unit emphasizes the abnormal period in the time series graph. The training data generating device according to claim 1 .

10. a selection unit that selects at least one process variable from among the plurality of process variables in response to an operation from a user, the generation unit generates the training data by excluding the abnormal period from the predetermined period for the time-series data corresponding to the selected process variable. The training data generating device according to claim 9 .

11. a display control unit that displays at least one of a first distribution related to the provisional training data, a second distribution related to the provisional test data, and a similarity between the first distribution and the second distribution on a display unit. The training data generating device according to claim 1 .

12. The detection unit detects the outlier from the provisional test data using a machine learning model trained using the provisional training data. The training data generating device according to claim 1 .

13. the acquisition unit acquires reference data having a period different from that of the operating data, The detection unit detects other abnormal values ​​from the operating data based on the reference data, the generation unit generates the training data by excluding other abnormal periods in which the other abnormal values ​​are detected from the predetermined period. The training data generating device according to claim 1 .

14. The computer Acquire operating data regarding the operating status of the equipment for a predetermined period of time; Dividing the driving data into at least provisional training data and provisional test data; Detecting outliers from the provisional test data based on the provisional training data; generating training data from the driving data by excluding an abnormal period in which the abnormal value is detected from the predetermined period; Training data generation method.

15. On the computer, an acquisition function for acquiring operation data relating to the operating state of the equipment for a predetermined period of time; a division function for dividing the driving data into at least provisional training data and provisional test data; a detection function for detecting an outlier from the provisional test data based on the provisional training data; a generation function of generating training data from the operation data by excluding an abnormal period in which the abnormal value is detected from the predetermined period; A training data generation program that achieves this.

Images